Research Collection

Doctoral Thesis

Three Essays based on Clickstream Data: Analyzing,Understanding and Managing Online Customer Behavior

Author(s): Becker, Ingo Frank

Publication Date: 2016

Permanent Link: https://doi.org/10.3929/ethz-a-010616045

Rights / License: In Copyright - Non-Commercial Use Permitted

This page was generated automatically upon download from the ETH Zurich Research Collection. For moreinformation please consult the Terms of use.

ETH Library

DISS. ETH NO. 23317

Three Essays based on Clickstream Data:

Analyzing, Understanding and Managing

Online Customer Behavior

A thesis submitted to attain the degree of

DOCTOR OF SCIENCES of ETH ZURICH

(Dr. sc. ETH Zurich)

presented by

INGO FRANK BECKER

Dipl.-Kfm. Univ., Technische Universität München

born October 27, 1982

citizen of Germany

accepted on the recommendation of

Prof. Dr. Florian von Wangenheim, examiner

Prof. Dr. René Algesheimer, co-examiner

2016

For Mum, Dad, Vito, and Kathi

Summary V

Summary

The internet is earth shaking. Two decades ago its global triumphal procession has begun and

is still ongoing to change how companies and customers interact. In opposition to traditional

media, the internet allows for gathering a plethora of data associated with millions of users,

companies encompassing a permanently evolving refraction of online marketing channels.

Novel business models and the refinement of old ones are adapting quickly and attempt to

harvest the most from this unique and overwhelming set of opportunities—however, often

flying blind. In this dissertation, we provide three individual studies to enlighten some parts of

this black spot.

Study 1 focuses on dissolving the attribution challenge, the question to what degree each

channel actually contributes to the marketing success of an advertising company. Advertisers

employ a variety of online channels to reach customers, however, often rely on simple and

imprecise heuristics such as “last click wins”. Recently, more elaborate academic models are

evolving, yet, generalizable insights that apply across industries on the effectiveness of

individual online channels in multichannel environments, and on their interplay, remain scarce.

In order to enlighten “what we may learn from advanced attribution”, we introduce a novel,

graph-based framework that represents customer path data as first- and higher-order Markov

walks and discuss it in the context of two well-established attribution heuristics, namely first

click and last click wins, as well as two logit benchmark models, embracing both scientific rigor

and practical application. The appliance of the proposed framework to four, large, real-world

data sets from three different industries allows for generalizations and cross-industry

comparisons. Our results differ substantially from the comparison models, and confirm as well

as refine previous research on singular data sets. For instance, we find indications for

idiosyncratic channel preferences (carryovers) and interactions within and across channel

categories (spillover). Thus, the framework scales down the gap between widely applied, but

Summary VI

inadequate methods, such as “last click wins”, and an ideal, but impractical attribution models

also taking into account data availability and managerial considerations.

Study 2 investigates on customer browsing behavior and aims to extract hidden purchase

intention manifested in the users browsing traits. While prior research has introduced valuable

studies on the effectiveness of individual online channels, spillover effects between two online

channels, and, partly, on multichannel effects also consolidating channels into channel

categories, the link between inter-exposure times of firm-customer contacts and purchase

propensity remains untouched. To address this gap, we develop a novel concept that implements

a time component between subsequent clicks. Routing from flow theory, we conceptualize

“focused attention” as users' browsing pattern, by aggregating intense browsing sessions

measured by timely close, successive, singular clicks into a novel concept, which we call the

micro-journey. Using a Cox proportional hazards model on four large scale individual user level

data sets, we reveal that micro-journeys are a well-suited predictor of conversion events. Users

with micro-journeys are more likely to convert—to equal proportions, they convert directly

after the micro-journey as well as thereafter with some delay, using single clicks. We further

characterize the micro-journeys and find meaningful browsing patterns that help to better

understand customers’ decision making progress including direct as well as indirect,

procrastinated conversion events. Thereby, this study introduces an innovative and valuable

construct adding a novel facet to advertisers’ increasing demand for a more holistic

investigation of online marketing effectiveness.

Study 3 focuses on users’ channel preferences in multichannel online environments.

Today, advertisers can select from a variety of online channels to engage with potential

customers, who, in turn, may exhibit preferences toward particular channels mirrored in their

pursuit (or ignorance) of channels on their path to purchase. Anticipating this proliferation of

online channels and the users’ browsing intentions make marketing decisions increasingly

Summary VII

complex. While we know from the offline world that advertisement exposure across different

types of media increases the advertiser’s credibility and recognition and, in consequence, sales,

we are still unaware if and how multichannel exposure in pure online environments translates

into conversions. Adding to research on dedicated online phenomena also including category

approaches that help to approximate the users’ (underlying) intentions, we apply Cox

proportional hazards model on the users’ full browsing history to elaborate on users channel

and category preferences and derive novel and interesting interaction effects. Across four data

sets, and opposed to the offline world, users clearly show idiosyncratic (homogeneous) channel

preferences. Although the results are more diverse with regard to channel taxonomies, we

derive novel, industry-specific and generalizable insights, for instance, that past exposures

toward informational channels may express information acquisition anteceding future

purchases. In that way, this study contributes to research on marketing effectiveness, and

supports advertisers in shaping their marketing activities.

The analysis of multichannel online environments is of pivotal importance for online

companies, especially in vigorous and complex environments. This dissertation contributes to

online marketing effectiveness theory and practice by conceptualizing user behavior and

translating it into conversion propensity, analyzing users’ browsing preferences manifested in

their browsing traits, and modeling and attributing online channel contributions. Thus, it not

only provides guidance and novel perspectives in marketing research, but also helps advertisers

to re-calibrate and enhance their marketing measures.

Zusammenfassung IX

Zusammenfassung

Das Internet verändert die Welt. In den vergangenen zwei Jahrzehnten hat sein Einfluss stark

zugenommen und gestaltet weiterhin maßgelblich die Beziehungen zwischen Kunden und

Unternehmen. Im Gegensatz zu traditionellen Medien ermöglichen Internet-Technologien das

Sammeln und Auswerten zahlreicher Kundendaten. Gleichzeitig können Unternehmen auf ein

sich fortlaufend differenzierendes Portfolio an Online-Kanälen für Ihre Marketingmaßnahmen

zurückgreifen. Vorhandene Geschäftsmodelle befinden sich im Wandel, Neue entstehen, und

stellen einzigartige und vielversprechende Möglichkeiten für Werbetreibende zur Verfügung—

allerdings bestehen nach wie vor Wissenslücken, die dazu führen, dass Unternehmen oft im

Blindflug agieren. Mit Hilfe von drei in sich selbstständigen Studien dieser Dissertation

möchten wir einen Betrag leisten Licht ins Dunkel zu bringen und Entscheider bei ihren

Marketingmaßnahmen zu unterstützen.

Die erste Studie befasst sich mit dem Thema „Kanal-Attribution“ und zielt folglich

darauf ab die Frage zu lösen, in welchem Maße jeder einzelne Online-Marketing-Kanal zum

Geschäftserfolg von Internet-Unternehmen in Multikanal-Umgebungen beiträgt. Unternehmen

nutzen eine Vielzahl von unterschiedlichen Online-Kanälen um Kunden anzusprechen, dabei

verlassen sie sich jedoch häufig auf einfache Heuristiken, wie beispielsweise auf die Methode

„der letzte Klick gewinnt“, welche lediglich den letzten Klick vor einer Transaktion

berücksichtigt. Auch wenn seit kurzem fortschrittlichere Attributionstechniken in der

Wissenschaft entwickelt wurden, so sind diese meist Kontext gebunden und lassen weder

Generalisierungen über Industrien hinweg zu noch bieten sie ein umfassendes Bild über

Interaktionen zwischen einzelnen Kanälen in Multikanal-Umgebungen. Um einen Schritt

weiter zu gehen und nicht nur eine weitere Attributionslogik vorzustellen, enthält diese Studie

zahlreiche Weisungen „was wir von fortgeschrittenen Attributionstechniken lernen können“.

Hierfür entwickeln wir eine Graphen-basierte Attributionstechnik, welche auf Markov-Ketten

Zusammenfassung X

erster sowie höherer Ordnung basiert, und diskutieren diese im Kontext von zwei in der Praxis

weit verbreiteten Heuristiken, „der erste Klick gewinnt“ und „der letzte Klick gewinnt“, sowie

zwei logistischen Regressionsmodellen, um die wissenschaftliche Rigorosität und die

praktische Relevanz zu gewährleisten. Vier umfassende, empirische Datensätzen von drei

unterschiedlichen Branchen helfen uns dabei Generalisierungen über mehrere Industrien

hinweg von Industriespezifika zu differenzieren. Unsere Resultate heben sich deutlich von den

beiden vereinfachenden Heuristiken ab und bestätigen und verfeinern bestehendes Wissen aus

wissenschaftlichen Untersuchungen basierend auf nur einem Datensatz. Beispielsweise

identifizieren wir Tendenzen für idiosynkratische Kanalpräferenzen von Online-Nutzern

(„Carryovers“) oder substantielle Interaktionseffekte zwischen Kanälen innerhalb einer

Kanalkategorie sowie über Kanalkategorien hinweg („Spillovers“). Diese Studie kann somit als

Fortsetzung der jüngst erschienenen Literatur zum Thema Attribution verstanden werden und

schließt eine relevante Wissenslücke zwischen der praktischen Anwendbarkeit von

simplifizierten Methoden, wie „der letzte Klick gewinnt“, sowie einem theoretischen Ideal, das

durch Datenverfügbarkeit und praktischer Akzeptanz limitiert wird.

In der zweiten Studie wird das Klickverhalten von Internet-Nutzern analysiert um darin

zugrundeliegende, aber verborgene Kaufabsichten besser sichtbar zu machen. Während

bestehende Studien wertvolle Beiträge leisten, um die Effektivität von einzelnen Kanälen,

„Spillover-Effekte“ zwischen zwei Kanälen, sowie, zu einem geringeren Teil, Effekte von

Kanalgruppierungen in sogenannten Kanalkategorien, zu verstehen, so gibt es bislang keinerlei

Online-Studien, welche die Zeitabstände zwischen einzelnen Unternehmen-Kunden-

Interaktionen in ihren Analysen berücksichtigen. Um genau an diesem Vakuum anzusetzen,

entwickeln wir im Rahmen dieser Studie ein neues Konzept, das eine Zeitkomponente

aufeinanderfolgender Klicks implementiert. Hierbei bedienen wir uns der „Flow-Theorie“ und

übersetzen ihr Element „focused attention“ in ein Verhaltensmuster von Internet-Nutzern,

indem wir intensiveres Surfverhalten messen, an Hand zeitlich kurz aufeinander folgender

Zusammenfassung XI

Einzelklicks, und dieses gesamthaft in ein neues Konzept überführen, die sogenannte „Micro-

Journey“. Unter Anwendung eines Cox Proportional Hazards Modells können wir bestätigen,

dass Micro-Journeys sowohl die Modellgüte verbessern als auch eine geeignete Ergänzung

darstellen, um Online-Einkäufe besser vorherzusehen. Internet-Nutzer, welche Micro-Journeys

in ihrem Surfverhalten aufweisen, tendieren häufiger zu Kaufentscheidungen als Nutzer ohne

Micro-Journey—hierbei konvertieren sie zu etwa gleichen Anteilen direkt nach der Micro-

Journey oder, mit etwas Abstand, danach unter Verwendung von Einzelklicks. Des Weiteren

differenzieren wir die Micro-Journey entsprechend weiterer messbarer Eigenschaften und

identifizieren bedeutsame Verhaltensmuster, die helfen können den Fortschritt im

Kaufentscheidungsprozess besser abzubilden, unter anderem unter Berücksichtigung direkter

sowie zeitlich verschobener Käufe. Das neue Konzept der Micro-Journey bildet somit eine

innovative und wertvolle Facette, um der steigenden Nachfrage Werbetreibender nach

ganzheitlichen Analysemethoden von Online-Marketing-Effektivität gerecht zu werden.

Daran anknüpfend veranschaulicht die dritte Studie zu Grunde liegende

Kanalpräferenzen von Online-Kunden in Online-Umgebungen. Heutzutage greifen Internet-

Unternehmer auf eine Vielzahl unterschiedlicher Online-Kanäle zurück, um mit potenziellen

Kunden zu interagieren. Kunden hingegen können individuell verschiedene Kanalpräferenzen

aufweisen, die sich in ihrer Klickneigung (oder ihrer Klickverweigerung) während Online-

Kaufvorgängen für Werbetreibende erkennbar widerspiegeln. Allerdings wird es zunehmend

komplexer, durch die Zunahme der Kanal-Diversität, die (Kauf-)Absichten potenzieller

Kunden in ihren Surfvorgängen zu verstehen und daraus Marketingmaßnahmen abzuleiten.

Obwohl aus der Offline-Welt bekannt ist, dass Werbebotschaften die über mehrere Kanäle

ausgestrahlt werden, das Vertrauen in eine Marke sowie ihre Wiedererkennungsrate steigern

und damit längerfristig Umsatzerfolge erhöht werden, so bleibt nach wie vor ungeklärt, ob und

wie die Nutzung mehrerer Kanäle in reinen Online-Umgebungen Kaufabschlüsse beeinflusst.

Um der Frage nach den Kanal- und Kanalkategorie-Präferenzen von Internet-Nutzern im

Zusammenfassung XII

Kaufentscheidungsprozess nachzugehen, wenden wir ein Cox Proportional Hazards Modell auf

vollständige Surfverläufe an, aufgezeichnet als „Customer Journeys“, und knüpfen somit

nahtlos an bestehende Literatur an, welche Nutzerverhalten selektiv, zum Teil auch unter

Einbeziehung von Kanalkategorisierungen, untersucht. Über alle vier modellierten Datensätze

hinweg bestätigt unsere Studie einen Widerspruch zu traditionellen Medien, welcher sich in

einer starken Tendenz zu idiosynkratische Kanalpräferenzen (homogene Kanalpräferenzen)

von Internet-Nutzern manifestiert. Obgleich unsere Ergebnisse, bezogen auf

Kanalkategorisierungen, weniger trennscharf sind, so generieren wir auch hier neue, sowohl

industriespezifische als auch generalisierbare Erkenntnisse, beispielsweise, dass ein

ausgeprägter Anteil informationeller Kanäle in der Surfhistorie eines Nutzers auf eine Phase

der Informationsakquisition durch den Nutzer hinweist, welche tendenziell zukünftigen Käufen

voraus geht. Durch diese sowie weiterer, neuer und relevanter Erkenntnisse bildet diese Studie

einen weiteren Baustein im Mosaik der Forschung rund um Online-Marketing-Effektivität und

trägt dazu bei, die Entscheider bei der Definition von Marketing Maßnahmen zu unterstützen.

Die Analyse von mehreren Online-Kanälen im Multikanal-Kontext ist von zentraler

Bedeutung für Unternehmen, insbesondere da Online-Umgebungen einem schnellen Wandel

mit zunehmenden Herausforderungen unterliegen. Somit leistet diese Dissertation einen

wertvollen Beitrag zur Weiterentwicklung von Online-Marketing-Effektivität in Theorie und

Praxis.

Acknowledgements XIII

Acknowledgements

I am delighted to take this opportunity to express my comprehensive thanks to my supporters.

First and foremost, I want to express my candid gratitude to my advisor, critic, and motivator

Florian von Wangenheim for his incommensurable support, brilliant ideas, and his

encouragement. A plethora of thanks are dedicated to my second advisor René Algesheimer for

his immediate support and all his efforts to bringing this dissertation to a success. Thank you.

I owe particular thanks to two people who accompanied me as my co-authors over the

past years: Eva Anderl for being my fellow in pushing our research again and again during the

ambitious scientific publication process. Another very special word of gratitude goes to Marc

Linzmajer. Thank you for your humor and infectious enthusiasm. It was great working with

you and I would do it again any time.

Furthermore, I deeply acknowledge the comprehensive and pragmatic support I received

from the intelliAd GmbH, my practice partners. Without their commitment a dissertation of this

kind would not have been possible. I want to especially mention Tobias Kiessling, Mischa

Rürup, Franz Graf, Darius Suryadi, and Alice Meier for their generous support.

I wish to send my sincere words of gratitude also to a handful of very special people.

Thank you Mum, thank you Dad, for all your support, patience and, above-all, your trust—in

the past, the presence and in the future. A huge thank you goes to two very smart people that I

may count among my best friends: Philipp, for backing me up whenever needed and Jens, for

the vital discussions on statistical challenges. Very special words of gratitude are reserved for

my little son, Vito, who was the most charming distraction from work. Finally, and foremost,

thank you, Kathi, for giving me your warm heart accompanying me on this and on all other

journeys life will ever invent.

Before you move on, I want to cordially thank you as an interested reader, for whom I

decided to write this manuscript.

Short Table of Contents XV

Short Table of Contents

Summary ................................................................................................................................... V

Zusammenfassung .................................................................................................................... IX

Acknowledgements ............................................................................................................... XIII

Short Table of Contents .......................................................................................................... XV

Table of Contents ................................................................................................................. XVII

List of Figures .................................................................................................................... XXIII

List of Tables ....................................................................................................................... XXV

List of Abbreviations ......................................................................................................... XXVII

1 Introduction ......................................................................................................................... 1

2 Mapping the Customer Journey: Lessons Learned from Graph-Based Online

Attribution Modeling ........................................................................................................ 29

3 Patterns that Matter: How Browsing Click Patterns as Micro-Journeys Influence

Customer Conversions ...................................................................................................... 69

4 Channels and Categories: User Browsing Preferences on the Path to Purchase ............. 121

5 Conclusion ....................................................................................................................... 163

6 References ....................................................................................................................... 175

7 Appendix ......................................................................................................................... 203

Table of Contents XVII

Table of Contents

Summary ................................................................................................................................... V

Zusammenfassung .................................................................................................................... IX

Acknowledgements ............................................................................................................... XIII

Short Table of Contents .......................................................................................................... XV

Table of Contents ................................................................................................................. XVII

List of Figures .................................................................................................................... XXIII

List of Tables ....................................................................................................................... XXV

List of Abbreviations ......................................................................................................... XXVII

1 Introduction ......................................................................................................................... 1

1.1 The Digital Era: Relevance, Development and Aspiration .................................. 1

1.2 Research on Online Marketing Effectiveness ....................................................... 2

1.2.1 Single online marketing channels ............................................................ 2

1.2.2 Multichannel Research and Attribution ................................................... 6

1.2.3 The Purchase Decision Process and Channel Categories ........................ 9

1.2.4 Research on Clickstream Data ............................................................... 14

1.3 Shortcomings and Research Scope ..................................................................... 18

1.3.1 Essay 1 – Mapping the Customer Journey: Lessons Learned from

Graph-Based Online Attribution Modeling .......................................... 19

1.3.2 Essay 2 – Patterns that Matter: How Browsing Click Patterns as

Micro-Journeys Influence Customer Conversions................................ 21

Table of Contents XVIII

1.3.3 Essay 3 – Channels and Categories: User Browsing Preferences on

the Path to Purchase .............................................................................. 23

1.4 Structure of the Dissertation ............................................................................... 25

1.5 List of Publications ............................................................................................. 27

2 Mapping the Customer Journey: Lessons Learned from Graph-Based Online

Attribution Modeling ........................................................................................................ 29

2.1 Introduction ......................................................................................................... 30

2.2 Research Background ......................................................................................... 34

2.3 Data ..................................................................................................................... 36

2.4 Model Development ........................................................................................... 40

2.4.1 Base Model ............................................................................................ 41

2.4.2 Higher-Order Models ............................................................................. 43

2.4.3 Removal Effect ...................................................................................... 43

2.5 Model Fit ............................................................................................................ 45

2.5.1 Predictive accuracy ................................................................................ 46

2.5.2 Robustness ............................................................................................. 49

2.6 Results................................................................................................................. 51

2.6.1 Attribution Results ................................................................................. 51

2.6.2 Interplay of Channels ............................................................................. 56

2.7 Discussion ........................................................................................................... 61

2.8 Outlook ............................................................................................................... 65

Table of Contents XIX

3 Patterns that Matter: How Browsing Click Patterns as Micro-Journeys Influence

Customer Conversions ...................................................................................................... 69

3.1 Introduction ......................................................................................................... 70

3.2 Conceptual Development of the Micro-Journey ................................................. 73

3.2.1 The Concept of Flow ............................................................................. 73

3.2.2 The Concept of the Micro-Journey ........................................................ 74

3.3 Empirical Setting and Extraction of the Micro-Journey ..................................... 76

3.3.1 Clickstream Data .................................................................................... 76

3.3.2 Extraction of the Micro-Journey ............................................................ 78

3.3.3 Characterizing the Micro-Journey ......................................................... 80

3.4 Model Development ........................................................................................... 84

3.4.1 General Model Formulation ................................................................... 84

3.4.2 The Micro-Journey as Predictor (Model 1) ........................................... 84

3.4.3 The Micro-Journey Characteristics (Model 2) ....................................... 88

3.4.4 The Micro-Journey and Procrastination (Model 3) ............................... 91

3.5 Estimation Results .............................................................................................. 93

3.5.1 Results of Model 1 – The Micro-Journey .............................................. 93

3.5.2 Results of Model 2 – The Micro-Journey Characteristics ..................... 96

3.5.3 Results of Model 3 – The Micro-Journey and Purchase Timing ......... 100

3.6 Robustness of the Estimation Results ............................................................... 102

3.6.1 Results of Model 1 ............................................................................... 103

3.6.2 Results of Model 2 ............................................................................... 104

Table of Contents XX

3.6.3 Results of Model 3 ............................................................................... 108

3.7 General Discussion ........................................................................................... 113

3.7.1 Theoretical Implications ...................................................................... 115

3.7.2 Marketing Implications ........................................................................ 116

3.7.3 Directions for Further Research ........................................................... 117

4 Channels and Categories: User Browsing Preferences on the Path to Purchase ............. 121

4.1 Introduction ....................................................................................................... 122

4.2 Conceptual Development .................................................................................. 128

4.2.1 Conceptual Model ................................................................................ 128

4.2.2 Categorization Approaches .................................................................. 131

4.3 The Four Data Sets ........................................................................................... 133

4.4 Model Development ......................................................................................... 136

4.4.1 General Model Formulation ................................................................. 136

4.4.2 Modeling Interaction Effects ............................................................... 138

4.4.3 Modeling Category Effects .................................................................. 139

4.5 Estimation Results ............................................................................................ 140

4.5.1 Present Channel Effects ....................................................................... 140

4.5.2 Past Channel Effects ............................................................................ 144

4.5.3 Homogenous and Heterogeneous Channel Interactions ...................... 146

4.5.4 Category Interactions ........................................................................... 149

4.6 Discussion ......................................................................................................... 153

4.7 Limitations and Research Directions ................................................................ 157

Table of Contents XXI

5 Conclusion ....................................................................................................................... 163

5.1 Implications ...................................................................................................... 163

5.2 Outlook ............................................................................................................. 169

6 References ....................................................................................................................... 175

7 Appendix ......................................................................................................................... 203

7.1 Definition of Online Marketing Channels (Essay 1 – 3) .................................. 207

7.2 Alternate Model Specifications (Essay 2) ........................................................ 209

7.2.1 Time-Dependent Covariates (Model 1) ............................................... 209

7.2.2 Continuous Covariates (Model 1) ........................................................ 210

7.2.3 Vector of Covariates (Model 2b) ......................................................... 211

7.3 Definition of R2D and R2

PH (Essay 2) ............................................................... 212

List of Figures XXIII

List of Figures

Figure 1 Dynamic Choice Process Based on Shocker et al. (1991) ......................................... 10

Figure 2 Structure of the Dissertation ...................................................................................... 26

Figure 3 Exemplary Markov Graph—Essay 1 ......................................................................... 42

Figure 4 Markov Graph for Data Set 1 (base model)—Essay 1 .............................................. 42

Figure 5 ROC Curves (within Sample)—Essay 1 .................................................................... 47

Figure 6 Prior Clickstream Approaches and the Concept of the Micro-Journey—Essay 2 ..... 71

Figure 7 The Micro-Journey versus Single Clicks—Essay 2 .................................................. 79

Figure 8 Conversions of All Journeys versus Journeys with Micro-Journeys—Essay 2 ......... 84

Figure 9 Direct versus Later Conversions after One Micro-Journey (in %)—Essay 2 ............ 91

Figure 10 The Transition Tree of the Sequential Models (Model 3)—Essay 2 ....................... 92

Figure 11 Conceptual Model of Relationships between Channels, Categories, and

Conversions—Essay 3 ............................................................................................. 130

List of Tables XXV

List of Tables

Table 1 List of Publications and Conference Contributions .................................................... 27

Table 2 Existing Research on Attribution Modeling—Essay 1 ............................................... 35

Table 3 Descriptive Statistics of the Data Sets—Essay 1 ........................................................ 37

Table 4 Definitions of Online Channels—Essay 1 .................................................................. 39

Table 5 Channel Distribution by Data Set—Essay 1 ............................................................... 40

Table 6 Removal Effect Calculation of Exemplary Markov Graph (in Figure 3)—Essay 1 ... 44

Table 7 Predictive Accuracy by Model and Data Set—Essay 1 .............................................. 48

Table 8 Removal effect: Average Standard Deviation as % of Average Removal Effect

(10-Fold Cross-Validation)—Essay 1........................................................................ 50

Table 9 Attribution Results in Comparison to Two Heuristic Models (in %)—Essay 1 ......... 52

Table 10 Estimation Results for Logit Model 1—Essay 1 ....................................................... 55

Table 11 Attribution Results for the Second Order Model by Data Set—Essay 1 .................. 57

Table 12 Descriptive Statistics of the Data Sets—Essay 2 ...................................................... 78

Table 13 Descriptive Statistics of the Journeys with Micro-Journeys—Essay 2 ..................... 80

Table 14 Descriptive Statistics of Micro-Journeys in Converting and Non-Converting

Journeys (with one Micro-Journey)—Essay 2 ........................................................... 83

Table 15 Estimation Results: The Micro-Journey as Predictor (Model 1)—Essay 2 .............. 94

Table 16 Estimation Results: The Micro-Journey Characteristics (Model 2)—Essay 2 ......... 99

Table 17 Estimation Results: The Effect of the Micro-Journey Characteristics on Direct

and Later Conversions (Model 3)—Essay 2 ............................................................ 102

Table 18 Robustness of the Results: The Micro-Journey as Predictor across Data Sets

(Model 1)—Essay 2 ................................................................................................. 104

Table 19 Robustness of the Results: The Micro-Journey Characteristics across Data Sets

(Model 2)—Essay 2 ................................................................................................. 107

List of Tables XXVI

Table 20 Robustness of the Results: The Effect of the Micro-Journey Characteristics on

Direct and Later Conversions across Data Sets (Model 3)—Essay 2 ...................... 112

Table 21 Categorization of Online Channels—Essay 3 ......................................................... 131

Table 22 Description of the Data Sets, Including Categories—Essay 3 ................................ 133

Table 23 Descriptive Statistics of the Data Sets, Including Channels—Essay 3 ................... 135

Table 24 Estimation Results: Present Channel Effects (Part 1/3)—Essay 3 .......................... 141

Table 25 Estimation Results: Past Channel Effects (Part 2/3)—Essay 3 ............................... 144

Table 26 Estimation Results: Channel Interactions (Part 3/3)—Essay 3 ............................... 147

Table 27 Estimation Results: Category Interactions (Table Part)—Essay 3 ......................... 150

Table 28 Correlation Matrix by Data Set—Essay 1 ............................................................... 203

Table 29 Estimation Results for Logit Model 2—Essay 1 ..................................................... 204

Table 30 Estimation Results: Full Table on Channel Effects (Part 1 – 3)—Essay 3 ............. 205

Table 31 Estimation Results: Full Table on Category Effects—Essay 3 ............................... 206

List of Abbreviations XXVII

List of Abbreviations

AIC Akaike Information Criterion

AUC Area under ROC Curve

B2C Business-to-Customer

BIC Bayesian Information Criterion

CIC Customer-Initiated Contact

DS Data Set

FIC Firm-Initiated Contact

HMM Hidden Markov Model

Info Informational

MJ Micro-Journey

Navi Navigational

ROC Receiver Operating Characteristic

SD Standard Deviation

SE Standard Error

SEA Search Engine Advertising

SEO Search Engine Optimization

SVAR Structural Vector Autoregression

TV Television

URL Uniform Resource Locator

WOM Word of Mouth

Introduction 1

1 Introduction

1.1 The Digital Era: Relevance, Development and Aspiration

From its first introduction in 1994 (D’angelo 2009), online advertising channels like search

engine marketing have become an essential part of many industries’ promotional mix (Yao and

Mela 2011), and advertisers today choose from a variety of online marketing vehicles,

including not only paid search and online display marketing, but also channels such as e-mail,

mobile, and social media advertising to reach consumers (Raman et al. 2012). Already in 2010,

three out of four CMOs and CSOs of leading European companies rate digital marketing and

sales topics as relevant or highly relevant for the future success of their business (McKinsey

and Company, Inc. 2010). The relevance of the topic, is also mirrored in online marketing

revenues, surpassing more than USD 160 billion globally in 2015 and accounting to more than

27 percent of the total media ad spending (eMarketer 2014a). In parallel, the ecommerce market

has gained momentum expected to surpass more than 1.2 billion online consumers (eMarketer

2013), boosting the market to a USD 1.7 trillion business and still growing at a two-digit

percentage rate annually (eMarketer 2014b).

Although first academic publications on the world wide web as advertising medium go

back to 1996 (Berthon, Pitt, and Watson 1996) and online advertising effectiveness has matured

to a subject of extensive research (Ha 2008; Kim and McMillan 2008), rapid market adaption

of digital media since 1994 has outpaced academic research (Klapdor et al. 2015). In particular,

most academic research focuses on selective issues and phenomenon often in a single channel

context such as search or display advertising. Comprehensive frameworks and studies in

multichannel settings are scarce, and often faced with data limitations such as aggregated or

simulated data sets (e.g., Breuer, Brettel, and Engelen 2011; Dalessandro et al. 2012).

Limitations in data availability and quality may be one reason for the divergent evolution of

research and practice to date. Notwithstanding, practitioners apply a plethora of channels to

Introduction 2

reach their customers (Raman et al. 2012) and, in succession, generating the demand for more

comprehensive marketing impact models based on individual-level user data (Hui, Fader, and

Bradlow 2009; Rust et al. 2004; Vakratsas and Ambler 1999).

Besides the proliferation of channels, the explosion of data has been identified as one of

the biggest challenges online marketers are faced with (IBM Institute for Business Value 2011).

In the following studies, we aim to approach both challenges by building on four large-scale,

real-world clickstream data sets, which allow us to investigate how users respond to

multichannel ad exposures on an individual level. In the following subsections, we approach

research on marketing effectiveness more broadly, and, thereafter, provide an overall view on

online studies building on clickstream data.

1.2 Research on Online Marketing Effectiveness

Overall research in online marketing is focused narrowly and touches upon various streams,

however, a universally valid theory does not exist to date. Nonetheless, a plentitude of scholar

have embraced the subject echoed in a myriad of publications. Regarding our research scope,

four fields are of utmost relevance delineating the following subsections: 1) Research on single

online marketing channels, 2) research based on multichannel data including attribution

modeling, 3) approaches on channel categorization relating to the purchase process, and 4)

recent research utilizing clickstream data.

1.2.1 Single online marketing channels

One main field of research on effectiveness in digital marketing relates to individual marketing

channels, especially to search including paid/sponsored (SEA) and unpaid/organic search

(SEO) as well as display marketing, but also to some degree covering social media, price

comparison, email/newsletter, affiliate and referrer marketing.

In paid search, according to Rutz and Bucklin (Rutz and Bucklin 2007) the advertisers’

remit is fourfold: 1) keyword selection, 2) bid management, 3) text ad design, and 4) landing

Introduction 3

page design. Extensive research has been conducted in selecting and optimizing keywords

(Ghose and Yang 2009; Jerath et al. 2011; Klapdor, von Wangenheim, and Schumann 2014;

Rusmevichientong and Williamson 2006; Rutz, Bucklin, and Sonnier 2012). With regards to

bid management, substantial research has been published focusing on the maximization of

marketing effectiveness within a given budget or alternate optimizing of budget decisions

(Abhishek and Hosanagar 2012; Chen, Liu, and Whinston 2009; Dar et al. 2009; Feldman et al.

2007; Katona and Sarvary 2010; Muthukrishnan, Pál, and Svitkina 2007). Specific research on

the design of text advertisements and landing pages is less extensive. Some cognate studies

evaluate the influence of the campaign’s content and position on its performance (Agarwal,

Hosanagar, and Smith 2011; Animesh, Viswanathan, and Agarwal 2011; Goldfarb and Tucker

2011a; Jerath et al. 2011; Rutz and Trusov 2011).

Furthermore, another stream of research processes the interrelation between diverse

factors in search. In paid search, Rutz and Bucklin (2011) propose a framework to capture

potential spillover effects from generic to branded search and show that spillover is asymmetric

with generic search positively affecting future branded search. Ghose and Yang (2010) modeled

cross-category purchases between product classes and see evidence of a considerable amount

of spillovers between the initial search and final purchase (intrinsic) and between cross-category

product purchases from different product classes (extrinsic). With regard to the interplay of paid

and organic search, some studies have been published (Rutz and Bucklin 2011; Xu, Chen, and

Whinston 2012; Yang and Ghose 2010). Ghose and Yang (2008) analyze how sponsored search

compares to organic search with respect to metrics such as conversion rates, order value and

profits and find positive, asymmetric effects between paid and organic search in their follow-

up study (Yang and Ghose 2010). Xu, Chen, and Whinston (2012) apply a game-theoretic

approach to analyze how the presence of organic search results as competing information

affects advertisers’ bidding behavior on sponsored advertisements and the equilibrium

Introduction 4

outcomes. Differentiated exposure in organic listings may improve social welfare, sales

diversity, consumer surplus and search engine benefits.

Turning to online display marketing, its value contribution is more complex to evaluate

since click-through rates to websites have gradually declined from 2% in 1995 to 0.3% in 2004

and are below 0.1% today (Chapman 2011; Cho and Cheon 2004; Fulgoni and Mörn 2009;

Hollis 2005; Sherman and Deighton 2001). Nonetheless, the relevance of display marketing for

advertisers remains undoubted making it revenue wise the second largest online marketing

channel after search (PriceWaterhouseCoopers 2014). Advancements in targeting technologies

allow for novel advertising measures such as retargeted display advertisements and give display

advertisements a stronger performance-oriented component (Lambrecht and Tucker 2013).

That being said, and the fact that the vast majority of users avoids clicking on (non-targeted)

display advertisements, so affecting them only by the mere ad impression (Chatterjee, Hoffman,

and Novak 2003; Drèze and Hussherr 2003), have led to two major research directions to

quantify the effectiveness of display advertising: 1) Indirect effects such as brand awareness,

and 2) direct effects such as immediate consumer response covering purchase intent (Hollis

2005; Qiu and Malthouse 2009). These two perspectives are rather complementary than

contradictory (Hollis 2005), since indirect effects may influence future consumer response.

Exhibiting indirect effects, multiple studies evaluate the changes in awareness creation covering

brand awareness, brand attitudes and purchase intentions as function of ad exposure (Briggs

and Hollis 1997; Cho, Lee, and Tharp 2001; Dahlen 2001; Drèze and Hussherr 2003; Gallagher,

Foster, and Parsons 2001). For instance, Briggs and Hollis (1997) recognized that online

displays raise advertising awareness and brand perceptions even though the user does not follow

them in a responsive click. Furthermore, academic research examines direct consumer effects

using click-through rates and website traffic (Chatterjee, Hoffman, and Novak 2003; Ilfeld and

Winer 2002; Sherman and Deighton 2001) or changes on purchase related effects such as

purchase intent (Braun and Moe 2013; Fulgoni and Mörn 2009; Goldfarb and Tucker 2011b;

Introduction 5

Lambrecht and Tucker 2013; Lewis and Reiley 2014; Manchanda et al. 2006; Qiu and

Malthouse 2009; Rutz and Bucklin 2012). For instance, Fulgoni and Morn (2009) find evidence

that sole exposure to an online advertisement positively influences online and offline sales as

well as additional metrics, for instance, brand site traffic, even with no clicks or minimal click

rates confirming the relevance of display marketing exposure. Lewis and Reiley (2014) detect

positive causal effects of online display advertising on a major retailer’s offline sales in a

controlled field experiment. Furthermore, display advertising may exhibit synergies with search

resulting in a larger effects size of combined display and search campaigns compared to two

separated channel homogenous campaigns. Technological advancements in digital marketing,

like dynamic retargeting or behavioral targeting, enable advertisers to tailor the ad’s content

based on an individual previous browsing behavior. As one of the first, Beales (2011) provides

a systematic empirical assessment on behavioral targeting measuring conversion rates and

revenues and identify behavioral targeting as more successful than non-targeted campaigns.

Using a proportional hazards model, Lambrecht and Tucker (2013) analyze retargeted display

advertisements and find, surprisingly, that on average less specific, generic retargeted content

is more effective than more specific retargeted marketing messages.

Research on other online channels is less extensive. For instance, some research has

been conducted on targeting and quasi-social networks (Provost et al. 2009), personalized

advertising and privacy controls on social networks (Tucker 2014), the effects of word-of-

mouth (WOM) and traditional marketing analyzing a social networking site (Trusov, Bucklin,

and Pauwels 2009), and determining influential social network users (Trusov, Bodapati, and

Bucklin 2010). Kumar et al. (2013) analyze the success of social media marketing efforts and

link WOM to actual sales illustrating that social media may support in generating sales,

increasing return-on-investment (ROI), and spread brand awareness. Turning to price

comparison websites, Baye et al. (2009) evaluate determinants of clicks received and price

dispersion. On email/newsletter marketing, Morimoto and Chang (2006) examine in a survey

Introduction 6

based study consumers’ attitudes toward unsolicited commercial online newsletters and postal

direct mails and show that unsolicited newsletters are perceived as more intrusive as postal

mail. In another study, Tezinde, Smith, and Murphy (2002) explore factors that influence users’

granting permission with regard to online newsletter marketing. In the context of permission-

based email marketing, Ansari and Mela (2003) analyze how e-customization at the individual

level affects email click-throughs and show positive effects of content-targeted marketing

approaches. One aspect they leave uncovered is the desired newsletter sending frequency. Due

to quasi-zero marginal cost, a well-balanced email communication against wear-out effects is

inevitable, a topic Bonfrer and Drèze (2009) firstly touched upon. Moreover, some studies have

investigated affiliate marketing. In a case study, Duffy (2005) analyzed its impact on

ecommerce companies focusing on success factors influencing the long-term relationship

between the affiliate and the vocal firm. Further studies concentrate on the efficiencies in

selecting online affiliate programs (Edelman and Brandi 2015; Papatla and Bhatnagar 2002)

and the definition of referral fees in affiliate marketing (Libai, Biyalogorsky, and Gerstner

2003). In a study on referrer/referral marketing, Guo (2012) develops an analytical model to

evaluate business potential and support decision making in incorporating online referral

marketing programs.

1.2.2 Multichannel Research and Attribution

More recently a body of literature covering multiple online channels has evolved, but not yet

reaching the breath and plurality as research on single channels. Based on aggregated data,

Breuer, Brettel and Engelen (2011) are the first to analyze long-term and interaction effects in

an online multichannel setting covering email, display, and price comparison advertising.

Surprisingly, they cannot confirm between-channel interaction effects for multichannel online

advertising, but explicitly call for a more detailed analysis of these effects.

Introduction 7

Fulgoni and Mörn (2009) reveal positive synergetic effects on retail sales with

campaigns combining search and display advertising in comparison to two separate single

channel campaigns. Applying a multivariate time series model, Kireyev, Pauwels, and Gupta

(2013) model interactions between paid search and display advertising and identify dynamics

that improve search clicks, search conversions and ROI over time.

From the offline world, we know of synergies between different advertising channels.

Jagpal (1981) was the first to present empirical evidence of synergies in multichannel

advertising, analyzing print and radio marketing, however, this study does not cover carryover

effects. In a controlled laboratory experiment, Edell and Keller (Edell and Keller 1989) find

synergies for television and print advertising in cross-media sequences. More recently, Naik

and Raman (2003) evaluate print-television synergies using an integrated marketing

communication model and suggest two effects of singular channel marketing: sales generation

and enhancement of other channels. In another study, Naik and Peters (2009) investigate within-

media and cross-media synergetic effects conveying offline and online channels—search and

display—and demonstrate how these synergies lift and optimize media spending. Finally,

Wiesel, Pauwels and Arts (2011) propose a conceptual framework to better allocate marketing

resources across media activities and channels, finding evidence of many bidirectional cross-

channel effects between offline and online channels. Research that combines online and offline

interaction effects is faced with substantial shortcomings in data availability compared to

research focused on online advertising purely, as it is difficult for advertisers to track individual

exposures to TV, print, or radio advertising to date. A notable exception is a study by Tellis et

al. (2005) evaluating micro-effects of TV advertising. They provide a model that helps to

understand which specific ad works on which TV channel, when, and how often. Although the

relevance of offline or hybrid studies seems less obvious for our research objectives, it provides

an adequate foundation for questions that relate to the effectiveness of multichannel user

preferences in online.

Introduction 8

Moreover, a research stream on attribution modeling is gaining momentum (Abhishek,

Fader, and Hosanagar 2012; Berman 2015; Dalessandro et al. 2012; Haan, Wiesel, and Pauwels

2013; Kireyev, Pauwels, and Gupta 2013; Li and Kannan 2014; Shao and Li 2011; Xu, Duan,

and Whinston 2014). Attribution modeling is focuses particularly on measuring the

effectiveness of individual channels in multichannel settings (Neslin and Shankar 2009) and

involves finding ways to measure "the partial value of each interactive marketing contact that

contributed to a desired outcome" (Osur 2012, p.3). From a publisher’s perspective, Jordan et

al. (2011) formulate an allocation mechanism based on multiple attribution that may support in

scheduling and pricing advertisements. Shao and Li (2011) introduce two attribution

approaches, a simple probabilistic model and a bagged logistic regression model. Building on

their work, Dalessandro et al. (2012) propose a causally motivated attribution methodology

based on cooperative game theory. Abhishek, Fader, and Hosanagar (2012) develop a dynamic

hidden Markov model (HMM) that captures a consumer’s deliberation process along the typical

stages of the purchase funnel. Li and Kannan (2014) propose a Bayesian attribution model and

measure short-term as well as long-term carryover (channel homogenous) and spillover

(channel heterogenous) effects of multiple channels using individual conversion path data.

Based on aggregate data by Kireyev, Pauwels, and Gupta (2013) analyzes attribution dynamics

for display and search advertising. Haan, Wiesel, and Pauwels (2013) propose a structural

vector autoregressive model, also based on aggregate data, to determine the effectiveness of

different online advertising channels. Applying a mutually exciting point process model, Xu,

Duan, and Whinston (2014) compute average conversion probabilities for diverse online

channels and reveals that the effect of display advertisements is underestimated compared to

search advertisements. Finally, looking at the dynamics between advertisers and publishers,

Berman (2015) proposes a Shapley value approach to elucidate the impact of different incentive

schemes on publishers' propensity to show advertisements and the consequential profits of

advertisers. Not from a modeling perspective, Tucker (2012) examines welfare consequences

Introduction 9

implied by the usage of attribution technologies and finds evidence for more conversions at a

lower price due to the ability to systematically substitute toward selected campaigns across

advertising platforms.

1.2.3 The Purchase Decision Process and Channel Categories

The basic idea of the procedural nature of purchase decisions is that when making a choice,

consumers follow at least a two-staged process (Hauser and Wernerfelt 1990). The concept

roots back to Howard (1963) and Campbell (1969), who initially introduced the concept of an

evoked set describing a set of considered brands in consumer responses. Wright and Barbour

(1977) shaped the term consideration set to define that brands a consumer will consider.

Although definitional clarity is clouded by the terminological ambiguity, the idea has proven

valuable eliciting numerous studies (Hauser and Wernerfelt 1990; Roberts and Lattin 1991,

1997). Albeit exact denotations diverge, research consent thereto that consumers confronted

with a multitude of brand narrow down their potential choice to a relevant set called the

consideration set (Alba and Chattopadhyay 1985; for a comparison of definitional differences

see Brown and Wildt 1992), which in consequence is subject for the purchase (Howard and

Sheth 1969; Parkinson and Reilly 1979). The concept of choice sets has been vastly applied in

marketing research (for a review see Roberts and Lattin 1997; Shocker et al. 1991). Yet, in an

online (multichannel) context, research relating to choice set theory remains scattered (Yadav

and Pavlou 2014). Some research has been conducted on the multi-staged nature of customer

decision processes within online web shops (Häubl and Trifts 2000; Moe 2006; Wu and

Rangaswamy 2003). Furthermore, Abhishek, Fader, and Hosanagar (2012) map data from an

online campaign launch onto the consumer’s deliberation process along the typical stages of

the conversion funnel, which they call disengaged, active, engaged, and conversion, and apply

a Hidden Markov Model to analyze how channels impact on stage transitions. More practice

oriented publications relate to the concept of customer decision processes (Court et al. 2009;

Edelman 2010; Mulpuru 2011). In search, analogous concepts have been mentioned as buying

Introduction 10

funnel (Jansen and Schuster 2011) or as purchasing funnel (Jordan et al. 2011). Although

dissent on exact definitions prevails, the general conception of choice sets may be exemplified

based on the widely-postulated definition by Shocker et al. (1991) (Figure 1).

Figure 1

Dynamic Choice Process Based on Shocker et al. (1991)

Users are not forming their purchase decision on the universal set, instead, they are only

aware of a certain range of alternatives that may satisfy their needs, forming the awareness set.

A subset of the awareness set, the consideration set, includes the alternatives which the user

seriously considers in making a purchase decision (Shocker et al. 1991; Wu and Rangaswamy

2003). There is ample evidence that these sets form and evolve dynamically, as users may

remove and add alternatives based on the processing of information they either already have or

they are confronted with externally (Hauser and Wernerfelt 1990; Shocker et al. 1991; Spiggle

and Sewall 1987; Wu and Rangaswamy 2003). Moreover, users may formulate a further subset,

the choice set, in which the remainder alternatives are compared against each other, before

conducting the final (purchase) choice maximizing their utility level. Albeit observational data

alone does not suffice for ascribing stages within the choice set process (Shocker et al. 1991),

Introduction 11

online clickstreams may allow for ratiocinations mapped onto (multi)channel usage along the

underlying purchase decision process.

Another field of online marketing research may be consolidated under categorization

approaches, assigning online channels to channel categories and aiming for a better

interpretation of user behavior. As the number of potential channel interactions increases

exponentially with the number of channels applied, also known as the curse of dimensionality

(Bellman 1961), simplifications such as channel categorizations have become inevitable to

examine certain effects, for instance, the interplay of channels. Although categorization

approaches do not necessarily link back to the theory of choice sets, the users’ informational

needs during this process are subject for periodical change (Payne, Bettman, and Johnson 1988),

thus, making them suitable as proxy to measure progression within customer decision

processes. So far, five categorization approaches have been applied in online marketing

research: Contact origin, browsing goal, degree of content integration, degree of

personalization, and branded vs. generic search contacts.

Contact origin describes whether the customer-firm interaction was initiated by the firm

or by the customer. A novelty in online advertising is that customer-firm contacts are often

initiated or “pulled” by the customer (Shankar and Malthouse 2007), whereas in traditional

media marketing activities are often “pushed” by the firm. Customer-initiated contacts (CIC)

convey paid and unpaid search, clicks on price comparison websites, as well as direct type-ins

in the address bar of the browser. In contrast, firm-initiated channels include for instance display

advertising, retargeting or email newsletters. Prior research has argued that customer-initiated

channels (CIC) are more effective than firm-initiated contacts (FIC) with regard to (predicting)

sales (Haan, Wiesel, and Pauwels 2013; Wiesel, Pauwels, and Arts 2011), as they are perceived

as less intrusive and comprise direct responses based on the customer’s actions (Shankar and

Malthouse 2007). Switches between channel categories may be interpreted as progression in

Introduction 12

the purchase funnel, especially, switches from firm-initiated to customer-initiated contacts, as

the user proactively antecedes in his or her browsing session after being exposed to firm-

initiated marketing measures. In addition, positive responses (clicks) on firm-initiated contacts

preceding customer-initiated contacts may indicate relevance or users’ being open for

marketing messages.

The browsing goal routes back to a classification developed in information retrieval

research and aims for anticipating the user’s underlying browsing intention by categorizing

channel exposures into navigational and informational contacts (Broder 2002; Broder et al.

2007; Klapdor et al. 2015; Rose and Levinson 2004). Survey based studies partly refer to a

tripartite taxonomy further differentiating transactional contacts (Broder 2002; Rose and

Levinson 2004), however, a twofold classification logic seems favorable in settings based on

observational data to reduce fuzziness and misclassifications (Shocker et al. 1991). The

browsing goal is seen as informational, if the user aims “to learn something by reading [..] web

pages” (Rose and Levinson 2004, p.15). In contrast, it is classified as navigational, if the user

intentionally accesses a certain web page (Broder 2002). Thus, direct type-ins and email

marketing are categorized as navigational, user interactions via display, price comparison or

affiliates are considered as informational contacts. Even though this classification leaves

considerable elasticity for interpretation, Klapdor et al. (2015) shows a positive relationship

between switches from informational to navigational contacts, arguing that users may have

narrowed down his or her choice set on the path to purchase.

More recently, Haan, Wiesel, and Pauwels (2013) suggest a taxonomy that classifies

channel exposures according to their level of integration into the websites original content. In

consequence, content-integrated marketing activities convey channels that are embedded as

integral part of a website, for example, listings on price comparison websites or blogs (Breuer,

Brettel, and Engelen 2011; Zhu and Tan 2007), or unpaid search results. In contrary, content-

Introduction 13

separated advertisements are rather tangentially related to the content and format of the website

such as display or paid. In a combined category approach, Haan, Wiesel, and Pauwels (2013)

show that content-integrated customer-initiated contacts (CICs) are more effective than

content-separated customer-initiated contacts (CICs) in driving purchase funnel progression, as

they are seen as less intrusive.

Another categorization approach refers to the degree of personalization of marketing

messages. Recent technological advancements in tracking techniques improve targeting

possibilities and, subsequently, catalyze novel options to personalize advertising messages

(Pavlou and Stewart 2000; Varadarajan and Yadav 2009). Personalized marketing messages

base on the user’s prior browsing traits or his or her disclosed characteristics, whereas non-

personalized marketing messages are broadcasted to a universal audience. Thus, personalized

marketing channels comprise retargeted display advertisements (Lambrecht and Tucker 2013),

as well as, paid and unpaid search results, as they originate from user-specific entered search

terms (Ghose and Yang 2009). Unexpectedly, in a study on retargeted banner advertisements,

Lambrecht and Tucker (2013) show that generic retargeted display advertisements are more

effective than specific retargeted banners, indicating an underlying psychological bias between

relevance and obtrusiveness.

Referring to search engine marketing, scholars have classified search requests into

branded and generic (non-branded) search terms. In this context, branded keywords include the

brand name of the corresponding online shop. In a multichannel context, this taxonomy has not

yet been applied, making it congruent with respect to the channel classification used in the

browsing goal taxonomy. Not surprisingly, prior research suggests that branded search terms

are more effective than generic search terms (Jansen, Sobel, and Zhang 2011). Furthermore,

based on aggregated data, Rutz and Bucklin (2011) find unidirectional spillover effects from

generic search activities affecting future branded search activities.

Introduction 14

1.2.4 Research on Clickstream Data

Electronic records of the online browsing activities of individual users has come to be known

as clickstream data (Bucklin and Sismeiro 2009), at times also called path data or journey data.

In consequence, they allow for tracing back which navigation path he or she has taken, and

uncover choices along their way. Thereby clickstreams may include individual browsing

behavior within one particular website as well as across several websites. Although a unifying

definition of the collective of tracked parameters and the technical data source is non-existent,

some characteristics are commonly applied. The major unit in recording clickstreams is the

page view which tracks whether a specific user was exposed to a given website (Bucklin and

Sismeiro 2009). In other words, in the event that a unique user clicks on an advertising message,

the link embedded into this ad forwards the user to the corresponding website, thus, logging the

respective data typically including the source of the click (e.g., website, online channel,

campaign), an exact time stamp, and a unique ID associated with the user. Additionally, a

distinction is made in site-centric and user-centric clickstream data (Bucklin and Sismeiro

2009). Site-centric path data may include detailed information of the site-user touchpoints

whenever a user enters a given website or may provide meticulous records of what a user does

when navigating the particular website. However, they fall short regarding the users’ activities

on other, for instance, competitors’ Web pages. Technologically, they are often generated using

server log files, or cookie tracking. Being subject for deletion, cookie generated data are hardly

impeccable, however, they persist as a widely-accepted industry standard and research has

verified that their shortcomings do not impose substantial complications (Drèze and Zufryden

1998; Flosi, Fulgoni, and Vollman 2013; Rutz, Trusov, and Bucklin 2011; Tucker 2012).

Alternate sources of clickstream data are internet analytics providers (e.g., ComScore), the

user’s internet service provider (ISP), Java applets installed to the user’s device. Thereby, the

recording captures all Universal Resource Locators (URLs) requests while the user navigates

the Web across all Web pages, thus, making it user-centric. Although tracking of all visited

Introduction 15

websites entails prospective modeling and managerial advantages (Padmanabhan, Zheng, and

Kimbrough 2001), potential drawbacks in data sampling might arise, especially, when

evaluating a particular website with comparably limited Web traffic or, even worse, when

analyzing customers’ purchase behavior of a dedicated online shop with by nature low visit-

conversion ratios. Furthermore, relating to user-centric as well as site-centric clickstreams, data

tracking occurs at the individual device level neither capturing potential multi-device usage nor

multiple individuals using one single device. Only experimental or laboratory data may rule out

these shortcomings with certainty and was discussed literature (Birnbaum 1999; Johnson 2001;

Mandel and Johnson 2002; McGraw, Tew, and Williams 2000; Moe 2006). Recent

technological tracking advancements also allow for tracing the mere exposure of marketing

messages even if the user avoids clicking, however, as of today no study has been published

dedicatedly analyzing this type of click and impression based path data.

Clickstream data not only enables scholars and practitioners to further elaborate on

renowned phenomena, but also permits to shed light into novel challenges routing from the

Internet (Bucklin and Sismeiro 2009). For instance, they contain more pertinent detail than

scanner panel data applied in the progression of choice set models early on (Bucklin et al. 2002).

Due to technological advancements, research based on naturally collected clicks, such as our

research, has recently gained traction.

One subfield of clickstream research addresses on-site browsing behavior. Huberman et

al. (1998) was one of the first to analyze clickstream data assuming that viewing of a website

is a function of the value derived from it and the cost implied. They illustrate that this cost-

benefit ratio is prone to better predict individual website usage and browsing behavior,

however, they ignore other influencing factors such as marketing exposure. Based on site-

centric data, Bucklin and Sismeiro (2003) showed users’ behavioral adaption and learning

effects that strive from repeated website visits. In particular, users tend to brows less subpages,

Introduction 16

however, with stable page view durations, overall resulting in a reduction of session duration.

Building on that, Danaher, Mullarkey, and Essegaier (2006) examined factors affecting website

visit duration in a user-centric, cross-domain study and finds that visit duration is mostly

situational, and only tangentially related to fundamental website characteristics. Applying a

cluster analysis on site-centric data, Moe (2003) classified visitors of an ecommerce retailer

along their specific browsing behavior and (sub)page visits and found four different browsing

strategies: directed buying, search/deliberation, hedonic browsers, and knowledge-building.

Montgomery et al. (2004) conducted another study on within-site browsing patterns. Unlike

Moe (2003), they utilize user-centric data to model transition choices between different page-

types of visitors navigating a given ecommerce bookseller and indicate that visitors may change

their goals while browsing.

Several studies on marketing exposure building on clickstream data have been

published. Chatterjee, Hoffman, and Novak (2003) build on site-centric clickstreams and apply

a binary logit model to investigate users’ click proneness on banner exposures. Their results

show that click rates of less click-prone users increase when exposed to banner advertisements

repeatedly and, in addition, indicate that click rates are independent from the position of banner

advertisements along their navigation path. In a study, partly based on clickstream data, Ilfeld

and Winer (2002) analyze sources of Web traffic and report that online advertising not only

impacts on website traffic, but also raises website awareness and brand equity. Danaher (2007)

introduced a stochastic approach to model and decompose user-centric page views across

diverse websites into frequency and reach. Using a hazard model approach, Manchanda et al.

(2006) analyze site-centric unique machine-level data of banner exposures that, for all tracked

journeys, ultimately resulted in a purchase decision. They investigated determinants affecting

time to purchase and showed that the total number of ad exposures as well as the number of

different website and page views in a given week are accelerating recurring purchases of

existing customers. In the context of retargeted banner advertisements, Lambrecht and Tucker

Introduction 17

(2013) applied a Cox proportional hazards model to illustrate how ad message specificity

influences purchase likelihood. Display advertising may not only induce purchase decisions,

but also influence browsing behavior. Building on site-centric individual-level data, Rutz and

Bucklin (2012) studied how display advertisements affect subsequent browsing behavior on an

automotive website and found segmented response effects of users—positive, negative, and

zero. In a multichannel context, Nottorf (2014) applied a binary logit with a Bayesian mixture

of normals approach to model a site-centric data set from a financial service provider and finds

differences in users’ click proneness of repeated exposures across varying types of display

advertisements as well as positive interactions display and search advertising. Furthermore,

several studies on purchase conversion or on choice set progression have been issued. Moe and

Fader (2004a; 2004b) developed a stochastic approach to link visit frequency and purchase

propensity. In a more detailed manner, Sismeiro and Bucklin (2004) delineate site-centric

individual-user data into three successively connected tasks of the on-site purchase process (i.e.,

completion of product configuration, input of personal data, order confirmation and payment)

and analyzed how behavioral user patterns impact on these three steps. The results revealed

significant sign reversals for various influencing parameters, for instance, exiting and returning

to the given website is positively associated with inputting personal data, however, negatively

associated with confirming the purchase by entering the payment information. Regarding

purchase funnel progression, Wu and Rangaswamy (2003) used site-centric data to investigate

on the formation of consideration sets. In a two-staged model, Moe (2006) connected product

viewings and product category choice and found predictive superiority of two-stage models

compared to a single-stage approach. More recently, Abhishek, Fader, and Hosanagar (2012)

analyzed stage transitions in a multichannel attribution approach based on a Hidden Markov

Model (HMM). Setting a stronger focus on multichannel attribution, various studies have

exemplified the usage of clickstream data (e.g., Haan, Wiesel, and Pauwels 2013; Li and

Kannan 2014; Xu, Duan, and Whinston 2014).

Introduction 18

For additional discussions on clickstream related literature and modeling prospects

please refer to Chatterjee Hoffman, and Novak (2003), Hui, Fader and Bradlow (2009), and

Bucklin and Sismeiro (2009).

1.3 Shortcomings and Research Scope

Overall, research in online marketing is relatively narrow focused leaving many facets

uncovered. On a consumer level, little is known on how the interplay of multiple advertising

channels influences purchase propensities. Moreover, advertisers struggle to link users'

browsing behavior to their expected conversion rates—especially in a multichannel

environment. Selective findings from prior research are weak in theory. They are mostly

induced from a stand-alone viewpoint or merely derived from the data, instead of anchored in

pre-known and well-established theoretical frameworks or underlying psychological constructs,

such as flow theory. The call for marketing impact models based on individual-level, single-

source data, aiming to identify optimal levels of marketing expenditures, or improving forecasts

of individual-level user click and conversion proneness remain insufficiently answered (Rust et

al. 2004). From a methodological perspective, little is known on how to amply and context-

specific analyze large-scale, consumer-level path (Bucklin and Sismeiro 2009; Chatterjee,

Hoffman, and Novak 2003; Hui, Fader, and Bradlow 2009). The adaption of scientifically rigor

marketing impact models into practical applicability being dissipated prevails challenging

(Leeflang and Wittink 2000; Lehmann, Mcalister, and Staelin 2011; Little 1970, 1979, 2004a;

Lodish 2001; Reibstein, Day, and Wind 2009). A simultaneous consideration of the above

outlined research areas, or leastwise amalgamating some of their circumferences, could help to

plot a more comprehensive picture of online marketing effectiveness. We approach these

research areas from plural directions. In doing so, we apply four detailed, individual-machine

level and site-centric clickstream data sets from three different industries not only including

converting users but also covering users that never performed a purchase: First, to investigate

on the so-called attribution challenge, which online channel contributes to what degree to the

Introduction 19

vocal firms’ marketing success. Second, routing from flow theory (e.g., Csikszentmihalyi and

Csikszentmihalyi 1988; Csikszentmihalyi 1975, 1977; Novak, Hoffman, and Yung 2000;

Webster, Trevino, and Ryan 1993), we suggest a novel concept of time-related browsing

respectively clicking patterns occurring within customer clickstreams, which we call micro-

journeys. Third, we examine interaction effects of user channel preferences (homogenous and

heterogeneous) closely reflecting a real-world multichannel context. All studies share the

detailed, real-world, and high qualitative data foundation, are situated in a multichannel context,

and are designed to satisfy scientific rigor as well as practical relevance. From a methodological

viewpoint, we propose two different, practically feasible methods. First, in a Markovian model

approach, we represent clickstream as Markov walk to evaluate channel relevance, second, we

leverage an in the online context less acknowledged, but in a broader scientific context well-

accepted proportional hazards model approach, to, inter alia, well-reflect the chronological

nature of clickstream data and purchase timing (Seetharaman and Chintagunta 2003).

Consequently, we further capture the notion from Day (2011) claiming a diverging capability

gap to handle the factually growing accessibility of complex data sets and respond to the recent

call from the Marketing Science Institute (2014) to give priority to the development of

analytical frameworks for data-rich environments. Thus, our three essays, seamlessly

complement existing research on online marketing effectiveness in a still evolving,

multichannel world.

1.3.1 Essay 1 – Mapping the Customer Journey: Lessons Learned from Graph-

Based Online Attribution Modeling

The proliferation of online marketing instruments as well as the growing prevalence of data

have made online marketing multifarious and increasingly complex (Leeflang et al. 2014;

Raman et al. 2012). Novel tracking solutions enable advertisers to collect individual-level

clickstreams capturing the exact browsing paths of each user (device) browsing their website

generating new enriching opportunities for their businesses (Bucklin and Sismeiro 2009). Most

Introduction 20

advertisers, however, struggle in benefitting from these opportunities, and still rely on simple

heuristics when it comes to attributing the correct credit to each online marketing channel,

including linear attribution, splitting the value contribution evenly across channels, or, even

more frugal, last click or first click wins approaches (Econsultancy 2012; The CMO Club &

Visual IQ 2014). While most concepts applied to practice only model successful user clicks

(e.g., last click wins) or, click paths, skipping user journeys that do not convert, more complex

attribution mechanism are publicly inaccessible and irreproducible (Dalessandro et al. 2012;

Tucker 2012). Albeit its high relevance, research on attribution has only recently gained

momentum (Abhishek, Fader, and Hosanagar 2012; Berman 2015; Dalessandro et al. 2012;

Haan, Wiesel, and Pauwels 2013; Kireyev, Pauwels, and Gupta 2013; Li and Kannan 2014;

Shao and Li 2011; Xu, Duan, and Whinston 2014), providing valuable approaches on “how to

measure channel effectiveness” in multichannel settings. Yet, the creation of generalizable

insights that include industry-specific as well as cross-industry findings on the effectiveness of

individual online channels, and on their interplay, in multichannel environments remains

untouched. Furthermore, scientific approaches developed to the attribution channels seem to

not disseminate sufficiently into practice so far, potentially, due to the fact that marketing

models demand more than scientific rigor (Little 2004a; 2000b; Lodish 2001; Wangenheim and

Wübben 2008).

In order to shorten this gap and to enrich previous research on attribution, we formulate

a novel, graph-based variant of attribution framework that represents the full set of customer

journeys (converting and non-converting) as first- and higher-order Markov walks, and debate

it in the context of two widely applied attribution heuristics, explicitly first click and last click

wins, as well as two logit benchmark models. Applying this approach to four, empirical, large

data sets covering three different industries, allows to compute the value contributions of

individual channels, channel interplay, and to postulate empirical generalizations relevant

across industries. Our results show substantial differences to the comparison models, yet,

Introduction 21

complement as well as generalize previous research based on one single data set (Li and Kannan

2014). For example, we observe that users, on their path to purchase, show idiosyncratic

channel preferences (carryovers) and derive meaningful interrelations between channels—both,

within and across channel categories (spillover). Forming an adequate compromise between

practical applicability and data availability, whilst keeping uppermost scientific standards, this

study provides valuable implications on “what we learn from attribution” and contributes to

solving the attribution puzzle. The research questions include:

1) What is an adequate attribution framework to determine the correct value

contribution of each online channel and their interplay, and how does it compare

to well-established mechanisms?

2) Based on that, what results and implications apply generalized across industries,

rather than in an industry context, including both, novel findings on channel

attribution as well as the interplay of online channels?

1.3.2 Essay 2 – Patterns that Matter: How Browsing Click Patterns as Micro-

Journeys Influence Customer Conversions

The evolution in online advertising culminates in two inter-connected, impending progresses:

The advancement of (tracking) technologies generating large quantities of information and,

given the proliferation of online channels, more complex user browsing and, thus, online

purchasing behavior (Bucklin and Sismeiro 2009; Leeflang et al. 2014). While the former,

mostly concealed from the user, enables advertisers to further ameliorate promotional content

and placement of their marketing messages (e.g., via real-time bidding), the latter, from a user

behavioral perspective, allows for multifarious browsing pattern, entailing an increased

complexity in advertiser-user interactions. In order to benefit from these developments as

marketer, it is pivotal to lay the foundations to extract implications manifested in the user

browsing traits recorded or accessible by the advertiser.

Introduction 22

This is exactly where our study ties up, to better uncover hidden purchase intention of

commonly unknown users navigating the Web and visiting focal advertisers’ websites. Despite

its relevance, prior research has set a strong focus on analyzing single online channel, especially

on search (e.g., Rutz, Bucklin, and Sonnier 2012) and display advertising (e.g., Braun and Moe

2013), to some degree on interactions within or across two channels (e.g., Yang and Ghose

2010) or, more general, on channel categories (e.g., Broder 2002), and, more scarcely, on a

multichannel contextual topics (e.g., Li and Kannan 2014). While all studies help to shed light

into relevant and interesting phenomenon, online (multichannel) research often builds on

aggregated data (Breuer, Brettel, and Engelen 2011), disallowing to interpret individual user

behavior or link it to well-established psychological constructs. Furthermore, if they apply

single-sourced clickstream data, none of them incorporate time-distances between single clicks,

instead, clicks are treated as equally distributed.

In response, we derive a novel concept to uncover hidden purchase intentions based on

the users’ click-timing. Building on the flow theory (Csikszentmihalyi and Csikszentmihalyi

1988; Csikszentmihalyi 1977; Novak, Hoffman, and Yung 2000), we capture “focused

attention” as users' browsing pattern, by conceptualizing intense browsing sessions, measured

by brief successive single clicks, into a novel concept, which we call micro-journey. We employ

a proportional hazards model on four large-scale individual user level data sets and exhibit that

our novel concept, the micro-journey, is well-suited to better predict user conversions. Users

utilizing micro-journeys as browsing patters are more inclined to conclude in a purchase event,

either directly after the micro-journey or procrastinated. Further characterizing the micro-

journeys reveals meaningful browsing patterns that help to better understand customers’

decision making progress. For instance, its relative position within the overall journey as well

as navigational contacts within the micro-journey are relevant predictors for converting users.

Introduction 23

The conceptualization of user behavior and its translation into conversion propensity is

highly relevant for theory and practice alike, especially in vigorous environments. By

introducing a theoretically grounded, behavioral click pattern, we enable advertisers to extract

user behavior on their path to conversion to dynamically improve their marketing measures and

correspond to the following research questions:

3) Are time-related click patterns, so-called micro-journeys, a suitable predictor to

better detect user journeys that are more likely to conclude in a purchase event?

4) In a more detailed manner, what are the characteristics attached to this user

browsing pattern, the micro-journey, that well indicate user journeys likely to

convert?

1.3.3 Essay 3 – Channels and Categories: User Browsing Preferences on the Path

to Purchase

Online marketing has become increasingly complex. Albeit the dyad of the proliferation of

online marketing channels and the advancements in tracking technologies summiting in vast

data quantities open up novel and enriching possibilities for marketers (Leeflang et al. 2014),

manifold uncertainties related to the interplay of online channels and the users’ browsing

behavior conflated with their subliminal (purchase) intentions arise or remain enigmatic. One

of these uncertainties is the actual channel and channel group preference of the user on the path

to purchase: Do users exhibit channel homogeneous or channel heterogeneous browsing

histories whenever they are inclined to form a purchase decision?

To date, research has approached this riddle in a two-fold manner. First, literature has

made use of the inheritance of preceded offline research and research on offline/online

synergies (e.g., Edell and Keller 1989; Jagpal 1981; Naik and Raman 2003). Multiple marketing

sources are associated with an increased credibility, processing motivation, which, in

consequence, results in enhanced brand recognition and purchase intent (MacInnis and Jaworski

Introduction 24

1989; Petty and Cacioppo 1986). Second, in a pure online context, research has examined

synergetic effects, however, emphasizing narrowly selected channels and their interplay

including search (e.g., Rutz and Bucklin 2011; Yang and Ghose 2010) or search and display

(e.g., Kireyev, Pauwels, and Gupta 2013). These studies exemplify the relevance of research

around user preferences and intermedia synergies, yet leaving multichannel aspects in an online

interplay mainly shadowed.

A related challenge in investigating channel preferences is associated with the raise of

channel diversity, which further provokes a phenomenon known as the “curse of

dimensionality” (Bellman 1961). It provokes novel complexities in analyzing data set, as data

points become increasingly sparse with the rise of dimensions (e.g., variables). Research has

encountered this phenomenon by categorizing channels resembling in character into

corresponding channel taxonomies (e.g., Broder 2002; Rose and Levinson 2004). Though

multiple taxonomies exist, research so far has focused on examining distinct topics merely in

isolation.

Thus, in this study we investigate these challenges in a three-fold and comprehensive

manner by linking the corresponding influencers to purchase propensity building on a

proportional hazards model: First, we measure the effect of present online channel exposures

and past channel stock on purchase propensity. Second, we analyze interactions between

present and past homogeneous and heterogeneous channel exposures to derive channel

preferences and, third, we comprehensively implement interactions of categories

simultaneously considering four relevant category approaches, namely, the contact origin, the

browsing goal, the degree of content integration, and the degree of personalization, within one

model to draw conclusions on the users’ channel group preferences. Based on four large-scale

individual-level clickstream data sets, we find novel and interesting results, such as strong

idiosyncratic channel preference of users across data sets and industries, controversial to prior,

Introduction 25

offline and hybrid (offline/online) research. By investigating the following two overarching

research questions, we derive findings that contribute to marketing effectiveness research and

can support practitioners in shaping their marketing activities in a dynamic multichannel

domain:

5) What are the users’ actual channel preferences while forming their purchase

decision online—do they exhibit homogeneous (idiosyncratic) or heterogeneous

(multi)channel preferences in their browsing records?

6) What interactions within the established channel categories are well-suited to

identify users forming their purchase decision in multichannel online

environment—including homogeneous and heterogeneous inter-category

effects?

1.4 Structure of the Dissertation

In Figure 2, we illustrate the systematic structure of this dissertation. First, in the introduction

(Chapter 1), we provide an overview over the digital era in describing its relevance, recent

developments and in pointing out the aspiration it poses on marketing scholars. Complementing

this chapter, we summarize the relevant literature on online marketing effectiveness and

elucidate research gaps, framing the broader scope of our research. Based on that, we continue

with Essay 1 (Chapter 2) on attribution aiming to introduce a novel variant of attribution models

and providing novel insights into the interplay of online marketing channels. In Essay 2

(Chapter 3), we introduce a novel concept, the micro-journey, which can substantially enhance

our understanding of online browsing behavior and its outcome. Routing from flow theory, the

micro-journey operationalizes the timing between chronological user clicks in order to translate

these click patterns into purchase propensity. Next, in Essay 3 (Chapter 4) we investigate the

users present and past channel preferences when browsing with an underlying purchase

intention. Finally, we sum up our findings and provide further research directions (Chapter 5).

Introduction 26

Figure 2

Structure of the Dissertation

Introduction 27

1.5 List of Publications

Table 1

List of Publications and Conference Contributions

Authors Title Conference / Journal Date Status

Anderl, E. M., Becker, I., v. Wangenheim, F., Schumann, J. H., and Graf, F.

Analyzing the Customer Journey: Attribution Modeling for Online Marketing Exposures in a Multi-Channel Setting

Theory & Practice in Marketing Conference (TPM), London, UK

31.05 – 01.06.2013

Published

Anderl, E. M., Becker, I., v. Wangenheim, F., Schumann, J. H., and Graf, F.

Analyzing the Customer Journey: Attribution Modeling for Online Marketing Exposures in a Multi-Channel Setting

Innovative Approaches to Measuring Advertising Effectiveness Conference, Wharton Customer Analytics Initiative, Philadelphia, PA, USA

16.05.2013 Published

Becker, I., Linzmajer M., and v. Wangenheim, F.

Modeling the Click Stream – How Micro Journeys as Browsing Click Patterns Influence Conversions

2015 INFORMS 37th Annual Marketing Science Conference (ISMS), Baltimore, MD, USA

18.06 - 20.06.2015

Published

Anderl, E. M., Becker, I., v. Wangenheim, F., and Schumann, J. H.

Mapping the Customer Journey: Lessons Learned from Graph-Based Online Attribution Modeling

International Journal of Research in Marketing (IJRM; VHB3: A)

Ongoing Conditionally accepted

Becker, I., Linzmajer M., and v. Wangenheim, F.

Patterns that Matter: Browsing Click Patterns as Micro-Journeys Influence Customer Conversions

Journal of Research in Marketing (JMR; VHB3: A+)

Ongoing Under review

Becker, I., Linzmajer M., and v. Wangenheim, F.

Channels and Categories: User Browsing Preferences on the Path to Purchase

Journal of Advertising (JA; VHB3: B)

Ongoing Under review

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 29

2 Mapping the Customer Journey: Lessons Learned from Graph-Based

Online Attribution Modeling

Eva Anderl, Ingo Becker, Florian von Wangenheim, Jan H. Schumann

Advertisers employ various channels to reach customers over the Internet, who often get in

touch with multiple channels along their “customer journey.” However, evaluating the degree

to which each channel contributes to marketing success and the ways in which channels

influence one another remains challenging. Although advanced attribution models have been

introduced in academia and practice alike, generalizable insights on channel effectiveness in

multichannel settings, and on the interplay of channels, are still lacking. In response, the authors

introduce a novel attribution framework reflecting the sequential nature of customer paths as

first- and higher-order Markov walks. Applying this framework to four large customer-level

data sets from various industries, each entailing at least seven distinct online channels, allows

for deriving empirical generalizations and industry-related insights. The results show

substantial differences from currently applied heuristics such as last click wins, confirming and

refining previous research on singular data sets. Moreover, the authors identify idiosyncratic

channel preferences (carryover) and interaction effects both within and across channel

categories (spillover). In this way, the study supports advertisers’ development of integrated

online marketing strategies.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 30

2.1 Introduction

Online advertising is essential to the promotional mix of many industries (Raman et al. 2012).

Today, advertisers employ a variety of online marketing channels1 to reach potential customers,

including paid search and display marketing, as well as e-mail, retargeted displays, affiliates,

price comparison, and social media advertising. At the same time, customers visit the

advertisers’ websites on their own initiative—for instance, by directly typing in the related web

address. Using various channels, many customers visit company websites multiple times before

concluding a purchase transaction (Li and Kannan 2014). Previous visits may influence the

users’ subsequent visits, such that the customer may return to a website through the same

channel (carryover effects) or through different channels (spillover effects). Given the

proliferation of online channels and the complexity of customer journeys,2 measuring the

degree to which each channel actually contributes to a company’s success is demanding.

Despite the widespread and ongoing practice of many advertisers to apply

comparatively simple heuristics (e.g., last click wins), such that the value is attributed solely to

the marketing channel directly preceding the conversion (The CMO Club & Visual IQ 2014),

this challenge of attributing credit to different channels (Neslin and Shankar 2009) has recently

begun to receive increased attention in academia and practice alike (Berman 2015). Academics

have proposed a variety of substantiated analytical attribution frameworks, including logistic

regression models (Shao and Li 2011), game theory-based approaches (Berman 2015;

Dalessandro et al. 2012), Bayesian models (Li and Kannan 2014), mutually-exciting point

process models (Xu, Duan, and Whinston 2014), multivariate time series models (Kireyev,

Pauwels, and Gupta 2013), structural vector autoregressive (SVAR) models (Haan, Wiesel, and

1 In this study, we use the term “online marketing channels” as an umbrella phrase referring to various online

marketing instruments, including search engine advertising, display, or social media advertising. 2 We define an online customer journey of an individual customer as including all touchpoints over all online

marketing channels preceding a potential purchase decision that lead to a visit of an advertiser's website.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 31

Pauwels 2013), and hidden Markov models (Abhishek, Fader, and Hosanagar 2012).

Furthermore, several industry players such as Adometry (Google), Convertro (AOL), or

VisualIQ have introduced a range of attribution methodologies (Moffett 2014). Even though

sophisticated attribution methods become accessible to a broader audience, and marketing

executives call for such performance measures (Econsultancy 2012), in practice the full

browsing history of a user is rarely taken into account when calculating channel effectiveness

in multichannel settings (Li and Kannan 2014).

Based on a survey among marketers using advanced interactive attribution offerings,

Osur (2012) reports that the two most widely applied objectives of attribution software are to

“measure the value and performance of digital channels” and to “measure how one digital

channel affects the performance of another [channel]” (p. 4). In this research, we address both

topics, but we extend the issue further by trying to identify generalizable answers to these

challenges. For advertisers, it is valuable to know what insights from attribution apply in a

company-specific context, and what insights may be generalized across companies (industries).

Such insights can help to better explain actual channel effectiveness and can also shed light on

the interplay of channels in multi-touch environments. Marketing scholars specifically call for

further research using customer-level path data across several firms and industries to determine

spillover effects in a more generalizable way (Li and Kannan 2014). Empirical generalizations

are important for both theory generation and evaluation, and can provide valuable guidance to

managers (Kamakura, Kopalle, and Lehmann 2014). For instance, if advertisers anticipate

idiosyncratic channel preferences among some users on their path to purchase, the advertisers

could transfer this knowledge into a more adequate channel selection. Insights into channel

sequences would add to multichannel research on channel strategy, efficiency, and

segmentation (Neslin and Shankar 2009), and thus go beyond the attribution problem.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 32

To achieve our research objectives, we needed to develop and apply an attribution

mechanism with the capability of determining the effectiveness of individual marketing

channels and deriving insights on the interplay of channels in a multichannel environment

across very different data sets and conditions. In particular, we suggest an attribution framework

based on Markovian graph-based data mining techniques, extending an approach originally

developed in the context of search engine marketing (Archak, Mirrokni, and Muthukrishnan

2010). We model individual-level multichannel customer journeys as first- and higher-order

Markov graphs, using a property which we call removal effect to determine the contribution of

online channels and channel sequences. The graph-based structure of our model reflects the

sequential nature of customer journeys, enabling insights into the interplay of channels.

Applying this framework to four large, real-world customer-level data sets from three different

industries enables us to derive both cross-industry generalizations and industry-specific

findings. In doing so, we make the following contributions:

First, we contribute novel insights into online marketing effectiveness of single channels

within a multichannel setting. We estimate our graph-based framework on four data sets from

different industries, and compare the results against two well-known heuristic attribution

techniques, namely first- and last-click wins, as well as two logit models. Prior research

indicates that heuristic approaches of attributing conversion to the very last (or first) click can

produce incorrect conclusions (Abhishek, Fader, and Hosanagar 2012; Li and Kannan 2014;

Xu, Duan, and Whinston 2014). A comparison of our results across four data sets enables us to

confirm and refine these results and move toward empirical generalizations. We find that firm-

initiated channels, where the advertiser initiates the marketing communication (Bowman and

Narayandas 2001; Wiesel, Pauwels, and Arts 2011), are consistently undervalued by the

heuristic attribution approaches. For customer-initiated channels, which are triggered by

potential customers, on their own initiative (Wiesel, Pauwels, and Arts 2011), the contribution

of paid search and direct type-ins is consistently overestimated by the last click wins approach.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 33

For other customer-initiated channels, additional factors such as industry and brand

characteristics seem to play a role. With regard to the logit models, results are more ambiguous,

so that the channel rankings vary considerably across data sets.

Second, higher-order models enable us to shed light on the interplay of channels in a

multichannel setting. By comparing the results of our attribution framework across data sets,

we generalize findings from prior literature indicating that a majority of channels exhibits

idiosyncratic channel carryover (Li and Kannan 2014). Furthermore, we observe spillover

effects both within and between channel categories. Customer-initiated channels show

substantial removal effects if they are followed by other customer-initiated channels, whereas

spillover effects between firm-initiated channels are, by and large, negligible. Spillovers

between customer-initiated and firm-initiated channels (and vice-versa) are more selective and

reach a moderate level.

Third, we propose a novel variant that adds to existing advanced attribution modeling

techniques (Abhishek, Fader, and Hosanagar 2012; Berman 2015; Haan, Wiesel, and Pauwels

2013; Kireyev, Pauwels, and Gupta 2013; Li and Kannan 2014; Xu, Duan, and Whinston 2014)

by representing customer path data as first- and higher-order Markov walks. This graph-based

approach, adapted from research on paid search (Archak, Mirrokni, and Muthukrishnan 2010),

represents a useful extension to the emerging attribution literature. Whereas first- and higher-

order models offer support in measuring channel contribution in a multichannel setting, higher-

order models, in particular, allow investigation of channel sequences and spillovers between

channels.

Finally, our framework is beneficial in the approach to several explicit problems that

online marketers confront. For instance, the framework may help to calibrate online channel

budgets and move toward an optimal budget allocation. If a channel’s budget share of a channel

is higher than its actual contribution, advertisers should readjust their budget splits.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 34

Furthermore, using our framework, advertisers can more accurately calculate the conversion

probability of a customer, given his or her previous customer journey. This information can be

used to support state-of-the-art applications such as real-time bidding decisions in advertising

exchanges.

2.2 Research Background

Academic research on attribution and the interplay of online channels has only recently gained

momentum. Jordan, Mahdian, Vassilvitskii, and Vee (2011) examine allocation decisions for

publishers, using multiple attribution approaches, and derive optimal allocation and pricing

rules for publishers who are selling advertising slots. In a study of the economic welfare

consequences of the use of attribution technologies, Tucker (2012) finds evidence for more

conversions at lower costs, due to the ability to systematically shift budget toward selected

campaigns, but does not, however, disclose details on the attribution methodology.

Furthermore, academic studies address the online attribution problem: Shao and Li

(2011) introduce two attribution approaches—a bagged logistic regression model and a simple

probabilistic model. Building on their work, Dalessandro et al. (2012) propose a more complex,

causally motivated attribution methodology based on cooperative game theory. Developing

from the basis of simulated campaign data, they find that advertisers tend to assign credit to

conversions that are driven by the users' volition to convert rather than on the actual influence

of the advertisement. Focusing on the interplay between advertisers and publishers, Berman

(2015) evaluates the impact of various incentive schemes and attribution methods on publishers'

propensity to show advertisements, and on the resulting profits of advertisers. He introduces an

analytical model based on the Shapley value and compares it to the last click wins heuristic.

Abhishek, Fader, and Hosanagar (2012) suggest a dynamic hidden Markov model (HMM) that

is based on individual consumer behavior and that captures a consumer’s deliberation process

along typical stages of the conversion funnel. They find that different channels affect consumers

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 35

in different stages of their decision process. For example, display advertisements usually impact

consumers early in the decision process, moving them from a disengaged state to activity or

engagement.

Table 2

Existing Research on Attribution Modeling—Essay 1

Study Methodology Channels

Shao and Li (2011) (1) Bagged logistic regression (2) Simple probabilistic model

Search; display; social media; email; video

Dalessandro et al. (2012)

Causally motivated methodology based on cooperative game theory (Shapley value) combined with logistic regression

Real data set: General prospecting (via seven different content providers); Simulated data set: General prospecting, retargeted display, search

Abhishek, Fader, and Hosanagar (2012) Dynamic HMM Display impressions generic; display impressions specific; display clicks generic; display clicks specific; search

Berman (2015) Analytical model based on cooperative game theory (Shapley value) combined with OLS regression

Different publishers: Two online magazines; two display ad networks, two travel search websites, one online travel agency, one media exchange network/retargeting

Haan, Wiesel, and Pauwels (2013) Structural vector autoregression (SVAR)

Television; radio; email; SEA product; SEA branded; retargeting; referrer; portals; comparison websites; SEO (as control); affiliate (as control)

Kireyev, Pauwels, and Gupta (2013) Multivariate time-series model (persistence modeling)

Display impressions; display clicks; search impressions; search clicks

Li and Kannan (2014) Bayesian model SEO; SEA; referrer; direct type-in; email; display

Xu, Duan, and Whinston (2014) Mutually exciting point process model Search; display; other (classified; affiliate)

Our study Markov graphs (first- and higher-order)

Four data sets from three industries: SEA; SEO; direct type-in; affiliate; display; price comparison; email; referrer; retargeting; social media; other

Note: Haan, Wiesel, and Pauwels (2013) base their study on aggregated data on a daily level, and further include several dimensions, for instance, whether the contact is firm- or customer-initiated or the degree of the creatives' content integration. Kireyev, Pauwels, and Gupta (2013) base their study on aggregated data on a weekly level.

Li and Kannan (2014) propose a Bayesian model for measuring online channel

consideration, visits, and purchases by using individual conversion path data and validating it

in a field experiment. They use the estimated carryover and spillover effects to attribute

conversion credit to different channels, and they find that the relative contributions of these

channels are significantly different from last click wins. By means of a mutually exciting point

process model, Xu, Duan, and Whinston (2014) calculate average conversion probabilities for

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 36

different online advertising channels, showing that the conversion rate measure underestimates

the effect of display advertisements, as compared to search advertisements. A multivariate time-

series model based on aggregate data by Kireyev, Pauwels, and Gupta (2013) analyzes

attribution dynamics for display and for search advertising. They derive spillover effects from

display toward search conversion; however, display advertisements also increase search clicks,

thereby increasing costs for search engine advertising. Finally, Haan, Wiesel, and Pauwels

(2013) propose a structural vector autoregression (SVAR) model, also based on aggregate data,

to determine the effectiveness of various offline and online advertising channels. Given that

prior research relies on single data sets, such that findings may be company- or industry-

specific, this study extends the existing literature on multichannel online advertising by

applying a novel attribution framework based on Markov graphs on four data sets from three

different industries (Table 2).

2.3 Data

Our research is based on four clickstream data sets provided by online advertisers, in

collaboration with a multichannel tracking provider. Clickstream data record each user's

Internet activity, and thus trace the navigational path the customer takes (Bucklin and Sismeiro

2009). For each visit to the advertiser’s website during the observation period, the data include

detailed information about the source of the click and an exact timestamp. Clicks either

represent a direct behavioral response to an advertising exposure, or result from the user

entering the advertiser’s URL directly into the browser, so these sources comprise all online

marketing channels, as well as direct type-ins. We also know for each visit whether it was

followed by a conversion—in this case a purchase transaction. We use these data points to

construct customer journeys that describe the click pattern of individual consumers across all

online marketing channels and their purchase behavior. Thus, we not only track successful

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 37

journeys ending with a conversion, but also journeys that never lead to a conversion, and we do

so within a timeframe of thirty days from the last exposure.

The data collection occurs at the cookie-level, such that we identify individual

consumers—or more accurately, individual devices. The use of cookie data suffers a number of

limitations, such as an inability to track multi-device usage or bias due to cookie deletion (Flosi,

Fulgoni, and Vollman 2013), yet cookies remain the industry standard for multichannel tracking

(Tucker 2012). We do not include information on offline marketing channels, because

measuring individual-level exposure to multiple offline media proves highly difficult in practice

(Danaher and Dagger 2013).

Table 3

Descriptive Statistics of the Data Sets—Essay 1

Description Data Set 1 Data Set 2 Data Set 3 Data Set 4

Industry Travel agency Fashion retail Fashion retail Luggage retail

Number of different channels 8 8 7 7

Number of clicks 1,478,359 1,639,467 1,125,979 615,111

Number of journeys 600,978 1,184,583 862,112 405,339

Thereof with length ≥ 2 206,519 170,914 142,039 105,031

Thereof with length ≥ 5 48,344 30,095 12,416 11,475

Journey length (SD) 2.46 (8.860) 1.38 (1.916) 1.31 (1.238) 1.52 (4.587)

Number of conversions 9,860 10,153 16,200 8,115

Journey conversion rate 1.64% 0.86% 1.88% 2.00%

Note: Standard deviations are in parentheses.

The advertisers that provide the data sets for this study operate in different industries:

Data Set 1 was provided by an online travel agency. Data Sets 2, 3, and 4 originate from

specialized online retailers selling apparel and luggage, respectively. All the advertisers in our

sample address a broad audience and are pure online players, so we can exclude online/offline

cross-channel effects. Each data set includes a minimum of 405,000 journeys per advertiser.

Their average length is 1.3 – 2.5 contacts, and between 0.9% and 2.0% of all journeys lead to a

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 38

successful conversion. In Table 3, we present detailed descriptions of the journeys in our data

sets.

All advertisers included in the evaluation distinguish seven or eight different online

channels, though across firms the channels that are used differ. Search engine advertising (SEA)

and search engine optimization (SEO) appear in all four data sets. Other channels employed by

the advertisers include direct type-in, affiliate marketing, display, price comparison, newsletter,

referrer, retargeting, social media advertising, and others. Table 4 provides an overview of the

online channels in our data, distinguishing between firm- and customer-initiated channels. As

online marketing contacts can be initiated either by customers or by the firm, the origin of the

contact is recognized as an important differentiator for online marketing channels (Haan,

Wiesel, and Pauwels 2013; Li and Kannan 2014; Wiesel, Pauwels, and Arts 2011). In firm-

initiated channels such as display advertising, the advertiser determines timing and exposures;

in customer-initiated channels, customers actively trigger the communication—for instance, by

performing a keyword search.

Table 5 provides information on the distribution of clicks across channels, illustrating

the variation within and between data sets. The frequency of channels varies considerably

across the four data sets. For example, though affiliate accounts for 46.1% of all clicks in Data

Set 2, it is the least frequent channel in Data Set 4 (0.3% of clicks). This variation also alleviates

endogeneity concerns. To rule out potential endogeneity, we furthermore conducted an analysis

of pairwise correlations between logit model input variables (see Section 2.5 for variable

specifications), which reveals low correlations between channels. The correlation matrices can

be found in the Appendix.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 39

Table 4

Definitions of Online Channels—Essay 1

Online channel Description

Channel type

Type-in Visits are classified as (direct) type-in if users access the advertiser’s website directly by entering the URL in their browser’s address bar, or by locating a bookmark, favorite, or shortcut.

Customer-initiated

Search (SEA / SEO)

A consumer searching for a keyword in a general search engine (e.g., Google) receives two types of results: organic search results ranked by the search algorithm, and sponsored search results, also known as paid search or search engine advertising (SEA). While organic search or search engine optimization (SEO) results are available for free, SEA clicks are sold via second-price auctions.

Customer-initiated

Price Comparison

Price comparison websites are vertical search engines that allow users to compare products by price and features. They aggregate product listings from a multitude of businesses, and direct users toward their websites.

Customer-initiated

Display Display advertising, respectively banner advertising, entails embedding a graphical object with the advertising message into a website. Timing and exposures of display banners are determined by the firm.

Firm-initiated

Newsletter Newsletter marketing, also known as email marketing, encompasses sending marketing messages toward potential customers using email.

Firm-initiated

Retargeting Retargeting is a subclass of display advertising that is personalized toward the user based on his or her browsing history. It aims to re-engage users who have visited an advertiser’s website, but did not complete a purchase.

Firm-initiated

Social Media

Social media advertising comprises a set of advertising platforms belonging to the field of social media, such as social networks (e.g., Facebook), micromedia (e.g., Twitter), or other (mobile) sharing platforms (e.g., Instagram). In one of our data sets, the advertiser uses targeted Facebook display advertisements, which we define as social.

Firm-initiated

Affiliate Affiliate marketing is a form of commission-based marketing in which a business (e.g., retailer) rewards the affiliate (e.g., a product review website) for referring a user toward the business’s website. As affiliate in our data sets may include both coupon websites that are customer-initiated and advertisements provided by affiliate networks that may be more firm-initiated, a clear differentiation between customer- and firm-initiated contacts across data sets is not possible.

Customer-initiated / Firm-initiated

Referrer Referral or referrer traffic covers all traffic that is forwarded by external content websites (with or without remuneration)—for example, by including a text link. As traffic sources vary across data sets, a clear differentiation between customer- and firm-initiated contacts across data sets is not possible.

Customer-initiated / Firm-initiated

Other All forms of advertising that do not clearly fit into one of the categories above, are gathered in a separate category which we call “other”.

Customer-initiated / Firm-initiated

Note: Affiliate and referrer may be customer-initiated in some cases (e.g., for coupon websites) or firm-initiated (e.g., advertisements by third parties).

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 40

Table 5

Channel Distribution by Data Set—Essay 1

Description

Data Set 1 (Travel) Data Set 2 (Fashion) Data Set 3 (Fashion) Data Set 4 (Luggage) Clicks / journey

Share of clicks Rank

Clicks / journey

Share of clicks Rank

Clicks / journey

Share of clicks Rank

Clicks / journey

Share of clicks Rank

Type-in n/a n/a n/a 0.29

20.7% 2 .25

18.9% 3 .13

8.5 % 3 (1.05) (0.81) (4.42)

SEA .58

23.4% 2 0.15

10.5% 4 .18

14.1% 4 1.13

74.3 % 1 (0.90) (0.72) (.55) (1.12)

SEO .17

6.8% 3 0.17

12.1% 3 .43

33.1% 1 .16

10.8 % 2 (0.61) (0.795) (.28) (0.54)

Price Comparison

.093.7% 4

00.1% 8 n/a n/a n/a

.042.5 % 5

(1.83) (0.053) (0.27)

Display 1.46

59.5% 1 0.02

1.2% 7 n/a n/a n/a n/a n/a n/a (8.75) (0.3)

Newsletter .03

1.2% 7 0.07

4.9% 5 .02

1.3% 6 n/a n/a n/a (0.24) (0.72) (.16)

Retargeting .01

0.4% 8 n/a n/a n/a .00

n/a n/a .02

1.1 % 6 (0.16) (.01) (.20)

Social Media n/a n/a n/a n/a n/a n/a .28

21.5% 2 n/a n/a n/a (.82)

Affiliate .06

2.5% 6 0.64

46.1% 1 n/a n/a n/a .00

0.3 % 7 (0.34) (1.02) (.16)

Referrer n/a n/a n/a 0.06

4.3% 6 .14

11.0% 5 .04

2.5 % 4 (0.26) (.38) (.21)

Other .06

2.6% 5 n/a n/a n/a n/a n/a n/a n/a n/a n/a (0.36)

Note: Standard deviations are in parentheses.

2.4 Model Development

We propose a graph-based Markovian framework to analyze customer journeys and derive an

attribution model, adapting an approach proposed by Archak, Mirrokni, and Muthukrishnan

(2010) in the context of search engine advertising. Markov chains are probabilistic models that

can represent dependencies between sequences of observations of a random variable. They have

a long history in marketing (Styan and Smith 1964) and have frequently been used to model

customer relationships (Homburg, Steiner, and Totzek 2009; Pfeifer and Carraway 2000). Other

applications include advertising frequency decisions (Bronnenberg 1998) and brand loyalty

(Che and Seetharaman 2009).

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 41

In our model, we represent customer journeys as chains in directed Markov graphs.3

A Markov graph � = ⟨�, �⟩ is defined by a set of states

� = �, … , � ( 1 )

and a transition matrix W with edge weights

��� = ���� = ������ = ��, 0 ≤ ��� ≤ 1, ∑ ��� = 1 ∀ ��� � . ( 2 )

Using this graph-based approach allows us to represent and analyze customer journeys

in an efficient way, as the size of the final graph does not depend upon the number of journeys

in the data set, but only on the number of states.

2.4.1 Base Model

Customer journeys contain one or more contacts across a variety of channels. In the base model,

each state si corresponds to one channel. If an advertiser employs three different channels (C1,

C2, and C3) in his online marketing mix, the model would include the three states C1, C2, and

C3.4 Additionally, all graphs contain three special states: a START state that represents the

starting point of a customer journey; a CONVERSION state representing a successful

conversion; and an absorbing NULL state for customer journeys that have not ended in a

conversion during the observation period. The full set of states S in our example would, then,

appear as follows: S = {START, CONVERSION, NULL, C1, C2, C3}.

The transition probability wij in the base model corresponds to the probability that a

contact in channel i is followed by a contact in channel j. For the first channel in each journey,

we add an incoming connection from the START state. If a customer journey ends in a

conversion, we connect the state representing the last channel in the journey to the

3 Called adgraphs by Archak, Mirrokni, and Muthukrishnan (2010). 4 As we do not make any assumptions on the channels used, we employ dummy channels in our examples. In

practice, the set of channels—and thus the set of states—depends on the actual channels used by the advertiser.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 42

CONVERSION state; otherwise, it leads to the NULL state. For modeling reasons, we always

add a connection from the CONVERSION state to the NULL state. Cycles in the graph are

possible, such as when a sequence of two identical channels appears in a customer journey.

Figure 3 shows an exemplary Markov graph based on four customer journeys. Figure 4 provides

a graphical structure of the simple model for Data Set 1.

Figure 3

Exemplary Markov Graph—Essay 1

Figure 4

Markov Graph for Data Set 1 (base model)—Essay 1

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 43

2.4.2 Higher-Order Models

Markovian models suggest that the present depends only on the first lag and do not incorporate

previous observations. However, because prior research suggests that clickstreams should not

be regarded as strictly Markovian (Chierichetti et al. 2012; Montgomery et al. 2004), we extend

the approach proposed by Archak, Mirrokni, and Muthukrishnan (2010) by introducing

alternative higher-order models in which the present depends on the last k observations.

Transition probabilities can thus be defined as follows:

�!�� = �|���� = ���, ���# = ��# , … , �� = � $ = �!�� = �|���� = ��� , ���# = ��# , … , ���% = ��% $.

( 3 )

For our implementation, we utilize the knowledge that a Markov chain of order k, over

some alphabet A, is equivalent to a first-order Markov chain over the alphabet Ak of k-tuples.

States in higher-order models, therefore, include k-tuples of states in the first-order models,

such that we can employ the same algorithms. Unfortunately, the number of independent

parameters increases exponentially with the order of the Markov chain, and quickly becomes

too large to be estimated efficiently with real-world data sets. Considering these implementation

issues in relation to algorithmic efficiency, we limit our analyses to Markov chains with a

maximum order of four.

2.4.3 Removal Effect

The representation as Markov graphs allows identifying structural correlations in the customer

journey data that can be used to develop an attribution model. Archak, Mirrokni, and

Muthukrishnan (2010) propose a set of ad factors to capture the role of each state, such as

Eventual Conversion(si)—that is, the probability of reaching conversion from a given state si.

Visit(si) is the probability of passing si on a random walk beginning in the START state. For

attribution modeling, we propose using the ad factor Removal Effect(si), defined as the change

in probability for reaching the CONVERSION state from the START state when we remove si

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 44

from the graph. As Removal Effect(si) reflects the change in conversion rate if the state si was

not present, it enables us to simulate counterfactual analysis on historical data, and is therefore

well suited for measuring the contribution of each channel (or channel sequence). Using the

assumption that all incoming edges of the state si that we remove are redirected to the absorbing

NULL state, Removal Effect(si) is equivalent to the multiplication of Visit(si) and Eventual

Conversion(si).5 The removal effect can thus be efficiently calculated using matrix

multiplication or by applying local algorithms provided by Archak, Mirrokni, and

Muthukrishnan (2010). Table 6 illustrates the calculation of removal effects for the exemplary

Markov graph presented in Figure 3. In this graph that is based on four distinct journeys, C1

has a visit probability of .75 and an eventual conversion probability of 1. Multiplying these two

values leads to a removal effect of .75 or 42.86% of the total removal effect in percent.

Compared to C1, both C3 and C4—again with an eventual conversion probability of 1—have

a lower removal effect of .5 (28.57%) due to their lower visit probability of .5. The removal

effect of C2, which only appears in non-converting journeys, is zero.

Table 6

Removal Effect Calculation of Exemplary Markov Graph (in Figure 3)—Essay 1

Channel Visit(si) Eventual Conversion(si)

Removal Effect(si) Removal Effect(si), in %

C1 .75 1.00 .75 42.86%

C2 .25 .00 .00 .00%

C3 .50 1.00 .50 28.57%

C4 .50 1.00 .50 28.57%

Removal Effect(si) can take values between 0 and the total conversion rate. However, as

most existing attribution heuristics use percentage values, we report removal effects per state

as percentages of the sum of all removal effects (excluding the special states START,

5 For a proof, see Archak, Mirrokni, and Muthukrishnan (2010).

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 45

CONVERSION, and NULL), when comparing our results to other models. Higher-order

models allow us to calculate removal effects for states representing channel sequences. To

calculate the removal effect for channel i in higher-order models, we calculate the mean removal

effect of all states containing channel i as the last observation or channel.

2.5 Model Fit

In the following, we evaluate the proposed framework with regard to predictive accuracy and

robustness and compare it to four benchmark models. Higher-order models outperform both the

last click wins and the first click wins attribution heuristic models, as well as a simple logit

model. In a comparison to a second logit model that includes order effects, they show similar

results.

We compare our approach to the last click wins and first click wins heuristics, as well

as to two different logit models. The specification of Logit Model 1 includes the number of

clicks for each channel i:

'()�*!+�$ = α + .�/� + .#/# + ⋯ + . / , ( 4 )

where xi is the number of clicks in channel i in the journey. This model does not reflect

the sequence of channels, but only the total number of clicks per channel. Therefore, we add a

second logit model which, in contrast, includes order effects:

'()�*!+��$ = α + 1 .!2��3$4�� + .!2��#$4�# + .!2���$4�3 + .2�4�2 ,

� �

( 5 )

where dit is a dummy variable for a click in channel i at position t, counting from the

end of the journey. The dummy variable dit is coded as 1 if channel i is present in position t;

otherwise it is set to 0. To make the number of variables in the logit model tractable, we include

the last four contacts of each journey. This definition improves the comparability to fourth-

order models, which take the last four contacts into account to determine the next transition.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 46

2.5.1 Predictive accuracy

Although attribution primarily takes a retrospective view, attribution models should be able to

correctly predict conversion events (Shao and Li 2011). In addition to ensuring scientific rigor,

this classification helps to persuade managers of the model’s credibility (Lodish, 2001), and

lays the foundation for further applications beyond attribution, such as real-time bidding. We

therefore use the 10-fold cross-validation, which is superior for measuring predictive

performance—both within and out of the sample—to leave-one-out validation or bootstrapping,

since all the data serve as the holdout once (Sood, James, and Tellis 2009). However, standard

metrics for classification accuracy, such as percentage correctly classified or log-likelihood, are

poor metrics for measuring classification performance in the case of unequal misclassification

costs or when class distribution is skewed (He and Garcia 2009). As the discriminative power

of these measures is limited in our context, where journey conversion rates do not exceed 2%,

we turn to alternative measures to evaluate predictive accuracy.

First, we choose the receiver operating characteristic (ROC) curve that decouples

classification performance from class distributions and misclassification costs. A ROC curve is

a two-dimensional graph; the true positive rate α is plotted on the x-axis, while the false positive

rate 1 - β appears on the y-axis (Bradley 1997). To compare our models, we reduce ROC

performance to a single scalar value, the area under the ROC curve (AUC), which we calculate

using linear interpolation (Bradley 1997). Figure 5 contains the ROC curves for all models

based on a within-sample evaluation of all journeys. As a second measure, we calculate the top-

decile lift for each model. Top-decile lift is defined as the proportion of the 10% of journeys

predicted to be most likely to convert that actually end in a conversion relative to the baseline

conversion rate (Neslin et al. 2006).

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 47

Figure 5

ROC Curves (within Sample)—Essay 1

Data Set 1

Data Set 2

Data Set 3

Data Set 4

In Table 7, we report both measures. Although the overall predictive accuracy varies

substantially between data sets, the relative predictive performance of the different model types

is comparable, leading to similar rankings of the model types. Within and out-of-sample

performance for all models is similar, indicating a low risk of overfitting. With the exception

of Data Set 2, the base model outperforms the first click wins heuristics and leads to results

similar to the last click wins approach, yet does not predict conversions as well as the two logit

models. Increasing the memory capacity substantially improves the predictive performance of

our graph-based models, such that they clearly outperform the base model as well as the

heuristic one-click approaches. In comparison to the logit models, the results move closer

together.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 48

Table 7

Predictive Accuracy by Model and Data Set—Essay 1

Model

Measure Sample Data Set

Base Model

Second Order

Third Order

Fourth Order

Last Click Wins

First Click Wins

Logit Model 1

Logit Model 2

AUC

Within sample

DS 1 .7398

(.0006)

.8214

(.0007)

.8323

(.0008)

.8357

(.0010)

.7398

(.0006)

.7133

(.0005)

.8185

(.0022)

.8278

(.0010)

DS 2 .7604

(.0006)

.8834

(.0006)

.8932

(.0005)

.8967

(.0005)

.7609

(.0007)

.7726

(.0008)

.8654

(.0011)

.8888

(.0004)

DS 3 .6079

(.0009)

.7933

(.0015)

.7996

(.0014)

.8040

(.0018)

.6079

(.0009)

.5980

(.0006)

.8037

(.0011)

.8041

(.0010)

DS 4 .6029

(.0010)

.7191

(.0029)

.7309

(.0009)

.7333

(.0009)

.6029

(.0010)

.5528

(.0007)

.7281

(.0013)

.7308

(.0011)

Out-of-sample

DS 1 .7396

(.0061)

.8207

(.0022)

.8305

(.0062)

.8294

(.0063)

.7406

(.0056)

.7136

(.0050)

.8185

(.0074)

.8274

(.0071)

DS 2 .7607

(.0056)

.8832

(.0054)

.8924

(.0044)

.8933

(.0050)

.7619

(.0056)

.7728

(.0071)

.8653

(.0086)

.8887

(.0037)

DS 3 .6079

(.0079)

.7930

(.0086)

.7988

(.0088)

.8006

(.0089)

.6093

(.0082)

.5990

(.0064)

.8036

(.0088)

.8040

(.0089)

DS 4 .6030

(.0093)

.7180

(.0083)

.7280

(.0080)

.7230

(.0072)

.6050

(.0080)

.5550

(.0063)

.7276

(.0084)

.7299

(.0084)

Top-decile lift

Within sample

DS 1 2.9056

(.0211)

4.7181

(.0249)

5.2798

(.0198)

5.3587

(.0203)

2.9056

(.0211)

2.4333

(.0203)

4.9226

(.0261)

5.1854

(.0500)

DS 2 2.4880

(.0407)

6.8070

(.0156)

6.9306

(.0146)

6.9546

(.0158)

2.4955

(.0312)

2.8692

(.0311)

6.6011

(.0211)

6.8893

(.0146)

DS 3 1.5065

(.0254)

5.1744

(.0207)

5.6426

(.0221)

5.6450

(.0232)

1.5065

(.0254)

1.4967

(.0188)

5.7731

(.0197)

5.7358

(.0186)

DS 4 2.8089

(.0137)

3.4708

(.1148)

4.1838

(.0134)

4.2897

(.0161)

2.8089

(.0137)

1.6820

(.0144)

4.0112

(.1042)

4.2194

(.0171)

Out-of-sample

DS 1 2.9075

(.1385)

4.6937

(.1593)

5.2393

(.1563)

5.2913

(.1704)

2.9175

(.1508)

2.4296

(.0934)

4.9260

(.1487)

5.1856

(.1499)

DS 2 2.4748

(.0742)

6.7889

(.1442)

6.9277

(.1178)

6.9113

(.1202)

2.5103

(.1078)

2.8709

(.1532)

6.6109

(.1684)

6.8978

(.1023)

DS 3 1.4896

(.0910)

5.1619

(.1433)

5.6311

(.1993)

5.6107

(.2055)

1.5418

(.1002)

1.5143

(.1368)

5.7714

(.1704)

5.7278

(.1661)

DS 4 2.8112

(.1074)

3.4556

(.1737)

4.1198

(.1273)

4.1553

(.1464)

2.8256

(.1114)

1.6897

(.1337)

4.0124

(.1332)

4.2090

(.1522)

Note: Standard deviations are in parentheses.

With regard to Data Set 1, Data Set 2 and Data Set 4, third-order and fourth-order

Markovian models slightly outperform the Logit Model 1 in AUC, and more substantially in

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 49

top-decile lift. For Data Set 3, which has the shortest journeys in our sample, the AUC

performance of the higher-order models is similar to Logit Model 1; however, they do not attain

the top-decile lift of the first logit model.

Overall, the second logit model has the best predictive abilities among the four

benchmark models. The third- and fourth-order models slightly outperform Logit Model 2 in

all measures for the Data Sets 1 and 2, collectively, yet the predictive performance metrics of

these models are remarkably close together, making it increasingly complex to clearly

denominate the model with the superior predictive performance. That said, the results differ by

metric for Data Sets 3 and 4: For Data Set 3, the second logit model shows a marginally better

performance for all metrics. For Data Set 4, the within and out-of-sample results are mixed.

Whereas Logit Model 2 performs slightly better for out-of-sample predictions, the fourth-order

models outperform all other approaches for within-sample prediction. In summary, the higher-

order models outperform both heuristic approaches and the first logit model, and are at a

comparable level with the more complex second logit model. Although we cannot attest to our

models’ clear superiority over all tested benchmark models, our framework achieves results

comparable to widely-accepted and tested models, making it well-suited for analyzing the

structural properties of customer journeys.

2.5.2 Robustness

Robustness is another important metric to evaluate model fitness (Little 2004b; Shao and Li

2011), as it conveys the ability of a model to deliver stable and reproducible results if the model

is run multiple times and is therefore indispensable for sustainable attribution results. For our

models, robustness applies to two measures. First, predictive accuracy should be robust across

all cross-validation repetitions. Table 7 lists the standard deviations of the predictive

performance measures for each model, as well as for the four models we use as a comparison.

The results imply low overall variation without systematic differences between models.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 50

Second—and even more important—the variable used for attribution modeling should provide

stable attribution results that offer a reliable basis for managerial decisions, such as budget

shifts. Therefore, we specifically test the robustness of the Removal Effect(si) ad factor. For

each model state si, we compute the average standard deviation of Removal Effect(si) across ten

cross-validation repetitions. We report the stability of the removal effects as percentages of the

average removal effect across all states, as the number of states per model and, correspondingly,

the mean Removal Effect(si) varies. We summarize these validation results in Table 8. For all

data sets in our sample, the average standard deviation as a percentage of the average removal

effect increases with model order, leading to a necessary trade-off between predictive accuracy

and robustness.

Table 8

Removal effect: Average Standard Deviation as % of Average Removal Effect

(10-Fold Cross-Validation)—Essay 1

Model

Data Set Base Model Second Order Third Order Fourth Order

Data Set 1 (Travel) 1.14% 1.92% 3.25% 5.43%

Data Set 2 (Fashion) 1.51% 2.10% 3.81% 7.57%

Data Set 3 (Fashion) 1.31% 1.72% 2.78% 5.15%

Data Set 4 (Luggage) 1.34% 1.80% 2.97% 4.67%

Combining our evaluation results with practical considerations, we recommend third-

order models for standard attribution analyses, as they balance predictive performance,

robustness, and algorithmic efficiency. Removal effects can be calculated efficiently in

O(│S│2) time (Cormen et al. 2009), and hence allow for frequent model updates. However, as

the number of states increases exponentially with the order of the Markov chain, we limit our

analyses to fourth-order models to allow for updates in near real-time.6 Using higher-order

6 However, depending on the use case, an advertiser may choose a different trade-off between predictive accuracy,

robustness, and algorithmic efficiency.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 51

models also yields additional insights into channel interactions, thereby further increasing

managers’ understanding of the interplay across channels.

2.6 Results

2.6.1 Attribution Results

Next, we take a closer look at the attribution results of our proposed framework and compare it

with the attribution heuristics most widely used in industry practice, namely last click wins and

first click wins (The CMO Club & Visual IQ 2014), and the two logit models, which show

comparable predictive abilities. As prior research indicates that heuristic attribution approaches

can lead to incorrect conclusions (Abhishek, Fader, and Hosanagar 2012; Li and Kannan 2014;

Xu, Duan, and Whinston 2014), a comparison of our results across four data sets enables us to

investigate structural differences and move toward empirical generalizations. Our analyses are

based on the full data sets and show the contribution of each channel toward final conversions.

Given our evaluation results, we use third-order Markov models in our comparison and

calculate the mean removal effect of all states containing channel i as the last observation or

channel. To enable comparisons to the first click and last click heuristic, we present our model

results as percentage values in Table 9.

First, we analyze, across data sets, the attribution results provided by our framework. In

sum, the customer-initiated channels (direct type-in, SEA, SEO, and price comparison)

accumulate the majority of the contribution (between 68% for Data Set 3 and 92% for Data Set

4), highlighting the general importance of companies’ online marketing success (Shankar and

Malthouse 2007). However, the actual contributions and corresponding rankings by channel

vary substantially across data sets. SEA, the strongest channel for Data Set 4 (60%) and Data

Set 1 (46%), receives less than 20% of the credit for Data Sets 2 and 3. For firm-initiated

channels, the results also show considerable variation. For example, a newsletter shows values

between 1% (Data Set 3) and 15% (Data Set 2). These differences may reflect the advertisers’

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 52

individual channel strategies, such that marketers need to individually derive and verify detailed

implications.

Table 9

Attribution Results in Comparison to Two Heuristic Models (in %)—Essay 1

Data Set 1 (Travel) Data Set 2 (Fashion) Data Set 3 (Fashion) Data Set 4 (Luggage)

Markov Graph

(Third Order)

Last Click Wins

First Click Wins

Markov Graph

(Third Order)

Last Click Wins

First Click Wins

Markov Graph

(Third Order)

Last Click Wins

First Click Wins

Markov Graph

(Third Order)

Last Click Wins

First Click Wins

% % % % % % % % % % % %

Type-in n/a n/a n/a 35.23% 43.91% 40.28% 27.55% 29.77% 25.51% 18.89% 22.02% 13.71%

SEA 46.29% 53.19% 56.36% 19.81% 22.27% 23.60% 18.53% 20.16% 20.70% 59.91% 60.98% 76.26%

SEO 19.54% 16.76% 16.67% 15.79% 13.66% 13.24% 22.03% 21.33% 21.12% 9.24% 7.79% 5.30%

Price Comparison 5.29% 4.78% 6.05% 0.21% 0.11% 0.12% n/a n/a n/a 3.56% 2.17% 2.21%

Display 0.93% 0.14% 0.21% 3.70% 1.79% 1.37% n/a n/a n/a n/a n/a n/a

Newsletter 4.47% 2.93% 4.28% 14.54% 8.76% 11.94% 1.21% 1.15% 1.32% n/a n/a n/a

Retargeting 1.25% 0.67% 0.78% n/a n/a n/a 0.07% 0.01% 0.00% 2.09% 1.72% 0.16%

Social Media n/a n/a n/a n/a n/a n/a 24.87% 20.73% 23.83% n/a n/a n/a

Affiliate 19.71% 20.17% 13.66% 8.72% 7.83% 6.87% n/a n/a n/a 4.41% 3.67% 0.42%

Referrer n/a n/a n/a 1.99% 1.67% 2.58% 5.72% 6.85% 7.52% 1.88% 1.65% 1.96%

Other 2.52% 1.36% 1.97% n/a n/a n/a n/a n/a n/a n/a n/a n/a

χ2 12,911 18,019 40,931 30,076 5,948 4,822 3,357 36,979

Df 7 7 7 7 6 6 6 6

P <.001 <.001 <.001 <.001 <.001 <.001 <.001 <.001

However, we observe significant differences when comparing the results of the Markov

model to those of the heuristic approaches, some of which apply consistently across data sets

and allow for generalizations of prior research. In general, the Markov model levels the channel

contributions’ amplitudes, distributing the contribution more evenly across channels. The

heuristic attribution models consistently assign less credit to the firm-initiated channels display,

retargeting, social media, and newsletter.7 Display as well as retargeting and social media

advertising, which—for the advertisers in our sample—can be regarded as specialized forms of

7 With the exception of Data Set 3, where the credit assigned to newsletter is higher in the first click model.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 53

banner advertising, may be underrepresented by one-click approaches, as banners increase

brand awareness (Ilfeld and Winer 2002; Sherman and Deighton 2001) and thereby reduce the

user’s cost of subsequent website visits through search channels (Li and Kannan 2014).

Furthermore, retargeting is, by definition, inappropriately reflected by the first click approach,

as it highlights specific products or services the user was browsing previously (Lambrecht and

Tucker 2013). For a newsletter, extant research reveals positive but weak short-term effects, yet

strong long-term effects (Breuer, Brettel, and Engelen 2011). Newsletters, stored in the inbox,

can be repeatedly accessed and may motivate subsequent visits through other channels. Single-

click heuristics may not be able to reflect these effects appropriately.

Similar to the Markov model results, customer-initiated channels accumulate the

majority of the contribution in the heuristic approaches. However, they leave a more ambiguous

picture on a channel level: compared to heuristic approaches, the Markov model consistently

assigns less credit to SEA, but more credit to SEO. Direct type-ins, when users directly access

the company website, get more credit by the last click wins approach, whereas price comparison

receives less credit. Our findings for SEA and direct type-in are consistent with prior research

on attribution: Xu, Duan, and Whinston (2014) show that the last click wins heuristic is biased

toward SEA. Li and Kannan (2014) find that the informational stock of other channel visits

catalyzes ultimate purchases via direct type-ins, such that type-in is overestimated by the last

click wins approach. However, in contrast to previous findings (Li and Kannan 2014), our

approach assigns more credit to SEO than the heuristic approaches. These results could be

explained by differences between product and retailer brands. The hotel chain in the study by

Li and Kannan (2014) has a strong “product” brand. In contrast, the advertisers in our sample

all act as retailers. For organic search, keywords containing product brands positively predict

direct conversions, yet there is no significant effect for retailer brands (Yang and Ghose 2010).

As a consequence, SEO would benefit less from spillovers from other channels. Similar to SEO,

price comparison gets more credit than in the last click wins approach. Breuer, Brettel, and

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 54

Engelen (2011) find that price comparison websites not only affect sales in the short term, but

compared to other online channels also have a moderate long-term effect, which may not be

fully captured by one-click heuristics.

The results for affiliate and referrers marketing are inconclusive. This corresponds to

the fact that a clear differentiation between customer- and firm-initiated contacts across data

sets is not possible. For Data Set 1, the affiliate channel includes coupon websites, which may

be frequented by more experienced customers who check for an available discount coupon

before finalizing their purchase. Although the last click wins attribution approach would allow

those affiliate websites to “free-ride” on previous contacts (Berman 2015), our framework

assigns less credit to the affiliate channel. For Data Sets 2 and 4, where the affiliate channel

includes more firm-initiated contacts, the Markov model assigns more credit to affiliate than

the heuristic approaches, which is in line with our finding that heuristic approaches

underestimate firm-initiated channels.

In our data sets, referrer mainly relates to aggregator websites that resemble search

engines in their functionality, and thus feature customer-initiated elements. Compared to the

first-click model, our framework constantly assigns less credit to referrer—similar to our results

for SEA. With regard to the last-click model, the results are mixed: While our model assigns

more credit to referrer for Data Sets 2 and 4, it assigns less credit for Data Set 3. Different levels

of brand awareness about the advertisers may explain these findings and should be investigated

in more detail.

The results of Logit Model 1, which we present in Table 10, reveal a larger variation

between data sets. The strongest predictors of conversion events are affiliate for Data Sets 1

and 4, price comparison for Data Set 2, and retargeting for Data Set 3. In contrast, our graph-

based model indicates that all mentioned channels play only a subordinate role. These

differences become especially evident when analyzing Data Set 2. According to the Markov

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 55

model, as well as to the first and last click wins approach, less than 1% of credit is attributed to

price comparison, which is the channel with the lowest number of contacts and is present in

only 47 of the 10,153 converting journeys. In contrast, price comparison has the highest

marginal effect (ME = 0.0035, p < .001) in the logit model. These discrepancies derive from

different underlying model mechanisms. Logit models do not aim to reflect the overall

contribution of a variable (i.e., channel), but rather focus on the predictive ability of a variable

with regard to a binary dependent variable (e.g., a conversion event). Thus, the logit models are

well-suited for calibrating the predictive capabilities of other frameworks—for example, our

graph-based approach.

Table 10

Estimation Results for Logit Model 1—Essay 1

Data Set 1 (Travel) Data Set 2 (Fashion) Data Set 3 (Fashion) Data Set 4 (Luggage)

B Exp(B) ME B Exp(B) ME B Exp(B) ME B Exp(B) ME

Type-in .259 *** 1.295 .0016*** .707*** 2.028 .0098*** .376*** 1.457 .0066***

SEA .443*** 1.557 .0000*** .417 *** 1.517 .0026*** .724*** 2.063 .0101*** .243*** 1.275 .0043***

SEO .388*** 1.474 .0000*** .196 *** 1.216 .0012*** .440*** 1.552 .0061*** .142*** 1.153 .0025***

Price Comparison .002* 1.002 .0000*** .548 *** 1.730 .0035*** .255*** 1.290 .0045***

Display -3.427*** .032 -.0004*** .188 *** 1.207 .0012***

Newsletter .616*** 1.852 .0001*** .277 *** 1.319 .0017*** .719*** 2.052 .0100***

Retargeting .310*** .1363 .0000*** .939 2.558 .0131 .195*** 1.215 .0034***

Social Media .504*** 1.655 .0070***

Affiliate 1.152*** 3.165 .0001*** -.195 *** .823 -.0012*** 1.848*** 6.346 .0327***

Referrer .108 *** 1.114 .0007*** .559*** 1.748 .0078*** .136** 1.146 .0024**

Other .032 1.033 .0000

Intercept -4.505*** .011 -5.124 *** .006 -4.978*** .007 -4.372*** .013

Observations 600,978 1,184,583 862,112 405,339

Note: * p < .05, ** p < .01, *** p < .001.

With regard to predictive accuracy, the second logit model performs better than the first

logit model. However, due to multicollinearity, the coefficient estimates for the last click are

not easy to interpret in an attribution context, because coefficients are indeterminate and the

standard errors of the estimates are large. As each journey includes at least one click, one of the

channel dummy variables for t=1 can be completely explained by a linear combination of the

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 56

other dummy variables, leading to perfect multicollinearity for t=1.8 For the other positions, the

results show different patterns in different channels across data sets that do not allow for

generalizations

In summary, our findings generalize prior research, showing that heuristics attribution

approaches tend to underestimate the contribution of selected firm-initiated channels (Li and

Kannan 2014; Xu, Duan, and Whinston 2014). For customer-initiated channels, the contribution

of SEA, and especially direct type-ins, is overestimated by the last click wins approach. For

other customer-initiated channels, additional factors such as industry and brand characteristics

seem to play a role, such that advertisers need to individually derive and verify detailed

implications, ideally on a more granular level. For example, it would be worthwhile to

separately analyze the contributions for keywords that contain product brands, retailer brands,

or no brand name at all.

2.6.2 Interplay of Channels

In addition to channel-level attribution, higher-order models offer a more detailed view of the

interplay of channels, which we illustrate using the second-order model in Table 11. Across all

data sets, the increase in overall purchase probability for most channels is highest right after the

START state, near the beginning of the journey9—which corresponds to the high share of one-

click journeys in our data.

8 As an alternative, it would have been possible to select a specific channel as a reference channel. We decided

against this solution, as this would have been counterintuitive for t=2 to t=4 where empty positions coded as zero are possible in journeys with fewer than four clicks.

9 A notable exception is retargeting, which explicitly targets customers who have previously visited the advertiser’s website (Lambrecht and Tucker 2013).

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 57

Table 11

Attribution Results for the Second Order Model by Data Set—Essay 1

Data Set 1 (Travel)

Current channel

Preceding channel

START SEA SEO Price Com- parison Display Newsletter Retargeting Affiliate Other

SEA 32.52% 14.27% 2.43% 0.34% 0.35% 0.48% 0.15% 0.84% 0.47%

SEO 7.78% 4.40% 4.75% 0.11% 0.07% 0.10% 0.06% 0.42% 0.08%

Price Comparison 3.07% 0.22% 0.05% 1.36% 0.03% 0.04% 0.01% 0.04% 0.04%

Display 0.62% 0.09% 0.02% 0.02% 0.42% 0.02% 0.01% 0.03% 0.01%

Newsletter 1.91% 0.27% 0.13% 0.02% 0.04% 0.99% 0.02% 0.07% 0.01%

Retargeting 0.34% 0.20% 0.06% 0.03% 0.02% 0.02% 0.24% 0.04% 0.01%

Affiliate 8.10% 2.61% 1.00% 0.23% 0.15% 0.20% 0.08% 5.31% 0.15%

Other 1.12% 0.27% 0.06% 0.01% 0.02% 0.02% 0.01% 0.03% 0.50%

Data Set 2 (Fashion)

Current channel

Preceding channel

START Type-in SEA SEO Price Com- parison Display Newsletter Affiliate Referrer

Type-in 16.85% 11.33% 2.21% 1.69% 0.04% 0.77% 2.49% 1.08% 0.35%

SEA 10.25% 1.52% 5.25% 1.97% 0.03% 0.35% 0.89% 0.64% 0.24%

SEO 6.25% 1.13% 1.93% 4.70% 0.02% 0.16% 0.43% 0.49% 0.39%

Price Comparison 0.07% 0.03% 0.03% 0.02% 0.03% 0.00% 0.01% 0.01% 0.00%

Display 0.62% 0.72% 0.33% 0.29% 0.01% 0.82% 0.23% 0.19% 0.06%

Newsletter 4.61% 1.84% 0.79% 0.50% 0.00% 0.22% 3.86% 0.18% 0.07%

Affiliate 4.24% 0.74% 0.52% 0.40% 0.02% 0.09% 0.12% 2.54% 0.14%

Referrer 1.26% 0.18% 0.16% 0.12% 0.01% 0.04% 0.07% 0.15% 0.19%

Data Set 3 (Fashion)

Current channel

Preceding channel

START Type-in SEA SEO Newsletter Retargeting Social Media Referrer

Type-in 13.26% 8.14% 1.49% 1.72% 0.06% 0.00% 2.36% 0.52%

SEA 11.07% 0.82% 3.74% 1.68% 0.09% 0.00% 1.00% 0.49%

SEO 12.23% 1.10% 1.61% 6.53% 0.05% 0.00% 0.62% 0.35%

Newsletter 0.71% 0.04% 0.04% 0.04% 0.29% 0.00% 0.03% 0.01%

Retargeting 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%

Social Media 12.74% 1.16% 0.63% 0.58% 0.03% 0.00% 8.27% 0.29%

Referrer 4.07% 0.00% 0.28% 0.36% 0.01% 0.00% 0.35% 0.77%

Data Set 4 (Luggage)

Current channel

Preceding channel

START Type-in SEA SEO Price Com-parison Retargeting Affiliate Referrer

Type-in 7.43% 4.30% 5.04% 0.83% 0.16% 0.20% 0.29% 0.17%

SEA 39.91% 1.37% 18.19% 1.72% 0.41% 0.28% 0.22% 0.19%

SEO 3.47% 0.38% 3.09% 2.26% 0.13% 0.03% 0.10% 0.06%

Price Comparison 1.20% 0.06% 0.57% 0.18% 0.72% 0.01% 0.02% 0.03%

Retargeting 0.10% 0.21% 1.01% 0.06% 0.03% 0.41% 0.00% 0.02%

Affiliate 0.28% 0.51% 1.21% 0.24% 0.18% 0.03% 0.92% 0.11%

Referrer 1.03% 0.08% 0.26% 0.03% 0.02% 0.00% 0.03% 0.19%

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 58

Sequences of identical channels show high removal effects in all data sets. For example,

in Data Set 1, affiliate preceded by affiliate has a percentage removal effect of 5.31%, whereas

the average removal effect for affiliate preceded by channels other than affiliate is only 0.63%.

For sequences of direct type-ins, which show comparably high removal effects, the results may

mirror Bowman and Narayandas’ (2001) finding on customers’ selective firm preferences

expressed by direct visits. However, the consistent importance of same-channel sequences

across data sets may further indicate idiosyncratic channel preferences for some users,

comparable to those found in multichannel relational communication (Godfrey, Seiders, and

Voss 2011). These channel preferences are also in line with related research findings showing

that previous visits through specific channels may reduce the cost for current visits through the

same channels (Li and Kannan 2014). Future research should investigate the existence of such

preferences, their antecedents, and their implications for multichannel online advertising.

Besides idiosyncratic channel preferences, we also find spillover effects between

channels both within and between channel groups. First, customer-initiated channels show

substantially higher removal effects if they are followed by another customer-initiated channel

than if they are followed by firm-initiated channels such as display and newsletter. With

increasing experience in online shopping, users anticipate the benefit and the occurred effort

associated with each online channel. In consequence, whenever a purchase need arises, more

directed users may sequentially use customer-initiated channels they are familiar with to reduce

search cost on their path to purchase (Li and Kannan 2014). Regarding the interplay of search

channels, prior research has argued that click propensities for paid and organic search results

are moderated by involvement (Jerath, Ma, and Park 2014). Our findings suggest that personal

channel preferences also play a role, as spillover effects from SEA to SEO are much stronger

than from SEO to SEA. Although consumers who have clicked on sponsored search

advertisements react to both paid and organic results, consumers clicking on organic search

results seem to generally prefer organic results. These results are consistent with Li and Kannan

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 59

(2014), but contrary to Yang and Ghose (2010), who find that organic search has stronger

spillovers toward paid search. Future research should investigate the role of channel preferences

and their interaction with involvement.

The interplay between search and direct type-in shows directional effects across data

sets, such that spillovers from both search channels, SEA and SEO, toward direct visits clearly

outreach the reversed effects. Users who visit the firm’s website via search are more likely to

return via direct type-in, potentially because preceding search contacts have shaped their

preferences toward a dedicated firm’s offering (Bowman and Narayandas 2001). For price

comparison, the results vary considerably across data sets. While for Data Set 1 (travel), price

comparison shows stronger spillovers toward both search channels, for Data Set 4 (luggage

retail), the corresponding spillovers are in the opposite direction. These variations between data

sets indicate industry-specific effects and may also be moderated by the customer’s price

sensitivity with regard to the respective product category (Mehta, Rajiv, and Srinivasan 2003).

Furthermore, we also find indications for spillover effects of customer-initiated channels

toward firm-initiated channels—and vice versa. These findings again generalize prior research

on multichannel customer journeys (Li and Kannan 2014; Xu, Duan, and Whinston 2014). The

removal effect for display, retargeting, social media, and newsletter is higher if these channels

are preceded by SEO, SEA, or direct type-in instead of other firm-initiated channels. Users

originating from customer-initiated channels may be more prone to response to pushed, firm-

initiated media than are users originating from firm-initiated channels.

The results further indicate moderate removal effects for firm-initiated channels

preceding customer-initiated channels. For instance, prior display visits show spillover effects

toward direct type-in and SEA, which is in line with previous research (Ilfeld and Winer 2002;

Li and Kannan 2014; Sherman and Deighton 2001). The results for newsletter show similar

patterns. Users who respond to newsletter-embedded links are more likely to return via SEA,

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 60

direct type-in or, above all, a newsletter which, except for direct type-in, corresponds to prior

research (Li and Kannan 2014).

According to our results, the spillover effects between firm-initiated channels are, by

and large, negligible. Except for some moderate spillovers between display and newsletter,

none of the removal effects between firm-initiated channels are noteworthy. These findings

correspond to extant research indicating that subsequent firm-initiated contacts reflect

stagnation in customer decision making and thus do not positively predict conversion

probabilities (Anderl, Schumann, and Kunz 2015).

On a channel level, it is interesting to analyze which channels induce the strongest

spillover effects—and which channels benefit most from spillovers from other channels. To

this end, we sum up the removal effects of all two-channel combinations that have channel C1

in the first position (corresponding to a column in Table 11), and compare them to the sum of

removal effects of all combinations ending with C1 (corresponding to a row in Table 11).

Across all data sets, the summarized removal effects of sequences starting with SEA, price

comparison, newsletter, and referrer consistently surpass the removal effects of sequences

ending in these channels. The opposite holds true for direct type-in and retargeting. A newsletter

exhibits strong long-term effects, yet only weak short-term effects (Breuer, Brettel, and Engelen

2011). Consequently, newsletter contacts may serve as facilitator for subsequent website visits

via other channels, namely direct type-in, SEA, SEO, and display, rather than as a direct sales

channel enabled by other channels (Li and Kannan 2014). By definition, retargeting focuses on

customers who have previously visited the advertiser’s website, and thus requires anteceding

channels. For SEO, display, and affiliate, the picture is mixed. As the diverging results for

customer-initiated channels may not be sufficiently explained by the channels’ contact origin

alone, a more differentiated view taking brand usage into account becomes appropriate (Anderl,

Schumann, and Kunz 2015). Whereas in some customer-initiated channels, such as direct type-

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 61

ins, customers actively use the advertiser’s brand to initiate the contact, other channels

(including price comparison), instead entail generic uses. Compared to generic customer-

initiated channels (e.g., price comparison), branded customer-initiated channels (e.g., direct

type-in) seem to benefit more from spillovers from other channels. Depending on the keyword

used, a search conveys both generic and branded uses; however, the proportion of generic

search queries in general outweighs branded search queries regarding both SEA and SEO

(Jansen and Spink 2009), such that—in aggregation—search channels appear closer to generic

uses, which supports the above generalization. A more detailed investigation of branded versus

generic contacts, as well as the role of brand awareness, would be a valuable extension of our

research. Besides increasing predictive performance, the application of higher-order models

thus enables a more detailed understanding of the interplay across channels, and allows

identifying several promising avenues for future research.

2.7 Discussion

In this study, we introduce an attribution framework based on first- and higher-order Markov

walks to examine the contribution of individual online channels and to shed light onto carryover

and spillover effects between online channels utilized by (potential) customers on their path to

purchase. The graph-based structure of our modeling approach mirrors the sequential nature of

individual-level customer journeys, enabling us to generate insights into the interplay of

channels in multichannel environments. Applying this model to four data sets from three

different industries, and comparing the results to existing heuristics, allows for cross-industry

generalizations as well as for industry-specific findings requested by recent literature

(Abhishek, Fader, and Hosanagar 2012; Li and Kannan 2014). To the best of our knowledge,

ours is the first study on attribution that seeks to derive empirical generalizations across

industries. Our results add to multichannel research on attribution (e.g., Li and Kannan 2014;

Xu, Duan, and Whinston 2014), research on channel strategy, and efficiency (Neslin and

Shankar 2009), and they bring forward valuable implications that help marketers to determine

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 62

channel contributions, recalibrate channel budgets, and draft targeting strategies. More

particularly, our research contributes to marketing theory and practice in the following ways.

First, we provide generalizable insights into the effectiveness of individual online

channels in a multichannel setting. We apply our graph-based framework to four empirical data

sets from three different industries, each entailing at least seven different online channels.

Whereas attribution results show considerable variation across data sets due to advertiser-

specific channel strategies, comparing our estimation results to two prominent attribution

heuristics (first- and last-click wins) and to two logit models enables us to derive empirical

generalizations on individual channel contributions. In this way, we add to extant attribution

literature (Abhishek, Fader, and Hosanagar 2012; Li and Kannan 2014; Xu, Duan, and

Whinston 2014). We find that firm-initiated channels are consistently undervalued by the

heuristic attribution approaches across industries. In contrast, the contribution of direct type-ins

and SEA is overestimated by these heuristic approaches across industries. For other customer-

initiated channels, especially SEO, additional factors such as industry and brand characteristics

seem to play a pivotal role. Given that differences between retailer and product brands as well

as keyword popularity influence consumers’ click behavior in search engines (Jerath, Ma, and

Park 2014), we conjecture that these results for SEO should not be generalized independent

from the corresponding brand. Thus, advertisers and scholars need to derive and verify detailed

implications for SEO, taking the individual brand into consideration. If the granularity of data

allows, isolating branded from generic search terms might be a useful extension to derive

further generalizations. Our framework would be well-prepared to handle data on different

levels of granularity, making it well-suited for this extension.

Second, higher-order models allow new insights into the interplay of channels along the

customer journey. Using our framework, we generalize prior findings indicating that a majority

of channels exhibit idiosyncratic channel carryovers (Li and Kannan 2014). The consistent

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 63

relevance of homogeneous channel sequences across data sets may mirror extant channel

preferences for some users, which is also shown in multichannel relational communication

(Godfrey, Seiders, and Voss 2011). Advertisers anticipating these preferences may improve

targeting measures for these customer segments. For search channels, paid search touches are

often followed by both paid and unpaid search contacts, while their unpaid equivalents are

merely followed by unpaid search contacts. These results endorse extant research (Li and

Kannan 2014); however, they contradict industry-specific findings by Yang and Ghose (2010).

Apparently, for some users who have previously accessed a website through paid search results,

SEO seems a valid choice if they intend to return. Thus, advertisers may benefit from more

integrated SEO and SEA strategies.

Moreover, we observe spillover effects within, as well as between, channel categories.

Customer-initiated channels exhibit substantial removal effects if they are followed by other

customer-initiated channels, whereas—in line with prior research (Anderl, Schumann, and

Kunz 2015) — spillover effects between firm-initiated channels are by and large negligible.

Spillovers between customer-initiated and firm-initiated channels are more selective and

are carried out at a moderate level in both directions. For instance, spillovers from SEA and

SEO toward direct type-in traffic consistently outreach the reversed effects, suggesting cross-

industry generalizability. Yet the results for price comparison vary considerably, indicating

industry-specific effects that may also be moderated by the customer’s price sensitivity (Mehta,

Rajiv, and Srinivasan 2003). Prior firm-initiated touches may further affect customer-initiated

contacts. Users who respond to display advertisements, retargeting, or newsletters tend to return

via SEA or direct type-in, confirming previous findings (Ilfeld and Winer 2002; Li and Kannan

2014; Sherman and Deighton 2001). Advertisers may therefore tactically utilize firm-initiated

channels to guide customers toward the advertised offering.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 64

Our findings are also useful at the channel level, particularly regarding the ability of

each channel to induce spillover effects—and to benefit from them. Across all data sets, we

observe that the removal effects of sequences that begin with SEA, price comparison,

newsletter, and referrer consistently outreach the removal effects of sequences ending in these

channels. For direct type-in and retargeting, the opposite holds true. To better explain these

results for customer-initiated channels, we conjecture that a more differentiated view that

includes branded channel usage may help to explain the differences (Anderl, Schumann, and

Kunz 2015). From that perspective, branded customer-initiated channels (e.g., direct type-in)

seem to benefit more from spillovers from other channels than do generic customer-initiated

channels (e.g., price comparison). The benefit to advertisers is that this consideration may result

in a better understanding of how channels catalyze one another, which is, in turn, pivotal to

creating integrated multichannel strategies.

Third, we develop a novel attribution variant that adds to existing attribution modeling

techniques (Abhishek, Fader, and Hosanagar 2012; Berman 2015; Haan, Wiesel, and Pauwels

2013; Kireyev, Pauwels, and Gupta 2013; Li and Kannan 2014; Xu, Duan, and Whinston 2014)

by representing multichannel online customer path data as Markov walks in directed graphs. In

addition to the first-order base model, we introduce higher-order Markov graphs, in which the

present depends on the last k observations. We calculate a property, namely the removal effect,

which is defined as the change in probability of reaching the CONVERSION state from the

START state when si is removed from the graph, to derive channel contributions. First- and

higher-order models can be used to measure channel contribution in a multichannel setting, but

higher-order models, in particular, allow investigation of channel sequences and spillovers

between channels. Our findings support previous research (Abhishek, Fader, and Hosanagar

2012; Berman 2015; Li and Kannan 2014), indicating that one-click attribution measures (such

as the last click wins approach), are neither suited to derive the actual contribution of individual

channels nor to determine carryover and spillover effects across channels.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 65

Finally, our framework provides a useful approach to a number of explicit problems that

online marketers confront. Because our model captures the contribution of each online channel,

it may help to calibrate online channel budgets and facilitate moves toward optimal budget

allocation (Raman et al. 2012). If the budget share of a channel is higher than its actual

contribution, advertisers should readjust their budget splits. Moreover, using our graph-based

framework, advertisers can more accurately calculate the conversion probability Eventual

Conversion(si) of a customer, given his or her previous customer journey. This information can

be used to support state-of-the-art applications such as real-time bidding decisions in ad

exchanges. In addition, knowledge of the state of each customer allows advertisers to better

predict conversion propensity through all potential channels. By more fully anticipating the

users’ channel preferences and the potential cost incurred by each channel, advertisers may be

able to reach customers through more efficient channels.

2.8 Outlook

Our research has several limitations that may serve to stimulate research on attribution and

online marketing effectiveness. We used four data sets from different industries, but some

findings may still be company-specific. The customer journeys in these data sets were short on

average, including a high number of one-click journeys. However, sophisticated attribution is

not required for journeys consisting of just a single click. In that case, both the last click wins

and the first click wins heuristics deliver objective results that would satisfy our criteria,

whereas longer journeys increase advertisers’ need to understand channel contributions.

Furthermore, although our framework should be well suited to handle such information, our

data sets do not include views and exposures to offline channels. We therefore recommend

applying this framework to further industries, and to integrate not just clicks but views and

offline channels.

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 66

To further evaluate the effectiveness of online marketing channels, companies should

consider revenues and profits from conversions, and—in the potential of a second step—the

customer lifetime value (CLV) of customers acquired. As Chan, Wu, and Xie (2011) show,

customers acquired through different online marketing channels differ in CLV. Our graph-

based approach is well suited for extension into such areas with additional data, so that further

research should include this information to advance attribution. The timing between contacts

along the customer journey would be another interesting extension of the research.

The attribution problem is, by definition, endogenic; it measures the relative

effectiveness of channels in a given setting (Li and Kannan 2014), so the results are conditional

upon a number of management decisions such as channels used, budget limits per channel, or

number of advertising creatives employed. Therefore, developing an optimal budget allocation

remains an iterative process. Correct attribution, however, is a necessary prerequisite for

managers to optimize their budget decisions. Subject to the availability of longitudinal data,

attribution results calculated using our framework could also serve as a basis for developing

optimization algorithms.

With regard to prediction, this study focuses on conversion probabilities. In addition to

predicting conversions within individual subjects, our methodology may also support prediction

of the rest of the journey (until conversion), given an initial browsing path, or allow advertisers

to anticipate the subsequent online channel exposure that is statistically best suited for

catalyzing conversions. The graph-based structure of our model, in combination with the panel

nature of our individual-level customer journey data, make our framework well-suited for this

extension.

Finally, a strict causal interpretation of customer journeys is difficult, because

alternative explanations may exist for correlations between conversions and advertising

exposure. Some channels, such as retargeting, explicitly try to target customers who have a

Essay 1—Lessons Learned from Graph-Based Online Attribution Modeling 67

higher propensity to purchase (Lambrecht and Tucker 2013). Even without special targeting,

observed correlations might be due to selection effects, such as an activity bias (Lewis, Rao,

and Reiley 2011). Establishing such a causal relationship would require large-scale field

experiments with randomized exposure. Such experiments are challenging to implement in

practice, especially in multichannel settings, but comparing our attribution modeling

framework against results of the experiments would be a valuable direction for building on our

research.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 69

3 Patterns that Matter: How Browsing Click Patterns as Micro-Journeys

Influence Customer Conversions

Ingo Becker, Marc Linzmajer, Florian von Wangenheim

Web users leave manifold traits while browsing and purchasing online, yet extracting

implications from these traits remains challenging. While research has proposed valuable

approaches to translate (multi)channel exposure into purchase propensity, the link between

inter-exposure times of online marketing contacts and online purchases remains unexplored.

Assuming that click sequences within short time intervals indicate purchase intent, the authors

introduce a novel concept, the “micro-journey”, which implements a time component into the

user’s browsing history operationalizing timing between subsequent clicks to uncover purchase

intentions. Grounded in flow theory, the micro-journey conceptualizes “focused attention” as

an indicator of users’ concentrated browsing sessions by linking clicks chronologically close to

each other. Applying a proportional hazards model to four customer-level data sets shows that

micro-journeys are well suited to improve the prediction of user conversions. Moreover, the

authors derive specific features of the micro-journey that support in identifying users prone to

convert—also analyzing direct and later/procrastinated purchases. Thereby, this research entails

a novel, valuable construct adding to advertisers’ increasing demand for a holistic analysis of

online marketing effectiveness.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 70

3.1 Introduction

As shopping went online, it opened up a completely new world of touchpoints. Before making

an online purchase decision, a customer may engage with marketing activities over several days

and through different media channels such as email, display advertisements, paid search

advertisements, social media, and direct visits to a website—“and all of these interactions can

play a role in the final sale” (Google 2013).

Contemporary research considers clickstream behavior10 as one of the most important

sources to predict online purchasing behavior (Moe 2003; Van den Poel and Buckinx 2005).

The richness in this data reveals, for instance, that some consumers repeatedly click on a display

advertising before purchasing online, while others prefer to utilize organic and paid search,

prompting companies to call for marketing impact models based on individual-level, single-

source data that help identify optimal marketing budget allocation (Rust et al. 2004). Due to

recent technological advancements, researchers gain access to such data sets mirrored in a

growing body of clickstream based literature on channel effectiveness (e.g., Ghose and Yang

2009; Li and Kannan 2014; Rutz and Bucklin 2011). That said, less is known about how time-

related information in (past) browsing behavior such as time-tagged users’ click patterns

influence conversions or help to forecast future conversions. Are consumer clicks random, or

might certain click patterns predict whether a consumer becomes a customer?

This article examines how time-associated clickstream patterns help to better

understand online purchasing behavior. Building on flow theory, we focus on a particular click

pattern that we call a micro-journey, which consists of aggregated, contiguous user clicks within

defined short time intervals. While prior approaches to aggregated data generally make no

assumption about the click sequence, some recent approaches to individual-level data focus on

10 Clickstream behavior refers to the transaction log that customers establish as they move around the web

(Chatterjee, Hoffman, and Novak 2003).

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 71

(single) channel effectiveness (e.g., Ghose and Yang 2009; Manchanda et al. 2006), channel

spillover effects (e.g., Rutz and Bucklin 2011), and, more sparsely, marketing effectiveness in

multichannel settings (Li and Kannan 2014), partly implementing the chronological click order.

These studies have paved the way for a better understanding of online purchase decisions and

interpreting the clickstream. Yet, considering the exact timing of clicks and how they are

embedded in the overall customer journey, may further complement these prior findings. As an

example of the more accurate treatment of timing, a journey of five clicks distributed over five

days should be treated differently to five clicks occurring within two hours, as they may imply

varying user intentions. Applying the concept of the micro-journey helps to disentangle time

factors of the click exposure and follows the browsing path of a particular user more precisely

(Figure 6).

Figure 6

Prior Clickstream Approaches and the Concept of the Micro-Journey—Essay 2

Controlling for several important drivers of the online clickstream and applying a

proportional hazards model on four distinct data sets, we examine how these specific click

patterns contribute to explaining final customer conversions. In particular, we assess the micro-

journey’s suitability in identifying users prone to convert (Model 1), evaluate the influence of

characteristics of the micro-journey (Model 2), and, in a dichotomous sequential model, analyze

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 72

which of these characteristics indicate direct purchases (succeeding the micro-journey) and

later/procrastinated purchases (Model 3).

First, we transfer a component of flow theory into an online clickstream context by

introducing the concept of the micro-journey that translates click patterns into purchasing

propensity. Our findings suggest that the micro-journey—based on an environmental

psychology approach—is a well-suited predictor and provides novel insights into how a time-

associated construct affects conversions (Model 1). Thereby, our research entails a novel,

valuable construct adding to advertisers’ increasing demand for a holistic analysis of online

marketing effectiveness (Novak, Hoffman, and Yung 2000; Yadav and Pavlou 2014).

In addition, considering various micro-journey characteristics (e.g., the relative

position) enables us to contribute novel insights into online marketing effectiveness of

touchpoint interrelations in multitouch/multichannel environments. While extant multichannel

research primarily illuminates channel interactions such as channel carryover/spillover effects

(Li and Kannan 2014; Rutz and Bucklin 2011), we add a new lens by combining the modeling

of off-site clickstreams with exposure time given in the user’s transaction log. We find that the

time distance may substantially influence the outcome of channel interactions. For instance,

spillover effects between navigational and informational channels show remarkable differences

when they apply in the short term versus the longer term (Model 2). In this way, we contribute

to the existing literature that addresses single and multiple channels but is based on aggregated

data, and overall clickstream-based models—models that do cover clicks or channel categories

but that do not cover time.

Moreover, the dichotomous sequential model allows for inferences on the effects of

micro-journey characteristics and purchase timing adding to research on time-related aspects in

purchase decisions and procrastination (Breuer, Brettel, and Engelen 2011; O’Donoghue and

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 73

Rabin 1999). We find indications for the existence of distinct user groups differing in their

mode of information acquisition, browsing behavior, and purchase decision-making (Model 3).

Along these lines, we find that (sequential) Cox proportional hazards models can be a

widely understandable alternative for analyzing the clickstream when trying to understand

converting customers. Using four large-scale data sets across three different industries, we

enable cross-industry comparisons and find that insights on channel effectiveness cannot be

generalized from results obtained from a single data set. Thereby, this study corresponds to the

demand for a widely understood and flexible framework (Rust et al. 2004) and provides a new

perspective analyzing individual-level, single-source path data (Hui, Fader, and Bradlow 2009).

Finally, our research may support in improving bidding strategies in real-time bidding

applications of advertising exchanges. As our results apply to known as well as unknown users

and help to identify users more prone to convert, advertisers may benefit from this knowledge

in the bidding decisions enabling them to more precisely target relevant customer segments.

3.2 Conceptual Development of the Micro-Journey

With the increasing number of online touchpoints that have found their way into our daily lives,

consumers invest considerable resources in online clicks. With regard to purchasing behavior,

time is a pivotal resource, so that consumers must make decisions regarding their use of time

in the purchase of goods and services (LeClerc, Schmitt, and Dubé 1995). A useful model from

environmental psychology that integrates a subjective feeling of time as a resource is the

concept of flow (Csikszentmihalyi and Csikszentmihalyi 1988; Csikszentmihalyi 1977).

3.2.1 The Concept of Flow

The concept of flow can be used to describe traditional “converting” customers with

psychological measures and is defined as “the holistic sensation that people feel when they act

with total involvement” (Csikszentmihalyi 1977, p.36). When people are in flow, they “shift

into a common mode of experience when they become absorbed in their activity. This mode is

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 74

characterized by a narrowing of the focus of awareness, so that irrelevant perceptions and

thoughts are filtered out, by loss of self-consciousness, by a responsiveness to clear goals and

unambiguous feedback, and by a sense of control over the environment” (Csikszentmihalyi

1977, p.72). In recent years, flow has also been studied in the context of information

technologies and computer-mediated environments, and has been recommended as a possible

metric for the online consumer experience (Ghani and Deshpande 1994; Hoffman and Novak

1996; Koufaris 2002; Novak, Hoffman, and Yung 2000; Webster, Trevino, and Ryan 1993).

Especially one component of the flow concept is relevant for conceptualizing consumers’

micro-journeys, namely, the concentration/attention focus. This component can be used as a

valid metric for measuring and analyzing the online consumer experience, as it aggregates

determinants of online shopping behavior related to time, cost and benefits, shopping contexts,

and personality traits11 to identify future converting consumers. But how can we extract this

psychological construct from the online consumer clickstream a priori?

3.2.2 The Concept of the Micro-Journey

To achieve the goal of extracting the psychological construct of concentration/attention focus,

the study of clickstream behavior shows what the consumer is looking for at each moment, as

the individual clickstream represents the behavioral responses of web consumers. An individual

who is in “flow” must concentrate on the activity. Therefore, concentration has been a

significant correlate or measure of flow in offline environments. Online consumer behavior also

demonstrates the pivotal role of concentration. Because web customers often have limited time

and information processing (e.g., Blackwell, Miniard, and Engel 2005), along with the interplay

of experiential and goal-directed behavior on the web (Hoffman and Novak 1996), they may

demonstrate a short attention span regarding the completion of a task like shopping a product—

11 Note that scattered findings on the named determinants of online shopping behavior already exist (for an

overview, see, for instance, Bosnjak, Galesic, and Tuten 2007; Moe 2003).

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 75

which would have a natural tendency to end with uncompleted purchases. Accordingly,

increased concentration is essential for greater customer conversion. When consumers shop in

the physical world, they must allocate most of their attention to that task. They must walk or

drive to stores, look through the product selections, interact with sales people, and make

purchases. Their ability to perform other activities while shopping is relatively limited

(Koufaris 2002). Using a computer or mobile device, customers are more prone to be distracted

from their shopping task regardless of whether they are at home (e.g., children, or television),

at the office (e.g., work, or colleagues), or on the move (e.g., traffic, or other people).

Furthermore, online activities such as email, instant messaging, or other websites can also divert

an online consumer’s attention. Consumers can take advantage of their computers’ capabilities,

carrying out multiple tasks in addition to shopping and, importantly, advertisers today can sort

through this multitasking to measure and anticipate their buying intentions or mere browsing

behaviors. Yet, concentration as an action of flow has been found to positively influence the

overall experience of computer users (Novak, Hoffman, and Yung 2000), and their intention to

use a system repeatedly (Webster, Trevino, and Ryan 1993). As well, studies have shown that

interruptions limiting concentration tend to reduce web users’ satisfaction with online shopping

(Xia and Sudharshan 2002). We therefore expect that high concentration, measured with and

operationalized as a micro-journey, would have a positive impact on conversion rates.

Applying the theoretical link of flow to purchase transactions, online customers are

expected to concentrate and focus along their path to purchasing a product. These characteristics

can be mapped from their clickstream behavior. Enormous potential exists in studying

individual behavior while navigating from page to page (Moe 2003). Hoffman and Novak

(1996) explained the customer online experience through the concept of flow. Sismeiro and

Bucklin (2004) later developed a model to predict online purchases by connecting the purchase

decision to browsing behavior such as (repeated) website visits, on-site page views, and time

spent per view; however, the time distance between subsequent visits (clicks) remains

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 76

unrecorded (Bucklin and Sismeiro 2009; Sismeiro and Bucklin 2004). To integrate a direct link

from clicking to buying, we build on a customer’s tendency to exhibit very focused browsing

patterns, indicative of the intense goal-directed (i.e., concentrated) motivation of the customer

to purchase a product online (e.g., Moe 2003). We expect that this basic psychological

mechanism antecedes most purchase decisions in an online context and can be represented by

the clicking behavior of customers—which we discuss in the context of a micro-journey. Study

of the micro-journey produces a snapshot of the consumer’s path to completed purchases and

helps to explain a significant amount of conversion rates’ variance.

To achieve the goal of identifying converting customers based on their first clicks, we

model individual-level multichannel customer journey data and analyze click patterns as

aggregated micro-journeys using Cox proportional hazards models (Cox 1972). In line with the

theoretical idea of an attention focus/concentration link, and comprising multiple clicks, micro-

journeys represent intense customer interactions that may occur when users are browsing in a

focused manner—for example, evaluating a purchase option or acquiring product information.

3.3 Empirical Setting and Extraction of the Micro-Journey

3.3.1 Clickstream Data

Our research is based on four clickstream data sets provided by online advertisers, in

collaboration with a multichannel tracking provider, in the areas of fashion retail, luggage retail,

and travel. The record of clickstream data for each user’s Internet activity traces the navigation

path (Bucklin and Sismeiro 2009). For each visit to the advertiser’s website during the

observation period, the data include detailed information about the source of the click, an exact

timestamp, and a unique user identifier that allows for relating clicks to individual users. Clicks

either represent a direct behavioral response to an advertising exposure, or result from the user

entering the advertiser’s URL directly into the browser, so these sources comprise all online

marketing channels as well as direct type-ins. We also tracked whether each visit was followed

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 77

by a conversion (purchase transaction) enabling us to analyze successful journeys ending with

a conversion as well as journeys that do not lead to a conversion within 30 days of the last

exposure. The number of online channels for advertisers differ, from nine to eleven, and

include: affiliate, display, newsletter, price comparison, referrer, retargeting, paid search

generic, paid search branded, unpaid search generic, unpaid search branded, social media, and

direct type-in.12 Branded search contacts are identified as searches in which the user mentions

the brand name of the advertiser, including searches with misspellings (Jansen, Booth, and

Spink 2008).We collect data at the cookie level, identifying individual customers—or more

accurately, individual devices. Though the use of cookie data presents several limitations, such

as an inability to track multi-device use, or bias due to cookie deletion (Flosi, Fulgoni, and

Vollman 2013), cookies remain the industry standard for multichannel tracking. All four

advertisers providing data sets are pure online players, so we can exclude online/offline cross-

channel effects.

We controlled the data sets for tracking errors, such as double-count of single clicks and

non-human click behavior (by web bots). Overall, we excluded approximately 2.5% of all

clicks. For our main analysis across all models, we selected the data set with the largest number

of clicks and journeys—in this case a fashion retailer—which allowed us to explore within and

across industry generalizations and comparisons. Application of these models to the three

remaining data sets verify robustness of the results—which we report in the Appendix. Our

final primary data set includes 1,635,724 user contacts contained in 1,184,582 individual

customer journeys, with a total conversion rate of 0.86%. Our data was collected over a major

tracking period of 31 days, though the collection period may exceed this time frame either

because a user journey continues (with browsing not exceeding 30 days between two

12 We provide definitions of the online channels analyzed in the Appendix section.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 78

subsequent clicks), or because a journey began before our major tracking period reached into

it. Using this method, the maximum time of our collected clicks was 78 days. Table 12 presents

detailed descriptions of all four data sets.

Table 12

Descriptive Statistics of the Data Sets—Essay 2

Data Set 1 Data Set 2 Data Set 3 Data Set 4

Industry Fashion Fashion Luggage Travel

Number of different channels 11 11 10 9

Number of different channels analyzed 10 9 9 9

Number of journeys 1,184,582 862,114 405,343 600,873

Journey length in clicks 1,38 1,30 1,48 2,30

(1,88) (1,23) (1,28) (5,20)

Number of conversions 10,153 16,201 8,117 9,861

Journey conversion rate 0.86% 1.88% 2.00% 1.64%

Number of clicks 1,635,724 1,122,838 601,417 1,380,190

Thereof Affiliate 754,355 2 1,699 36,487

Thereof Display 10 - - 787,743

Thereof Newsletter 79,453 12 8 17,204

Thereof Price Comparison 2,205 - 15,003 90,302

Thereof Referrer 70,394 123,952 15,043 -

Thereof Retargeting 20,059 31 7,011 5,697

Thereof SEA Generic 120,542 119,374 442,617 287,346

Thereof SEA Branded 50,982 37,493 8,284 54,481

Thereof SEO Generic 123,850 330,081 62,179 73,492

Thereof SEO Branded 75,030 42,380 3,866 27,438

Thereof Direct Type-in 338,844 213,053 45,707 -

Thereof Social Media - 241,462 - -

Thereof Other - 14,998 - - Note: Standard deviations are in parentheses. The different number of channels by advertiser derives from the individual channel propensity and selection of each advertiser. Rare channels are removed from the models as they are underrepresented and insignificant (DS 1: Display; DS 2: Affiliate, Newsletter; DS 3: Newsletter).

3.3.2 Extraction of the Micro-Journey

We use the information of the precise time tag associated with each individual click to construct

conglomerates of successive clicks occurring within pre-defined time intervals, which we call

micro-journeys. In particular, as long as two contiguous clicks do not exceed an interval of 30

minutes, they add up to a micro-journey, which may therefore comprise multiple clicks, yet,

not less than two clicks (Figure 7). Applying this particular time interval mirrors prior literature

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 79

on online purchasing behavior jointly approximating an average website visit duration to 30

minutes of interested customers (Berendt et al. 2001). Li and Kannan (2014) investigated

conversions in a multichannel online marketing environment. Their field study treats all

“contiguous visits through the same channel within 30 minutes […] as a single visit” (p. 46).

Catledge and Pitkow (1995) measured mean inactivity time within a site to be 9.3 minutes;

adding 1.5 standard deviations, they determined a 25.5 minute cut-off for the duration of a visit.

Most web applications use this figure rounded up to 30 minutes as the maximal length of a

session (Cooley, Mobasher, and Srivastava 1999; Spiliopoulou and Faulstich 1999). The time

that users spend reading and processing the contents of any single page varies within certain

limits. If a long time elapses between one request and the next, it is likely that the latter request

represents a new visit. In an empirical study of the impact of web experiences on virtual buying

behavior, Constantinides and Geurts (2005) give their participants a maximum time of 30

minutes to complete an online search and make a decision about a product and the online

vendor, confirming the 30-minute heuristic for the length of a single customer journey online.

Figure 7

The Micro-Journey versus Single Clicks—Essay 2

Based on the observation of the empirical evidence in the relevant literature, when two or more

sequential clicks are observed within a 30-minute time period, we group those clicks as a micro-

journey. We do so with the understanding that users who interact with an advertiser within a

30-minute period reflect a focus for gathering information on a specific product, or for an intent

to conclude a purchase decision. If users spend more than 30 minutes on a session, they will

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 80

likely not complete that session, but will (if at all) continue their purchase path in a later

browsing session.13 Consequently, a unique customer journey may range from no micro-

journey at all, up to multiple micro-journeys, depending on the user’s clicking behavior. Table

13 illustrates descriptions of the journeys that have micro-journeys. Journeys with micro-

journeys show a higher conversion rate compared to all journeys, and especially as compared

to journeys with single clicks only (see Table 12 and Table 13).

Table 13

Descriptive Statistics of the Journeys with Micro-Journeys—Essay 2

Data Set 1 Data Set 2 Data Set 3 Data Set 4 Industry Fashion Fashion Luggage Travel Total number of micro-journeys 106,812 92,813 80,993 243,615 Journeys with micro-journeys 90,333 86,508 73,560 151,212

Thereof number of conversions 3,533 5,192 3,053 3,503 Thereof conversion rate 3.91% 6.00% 4.15% 2.32%

Journeys with 1 micro-journey 81,087 82,083 67,780 118,441 Thereof number of conversions 2,204 3,741 2,485 2,624 Thereof conversion rate 2.72% 4.56% 3.67% 2.22%

Journeys with 2 micro-journeys 6,274 3,466 4,674 16,618 Thereof number of conversions 623 871 429 511 Thereof conversion rate 9.93% 25.13% 9.18% 3.07%

Journeys with 3 micro-journeys 1,608 587 797 6,249 Thereof number of conversions 259 288 102 204 Thereof conversion rate 16.11% 49.06% 12.80% 3.26%

Journeys with >3 micro-journeys 1,364 372 309 9,904 Thereof number of conversions 447 292 37 164 Thereof conversion rate 32.77% 78.49% 11.97% 1.66%

3.3.3 Characterizing the Micro-Journey

The first statistical model is focused on the general potential of micro-journeys to support

forecasting of purchase events. The intent of the second and third models, however, is to

describe the micro-journey in more depth, and to show how its characteristics might affect

purchase decisions. We therefore cascade the micro-journey covariate into further predictors

that represent its characteristics.

13 To exclude effects related to the interval pre-definition, we analyzed various time values between successive

clicks from 15 minutes to a maximum of 240 minutes. Results between 20 and 60 minutes were stable across data sets; 30 minutes seemed most appropriate.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 81

The number of micro-journey clicks. Several empirical studies have shown that the number of

advertising exposures increases advertising response (e.g., Pedrick and Zufryden 1991; Tellis

1988). Moreover, Klapdor et al. (2015) empirically verify that journey length, measured in

clicks, is well-suited to better identifying converting customers. The underlying mechanism

may relate to flow theory, such that more clicks may express a higher degree of telepresence

and time distortion, a continuation of focused attention (Novak, Hoffman, and Yung 2000). We

transfer these findings into the concept of micro-journeys and implement their length, measured

in clicks, as additional predictor.

The number of micro-journey channels. Various studies have demonstrated positive effects

associated with exposure to advertising messages on multiple channels, compared to repeated

exposure on a single channel. For instance, Wiesel, Pauwels and Arts (2011) find evidence for

bidirectional cross-channel synergies between offline and online channels that help to better

allocate marketing resources In an online context, Klapdor et al. (2015) show indication of a

positive link between channel exposure and purchase likelihood.

The micro-journey duration. We defined 30 minutes as the maximum time interval between

two subsequent clicks to form a micro-journey (Berendt et al. 2001). However, micro-journeys

may be shorter or exceed this period if they contain more than two clicks. Although we know

from prior research that the time duration of the full journey does not affect users’ conversions

(Klapdor et al. 2015), this finding may not necessarily hold for a time-delimited period such as

the micro-journey. Novak, Hoffman, and Yung (2000) show that time-associated factors—

telepresence and time distortion—positively influence flow, which in turn defines a compelling

browsing experience. In consequence, a highly attentive user may be less time-focused,

indicating that the duration of the micro-journey may influence its impact on purchases.

Assuming that the micro-journey is an integral part of the information acquisition process and

expresses the consumer‘s interest in a product category, it may directly link to purchases, as

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 82

information acquisition constitutes a major portion of duration time in purchase decision

processes (Putsis and Srinivasan 1994).

The relative position of the micro-journey. The relative position of the micro-journey within

the total journey can express browsing state, progress, or finalization in the choice set formation

process (Shocker et al. 1991). Position may indicate purchase involvement (Beatty and Smith

1987), interest in the product category (Putsis and Srinivasan 1994), and the process of

information acquisition (Murray 1991). Consequently, the micro-journey’s position may be an

indicator of the user’s current position within the decision process expressing the (shorter or

longer) duration until purchases (Putsis and Srinivasan 1994). As the journey length in clicks

has been found to be associated with progression and purchase propensity, as opposed to the

time duration of the total journey (Klapdor et al. 2015), we model the relative position of the

micro-journey as the relative proportion of its first click and the overall journey length measured

in clicks.

Browsing goal. Moreover, to uncover and describe the user’s (hidden) browsing intention in

ways that better predict conversion events, we rely on a taxonomy developed in information

retrieval research and follow a channel categorization approach that is constructed to describes

the user’s browsing goal (Broder 2002; Jansen, Booth, and Spink 2008). For a user who wants

to gather information by reading a website, the browsing goal is assumed to be of informational

nature (Rose and Levinson 2004), and of navigational nature if the user aims to visit a specific

website (Broder 2002). Categorization approaches help to disentangle complex channel settings

by classifying and aggregating innate channels.14 To allow for inferences in the underlying

choice set formation process (e.g., Hauser and Wernerfelt 1990; Roberts and Lattin 1991, 1997),

we implement the switches from informational to navigational contacts (and vice versa) as

14 In Essay 3 (Section 4.2.2), we provide further information on different category approaches including a detailed

table of established channel-category links (Table 21).

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 83

subsequent clicks in different channel categories that indicate progress in browsing behavior:

A user who first visits a retailer’s website through informational channels and later returns via

a navigational channel may have narrowed down the choice set in the deliberation process, and

in returning to the specific website follows a dedicated purpose (e.g., purchase).

Table 14 shows micro-journey characteristics in converting versus non-converting

journeys with exactly one micro-journey.15 While the number of clicks and the number of

channels are comparable, micro-journeys that precede conversion events seem to be of longer

duration, are positioned later in the overall journey, and more often comprise category switches.

Table 14

Descriptive Statistics of Micro-Journeys in Converting and Non-Converting Journeys (with

one Micro-Journey)—Essay 2

Data Set 1 Data Set 2 Data Set 3 Data Set 4 Industry Fashion Fashion Luggage Travel Number of non-converting micro-journeys 78,883 78,342 65,295 115,817

Micro-journey length in clicks 2.43 2.30 2.26 2.56

(1.24) (0.95) (0.73) (1.97)

Micro-journey number channels 1.25 1.15 1.23 1.16

(0.46) (0.36) (0.43) (0.37)

Micro-journey duration in seconds. 302 270 427 333

(391) (366) (441) (415)

Micro-journey relative position 0.08 0.05 0.07 0.09

(0.21) (0.17) (0.19) (0.22) Micro-journeys with navigational switch 22.5% 10.2% 3.6% 7.2% Micro-journeys with informational switch 78.6% 83.6% 87.2% 77.6%

Number of converting micro-journeys 2,204 3,741 2,485 2,624

Micro-journey length in clicks 2.49 2.64 2.32 2.29

(1.34) (1.41) (0.85) (0.71)

Micro-journey number channels 1.44 1.18 1.36 1.32

(0.56) (0.4) (0.52) (0.49)

Micro-journey duration in seconds 495 398 623 613

(499) (447) (487) (531)

Micro-journey relative position 0.31 0.21 0.14 0.25

(0.31) (0.29) (0.26) (0.31) Micro-journeys with navigational switch 38.0% 14.3% 15.1% 30.9% Micro-journeys with informational switch 49.0% 64.2% 76.4% 35.9%

Note: Standard deviations are in parentheses. Sample includes micro-journeys from journeys with exactly one micro-journey. Since the advertisers’ channel mix differs to a certain extent, relative shares of navigational and informational switches are subject to channel variation.

15 We include only user journeys with exactly one micro-journey, as the co-existence of multiple micro-journeys

may provoke biases.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 84

3.4 Model Development

3.4.1 General Model Formulation

We develop three models to deepen our investigation of the effectiveness of micro-journeys

and to control for effects that are well established in marketing research. First, we explore

whether micro-journeys are an appropriate indicator to foresee customer purchases. Our second

model details the micro-journey to identify the micro-journey characteristics that in particular

affect conversion events. The third model is of sequential nature in order to investigate the

characteristics of micro-journeys that affect direct or later purchases.

3.4.2 The Micro-Journey as Predictor (Model 1)

In our first model, we empirically measure the effect of micro-journeys as a specific pattern of

concentrated browsing behavior, and control for further effects that are known to influence

purchase likelihood. Figure 8 plots the conversion rate of customer journeys that contain micro-

journeys against the ones that do not contain micro-journeys. This initial evidence suggests that

journeys, including the specific browsing pattern of micro-journeys, are more likely to conclude

in a purchase decision, in contrast to journeys of users without this browsing behavior.

Figure 8

Conversions of All Journeys versus Journeys with Micro-Journeys—Essay 2

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 85

Our first analysis in Figure 8 does not control for other effects that can influence

purchase behavior, such as channel exposure, navigational or informational contacts, or journey

length. Therefore, we apply a Cox proportional hazards model (Cox 1972), which has been used

in preceding studies on online marketing effectiveness (Lambrecht and Tucker 2013;

Manchanda et al. 2006). In general, the dependent variable in a proportional hazards model is

the time T until the occurrence of an event, which in our case is a binary conversion event.

Furthermore, the proportional hazards model, common in medical sciences, allows for (right)

censoring,16 reflects sequentially occurring covariates (e.g., medication, advertisement), and

measures covariates‘ effect on the time to an event (e.g., conversions). The proportional hazards

model, therefore, is well suited for analyzing clickstream data, and has an advantage over binary

regression models such as logit or probit models (Collett 2015).

The proportional hazards model formula defines the hazard (event) at time t as the

product of two quantities. First, the underlying baseline hazard, h0(t), describes, at the baseline

level of covariates, the risk of an event per time unit. The second quantity consists of the

exponential expression e to the linear sum of βiX i, where the sum is over the p predictor X

covariates. The latter describes the responsive effect of the explanatory or predictor covariates

on the hazard. A relevant feature of this definition is that the baseline hazard is a function of

time t, but excludes the explanatory X covariates. In contrast, the exponential expression shown

above involves the X covariates, but excludes t. Thus the explanatory X covariates are defined

as time-independent. We select a Cox proportional hazards model as our standard model

because of its property of a baseline hazard, h0(t), as an unspecified function, which is favorable

for our data sets, as we are unaware of a particular form of the hazard that we could represent

16 Note that under specific conditions conversions may be unrecorded—for example, if successive clicks within

one journey occur outside the observation period of 30 days, or if a user uses multiple devices. Thus, the occurrence of censored data may apply.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 86

by applying parametric models. This specific feature makes the Cox model a semiparametric

model and increases flexibility (Seetharaman and Chintagunta 2003).17 We further control for

parametric models including exponential, Weibull, and Gompertz distributions, as well as a

logit model, leading to comparable results. The hazard function of the Cox model for customer

i h6!t, X$ is

h6!t, X$ = h9!t$ × exp !1 β?X6?$,@

? �

( 6 )

with X = (X1, X2, …Xp) as predictor variables for customer i. We specify the predictor

variables for our base Model 1a for customer i as follows:

exp !1 β?X6?$@

? �= exp !β�Affiliate6 + β#Display6 + β3Other6 + β2Newsletter6

+ βMPriceComparison6 + βTReferrer6 + βVRetargeting6

+ βXPaidSearchGeneric6 + β\PaidSearchBrand6

+ β�9UnpaidSearchGeneric6 + β��UnpaidSearchBrand6

+ β�#Social6 + β�3TypeIn6 + β�2NaviSwitch6 + β�MInfoSwitch6

+ β�TTotalNoChannels6 + β�VTotalNoClicks6$.

( 7 )

In general, by disentangling our data set on the individual-user level, we follow current

industry practice and academic standards by using one of the most disaggregated units (Tellis

and Franses 2006). The time covariate T is aggregated over days, which is in line with previous

research (Lambrecht and Tucker 2013) and allows for including the chronological time effect

of the predictor covariates. Using model 1a as a basis, we introduce one additional covariate in

17 The semiparametric Cox model makes no assumption regarding the hazard over time; instead, the hazard is

derived as multiplicative replica of all subjects from the data, making it beneficial over parametric models if the shape of the hazard function is unknown.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 87

our Model 1b to investigate our concept of the micro-journey, leading to the following vector

of predictor covariates for customer i:

exp !1 β?X6?$@

? �= exp !β�MicroJourney6 + β#Affiliate6 + β3Display6 + β2Other6

+ βMNewsletter6 + βTPriceComparison6 + βVReferrer6

+ βXRetargeting6 + β\PaidSearchGeneric6

+ β�9PaidSearchBrand6 + β��UnpaidSearchGeneric6

+ β�#UnpaidSearchBrand6 + β�3Social6 + β�2TypeIn6

+ β�MNaviSwitch6 + β�TInfoSwitch6 + β�VTotalNoChannels6

+ β�XTotalNoClicks6$.

( 8 )

Models 1a and 1b are identical, with the exception of the additional covariate measuring

the effect of the micro-journey, making them nested models that allow for comparison tests.

We compare the goodness-of-fit measures of the two models—with and without micro-

journey—and further interpret the coefficient of the micro-journey covariate. The expression

β1 measures the effect of the micro-journey on the dependent covariate time to conversion T.

The supplementary predictor covariates represent a full model, and control for additional known

effects that may influence conversion likelihood: β2 to β14 measure the effect of the user´s

channel exposure,18 β15 and β16 include navigational and informational switches that the

customer conducts while browsing; and β17 and β18 control for the total number of channels as

well as clicks in the customer journey. The predictor terms β1 to β16 are defined as binary

covariates. In consequence, due to the aggregation of the dependent covariate T on a daily level,

multiple occurences within the same day may not be fully represented in our models. To control

18 As advertisers may apply different sets of online marketing channels, the channel exposure differs between data

sets.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 88

for the effect of multiple occurrences (for instance, occurrences of the micro-journey or channel

clicks within one day), we further verifed the robustness by computing our model using

continuous covariates that represent the exact number of micro-journeys, channel exposures,

and navigational and informational switches for each day. We provide this model specification

in the Appendix section. Furthermore, all predictor covariates are modeled time-independent,

because our tests suggest that the proportional hazard condition holds across all data sets and

covariates, and verifies the assumption that the covariates are multiplicatively related to the

hazards. In other words, the predictors’ occurrence (and non-occurrence) within customer

journeys may affect the course of the corresponding hazards, yet, keeps them proportional to

one another. One potential explanation of this phenomenon may be that the user, before entering

a micro-journey, already has a more definite idea about buying a product/service, and therefore

is more likely to enter a browsing state of high concentration focus (flow). Thus, the micro-

journey itself may express the user’s prior purchase intention, rather than only raising the

purchase likelihood during its occurrence. Nevertheless, we additionally controlled for

modeling of all time-varying covariates, which leads to analogous results, thereby verifying the

robustness of our model. We provide the specification of the time-dependent model in the

Appendix section.

3.4.3 The Micro-Journey Characteristics (Model 2)

As micro-journeys exhibit various characteristics, our detailed analysis explores the differences

between the various aspects of micro-journeys, and investigates their ability to predict

conversion events. We analogously apply a Cox proportional hazards model, but include a

further set of covariates to measure micro-journey characteristics (which we introduced in the

section “Characterizing the Micro-Journey”). These characteristics are: (1) the number of

micro-journey clicks, (2) the number of micro-journey channels, (3) the micro-journey duration,

(4) the relative position of the micro-journey, and (5) the browsing goal of the micro-journey.

Again, we control for parametric models and for robustness by including continuous covariates,

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 89

a logit model, as well as time-dependent covariates. We formulate two different models, which

allows us to first measure the effect of the micro-journey covariate and, second, to rule out

potential definititional biases related to the micro-journey characteristics. Therefore, Model 2b

includes the identical sample from Model 1, allowing us to further include and interpret the

micro-journey covariate. Additionally, and in opposition to Model 2b, we add Model 2a with a

sample that has journeys with exactly one micro-journey.19 On this basis, we can elude the

interdependecies and potential definitional biases of micro-journey characteristics that can

occur when one particular user journey comprises more than one micro-journey on the same

calendar day. If one user has multiple micro-journeys on the same day, the covariates measuring

the micro-journey characteristics may not be identified exhaustively, as they are aggregated

over days and, in consequence, one micro-journey needs to be selected over the other. While

we can completely rule out this selection bias for Model 2a, we handle it in Model 2b by

including the peak values of the micro-journey characteristics for users with more than one

micro-journey on the same day. For instance, if a user has two micro-journeys on the same

day—one journey with three online channels involved and a second journey with only one

channel—the covariate measuring the number of channels of the micro-journey will take the

larger value (here, three). In consquence of the possibility of imprecise definitions, we include

Model 2b primarily, in order to verify that the micro-journey covariate remains significant and

positive even within a model that includes additional predictors. With Model 2a, our objective

is to interpret the micro-journey characteristics, and to exclude the micro-journey covariate, as

it always takes the value one due to the sample selection. While the general formula of the

19 In a related approach, Lambrecht and Tucker (2013) narrowed the samples in order to detail the analyses, and

modify the dependent variable in Cox models to introduce time lags.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 90

hazard function h6!t, X$ remains unchanged, for Model 2a we adapt the vector of covariates

within the exponential component of the formula for customer i accordingly:20

exp !1 β?X6?$@

? �= exp!β�MicroJourneyClicks6 + β#MicroJourneyChannels6

+ β3MicroJourneyDuration6 + β2MicroJourneyPosition6

+ βMMicroJourneyNaviSwitch6 + βTMicroJourneyInfoSwitch6

+ βVAffiliate6 + βXDisplay6 + β\Other6 + β�9Newsletter6

+ β��PriceComparison6 + β�#Referrer6 + β�3Retargeting6

+ β�2PaidSearchGeneric6 + β�MPaidSearchBrand6

+ β�TUnpaidSearchGeneric6 + β�VUnpaidSearchBrand6

+ β�XSocial6 + β�\TypeIn6 + β#9NaviSwitch6 + β#�InfoSwitch6

+ β##TotalNoChannels6 + β#3TotalNoClicks6$.

( 9 )

As models 2a and 2b have an identical vector of covariates, except for the general micro-

journey covariate, we describe the predictor covariates of Model 2a. Due to different samples,

we intentionally refrain from interpreting the goodness-of-fit criteria of these models. The

coefficients β1 to β6 include the effects of micro-journey characteristics that include: the number

of clicks (β1), the number of different channels (β2), the duration measured in time (β3), the

relative position within the whole journey (β4), the occurrence of a navigational switch (β5), and

the respective informational switch within the micro-journey (β6). While β7 to β19 control for

channel effects, β20 and β21 control for navigational and informational switches throughout the

entire journey, and β22 and β23 control for the total number of different channels and clicks in a

manner similar to that of Model 1.

20 The vector of covariates for Model 2b is provided in the Appendix section.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 91

3.4.4 The Micro-Journey and Procrastination (Model 3)

Through our third model, we aim to better identify whether users show micro-journeys in their

browsing patterns when they finalize their purchase decisions, or whether they simply acquire

information first, which is indicated by a micro-journey, and then procrastinate in their purchase

decision (O’Donoghue and Rabin 1999). This differentiation between direct purchase and

procrastination is especially relevant for marketers who may observe micro-journeys while

tracking potential customers, in order to calibrate marketing measures toward individual users.

Figure 9

Direct versus Later Conversions after One Micro-Journey (in %)—Essay 2

In order to investigate which characteristics of micro-journeys indicate direct purchases

versus later purchases, we again limit our sample to journeys with exactly one micro-journey,

enabling a clear differentiation between direct and later purchases. If a user shows more than

one micro-journey for the browsing path and, for example, converts directly with the last micro-

journey, it is difficult to define the dependent covariate T, time to purchase, for the journey as

it would include micro-journeys that convert both directly and later. Based on the sample with

exactly one micro-journey, Figure 9 reveals that large shares of journeys convert directly after

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 92

the micro-journey as well as later—after single clicks—which indicates that users not yet

converted with a micro-journey are still relevant for marketing measures.

This analysis neither conceals the micro-journey characteristics that lead to the direct or

the later purchases, nor controls for other pre-known influences. Therefore, we turn to a

sequential Cox proportional hazards model. The application of a sequential Cox model is

inspired by the sequential logit model, which consists of separate logistic regression for each

transition step on the sub-sample in charge of making a binary decision (Buis 2011).

Figure 10

The Transition Tree of the Sequential Models (Model 3)—Essay 2

The setup of our third model resembles Model 2a: The sample and the vector of

covariates remain unchanged, and the predictor for the micro-journey itself is excluded because

it takes a value of one across all journeys. However, due to the sequential nature of the third

model, we replace the dependent covariate T, time to purchase, with T1, time to direct purchase

after the micro-journey (transition 1) and T2, time to later purchase after the micro-journey

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 93

(transition 2).21 In this way, these two transitions reflect the chronological nature of the purchase

decision process. First, the user directly purchases after the micro-journey or does not purchase

directly thereafter. Second, when the user does not purchase directly after the micro-journey,

he or she purchases later or does not purchase at all during the tracking period (Figure 10).

For robustness, we again control for parametric models, for continuous covariates, as

well as for models with time-dependent covariates. We also apply a sequential logit model that

follows the definition from prior literature (e.g., Buis 2011), and calculate a multinomial logit

model with an equivalently defined dependent variable: no purchase, direct purchase, and later

purchase. Again, all models show comparable results.

3.5 Estimation Results

3.5.1 Results of Model 1 – The Micro-Journey

Comparison of nested models. Table 15 details the results of the nested Model 1 for our primary

data set, including the predictor covariates as well as the goodness-of-fit results, reporting log-

likelihood, the Akaike Information Criterion (AIC, Akaike, 1974), the Bayesian Information

Criterion (BIC, Schwarz, 1978), and two R2 statistics. To compare the nested models 1a and 1b,

we include AIC and BIC, as they apply a penalty that adjusts for an increasing number of

estimated parameters, which discourages overfitting (Burnham and Anderson 2002). For hazard

models, no criterion to illustrate model fit is commonly agreed upon (Scherer, Wünderlich, and

von Wangenheim 2015). To simplify interpretation, we further add two measures that resemble

the explained variation from regression models, R2D (Royston and Sauerbrei 2004) and R2

PH

(Hosmer, Lemeshow, and May 2008; Royston 2006).22 For the logit model, we report

21 Direct conversions are defined as conversions that occur directly after a micro-journey. Later conversions are

defined as conversions that happen after a micro-journey, however, with at least one single click between the micro-journey and the conversion event.

22 The mathematical definition is illustrated in the Appendix section.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 94

McFadden’s (1974) R2, following Menard’s (2000) argument, that it satisfies most of

Kvålseth’s (1985) criteria for an appropriate R2.

To simplify the results presentation, we primarily report results of the Cox model.

Across all five criteria, Model 1b (which includes the micro-journey as an additional predictor)

shows a superior model fit over Model 1a, confirming that adding the micro-journey improves

model fit. Performing the likelihood-ratio test, the improvement in log-likelihood in Model 1b

over Model 1a is significant, χ2 (1) = 181.91, p < .0000. The model fit according to both AIC

and BIC is substantially lower for Model 1b compared to Model 1a, providing strong evidence

that Model 1b has an advantageous predictive performance over Model 1a (Kass and Raftery

1995; Wasserman 2000). The explained variation, R2D and R2

PH, both improve for Model 1b

over Model 1a. The results remain robust in an analogous logit model.

Table 15

Estimation Results: The Micro-Journey as Predictor (Model 1)—Essay 2

Data Set 1 (Fashion) Model 1a Model 1b

Variable B SE 95% CI Exp(B) B(Logit) B SE 95% CI Exp(B) B(Logit)

MicroJourney 0.464*** 0.03 [0.399, 0.529] 1.590 0.582***

Affiliate -0.807*** 0.05 [ -0.913, -0.701] 0.446 0.003 -0.975*** 0.06 [-1.084, -0.865] 0.377 -0.062

Display Newsletter 0.664*** 0.04 [ 0.588, 0.739] 1.943 0.396*** 0.468*** 0.04 [0.386, 0.505] 1.597 0.260***

PriceComparison 0.623*** 0.220 [ 0.192, 1.054] 1.865 0.964*** 0.376* 0.22 [-0.057, 0.809] 1.456 0.856*** Referrer -0.131* 0.07 [ -0.272, 0.009] 0.877 0.209*** -0.266*** 0.07 [-0.408, -0.123] 0.766 0.176***

Retargeting 0.369*** 0.070 [ 0.231, 0.507] 1.446 0.353*** 0.180** 0.07 [0.039, 0.321] 1.197 0.301***

PaidSearchGeneric 0.438*** 0.050 [ 0.340, 0.537] 1.550 0.459*** 0.261*** 0.05 [0.159, 0.363] 1.298 0.413***

PaidSearchBrand 1.211*** 0.03 [ 1.144, 1.278] 3.357 1.090*** 1.033*** 0.04 [0.961, 1.105] 2.809 1.016*** UnpaidSearchGeneric -0.286*** 0.060 [ -0.403, -0.169] 0.751 -0.338*** -0.531*** 0.06 [-0.654, -0.409] 0.588 -0.475***

UnpaidSearchBrand 0.759*** 0.03 [ 0.692, 0.827] 2.136 0.653*** 0.562*** 0.04 [0.488, 0.636] 1.754 0.542***

TypeIn 1.029*** 0.03 [ 0.966, 1.092] 2.798 0.406*** 0.910*** 0.03 [0.844, 0.976] 2.484 0.403***

NaviSwitch -0.470*** 0.03 [ -0.525, -0.415] 0.625 1.889*** -0.472*** 0.03 [-0.527, -0.417] 0.624 1.779*** InfoSwitch -0.654*** 0.05 [ -0.744, -0.563] 0.520 0.020 -0.647*** 0.05 [-0.737, -0.556] 0.524 -0.088*

TotalNoChannel 0.873*** 0.01 [ 0.858, 0.888] 2.394 0.875*** 0.01 [0.860, 0.891] 2.399 TotalNoClicks -0.004*** 0 [ -0.006, -0.002] 0.996 0.042*** -0.006*** 0 [-0.008, -0.004] 0.994 0.032***

TotalDuration 0.000*** 0.000***

Constant -6.838*** -6.719***

N 1,184,582 1,184,575 1,184,582 1,184,575

Observations 1,398,267 1,398,267 Time at risk 1,863,964 1,863,964

Log likelihood -103,305.6 -41,095.5 -103,214.6 -40,939.3

AIC 206,641.2 82,220.9 206,461.3 81,910.6

BIC 206,823.5 82,400.7 206,655.7 82,102.3

R2 (D) 0.622 0.628 R2 (PH) 0.779 0.782 R2 (McFadden) 0.297 0.299 Note: * p < .10, ** p < .05, *** p < .01; In the logit model, display predicts failure perfectly and 7 subjects are dropped. Display was removed; The results for Data Set 2 to Data Set 4 are reported in the Appendix.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 95

Relevance of the micro-journey. Referring again to the results of models 1a and 1b in Table 15,

we report statistical significance and 95% confidence intervals for the predictor coefficients.

As p-values that do not scale up well for extensive data sets lead to deflation, the information

transported via confidence intervals is considered more precise (Lin, Lucas Jr., and Shmueli

2013). The predictor of the micro-journey in Model 1b has a significant and strong positive

effect on purchase events (Cox Model 1b: MicroJourney b = 0.464, p < .01). Thus, adding

micro-journeys to a model not only improves model fit (illustrated in the section above), but

also indicates user journeys that are more likely to conclude in a purchase transaction. This

result also holds for the logit model (Logit Model 1a: MicroJourney b = 0.582, p < .01).

Looking at the control covariates, we find very similar effects for both Model 1a and

Model 1b, highlighting the stability of our results. While the total number of different channels

has a positive effect on purchase events (Cox Model 1a: TotalNoChannel b = 0.873, p < .01;

Cox Model 1b: TotalNoChannel b = 0.875, p < .01), the total number of journey clicks has a

significant, but negligible effect (Cox Model 1a: TotalNoClicks b = -0.004, p < .01; Cox Model

1b: TotalNoClicks b = -0.006, p < .01; Logit Model 1a: TotalNoClicks b = 0.000, p < .01; Logit

Model 1b: TotalNoClicks b = 0.000, p < .01). The former results, regarding total number of

different channels, endorse the results from previous research—showing that advertisement

exposure across various channels affects purchase events positively (Naik and Raman 2003;

Wiesel, Pauwels, and Arts 2011). The latter results are interesting and relevant to our research.

First, they contradict the results from preceding research on display marketing that the length

of the user journey, measured in clicks, does matter in predicting purchase events (Cho, Lee,

and Tharp 2001). Second, the results reveal that only the length of the user journey is an

insufficient predictor to forecast purchase events. Journeys with more clicks better fulfill the

preconditions of containing a micro-journey, especially one or multiple micro-journeys that

have a large number of clicks. However, the total number of journey clicks itself is not a

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 96

decisive predictor, so that the concentration of these clicks, measured with micro-journeys,

seems more suitable to distinguish converting journeys from other journeys.

Interestingly, navigational and informational switches show negative effects on the time

to purchase for the Cox model. Interpreting these coefficients that measure the browsing goal

of a user emphasizes that their estimates are of a relative nature, and are embedded in a full

model. In comparison with the other predictors in the model, they show a negative effect.

However, our data tracks only users who visit the advertiser’s website. A negative effect related

to navigational switches, for example, does not imply that a user with a navigational switch has

a lower purchase propensity compared to users with no advertisement exposure at all. Still,

navigational switches show a smaller negative effect on purchase events compared to

informational switches across both Cox models, with the switches being in comparable order

with previous research (Klapdor et al. 2015). Hereby, the results from the logit deviate from the

Cox model. While navigational switches have a strong positive effect on purchase events, the

results from the predictor of informational switches becomes insignificant for Model 1a, though

only less significant for Model 1b. The difference in the results may originate from the different

setup and level of detail in the model. While the Cox model can include a predictor multiple

times, as it is aggregated over days and thereby can reflect their chronological order, the logit

model simply includes whether a switch (navigational or informational) is observed at least

once per journey. Moreover, switches that occur on the same day as micro-journeys, and the

potential coincidence of the two, are disregarded in the logit model. Because of these model-

specific limitations, we have a stronger belief in the results derived from the Cox model.

3.5.2 Results of Model 2 – The Micro-Journey Characteristics

Whereas Model 1 shows that micro-journeys are a suitable predictor for converting journeys,

Model 2b includes the full sample identical to Model 1, the predictor of the micro-journey, and

additional micro-journey characteristics.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 97

To avoid any confusion, both models—Model 2a and 2b—have a dedicated purpose and

are not appropriate for comparison in the way that Model 1a and Model 1b are. For the micro-

journey characteristics, we focus our interpretation on Model 2a, and for the micro-journey

predictor we focus on Model 2b (Table 16). In the latter, the predictor of the micro-journey

itself remains positive and significant, which is in line with Model 1b, which disregards micro-

journey characteristics (Cox Model 2b: MicroJourney b = 1.146, p < .01). In Model 2a, all

predictors of the micro-journey characteristics are significant, except the predictor on the

number of micro-journey clicks. In the Cox model, the relative micro-journey position has the

strongest positive effect on purchases (Cox Model 2a: MicroJourneyPosition b = 0.918, p <

.01). Furthermore, we find significant positive effects with navigational and informational

switches (Cox Model 2a: MicroJourneyNaviSwitch b = 0.586, p < .01;

MicroJourneyInfoSwitch b = 0.707, p < .01). While the effect of the micro-journey duration (in

time) is significant and negligible (Cox Model 2a: MicroJourneyDuration = 0.001, p < 0.1), the

number of channels within the micro-journey has a significant negative effect (Cox Model 2a:

MicroJourneyChannels b = -0.236, p < .01). Looking at Model 2b, the Cox model shows mostly

comparable and significant results; however, the magnitude of the effects changes to some

degree. The number of micro-journey channels remains significant and negative (Cox Model

2b: MicroJourneyChannels b = -0.647, p < .01), and navigational and informational switches

remain positive (Cox Model 2b: MicroJourneyNaviSwitch b = 1.267, p < .01;

MicroJourneyInfoSwitch b = 1.145, p < .01). The number of micro-journey clicks becomes

significant, but has a relatively limited effect on the time to conversion (Cox Model 2b:

MicroJourneyClick b = 0.054, p < .01), likewise the micro-journey duration (Cox Model 2b:

MicroJourneyDuration b = 0.000, p < .01). One potential explanation of the higher magnitude

of the coefficients may lie in the relative nature of their estimates, which depend on the sample

selection. Whereas Model 2a analyzes journeys with exactly one micro-journey, Model 2b

includes the full sample (i.e., also journeys without micro-journeys). When analyzing the full

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 98

sample, the reference values of the estimates differ between the models, yet show identical

directions—apart from the micro-journey position—which becomes significant and negative

(Cox Model 2b: MicroJourneyPosition b = -0.578, p < .01). Differing from Model 2a, Model

2b includes journeys with multiple micro-journeys. The descriptive statistics in Table 13 show

that the conversion rate increases with the number of micro-journeys for Data Sets 1, 2, and 3.

If a journey includes multiple micro-journeys, the micro-journeys are located at different

relative positions within the overall journey, which may be the cause of this deviation from

Model 2a. We conclude that the micro-journey position seems to play a pivotal role, as do

navigational and informational switches. Prior literature leads us to expect a positive

relationship between navigational switches and purchase events, and a negative relationship

between informational switches and purchase events (Klapdor et al. 2015), as navigational

switches may be interpreted as progression and informational switches as regression in the

purchase decision process. Again, as in Model 1, our results are counterintuitive: Navigational

and informational switches as controls show negative effects on conversion events, while the

order of effects remains as expected from previous findings (Klapdor et al. 2015). Interestingly,

we find that both switches are associated with a strong positive effect if they occur within a

micro-journey. That is especially remarkable given that the number of channels that occur

within a micro-journey show a negative effect since switches require at least two distinct

channels—and the number of channels of the entire journey shows a positive effect. Obviously,

within the micro-journey, category switches show a positive sign, potentially partly layered by

the positive effects of the micro-journey itself. This overlaying effect seems to diminish if too

many channels (potentially from one channel group) become involved. Thus, within the short

time intervals zoned by the micro-journey, exposure toward multiple channels show negative

coefficients—yet, except for channel exposure in switches. This may imply rather idiosyncratic

channel preferences, at least, in short time intervals. Within the micro-journey, however, both

informational and navigational switches are positively associated with time to purchase events

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 99

at a comparable level. Thus, switches between these categories may be interpreted as intense

browsing patterns, either leading to progression in the deliberation process (navigational

switch) or, potentially, indicating further information acquisition (informational switch), which

may also express purchase propensities. The results of the logit model indicate that the

coefficients are more closely related to existing research, which may be due to the formerly

applied logit model (Klapdor et al. 2015). Additionally, we provide more generalized findings

across data sets in the Section 3.6.

Table 16

Estimation Results: The Micro-Journey Characteristics (Model 2)—Essay 2

Data Set 1 (Fashion) Model 2a Model 2b

Variable B SE 95% CI Exp(B) B(Logit) B SE 95% CI Exp(B) B(Logit)

MicroJourney 1.146*** 0.091 [0.968,1.323] 3.146 0.586***

MicroJourneyClicks 0.009 0.013 [-0.017,0.035] 1.009 0.025* 0.054*** 0.011 [0.034,0.075] 1.055 0.034***

MicroJourneyChannels -0.236 *** 0.067 [-0.367,-0.105] 0.790 0.101 -0.647*** 0.063 [-0.770,-0.525] 0.524 0.027

MicroJourneyDuration 0.001 *** 0.000 [0.001,0.001] 1.001 0.000*** 0.000*** 0.000 [0.000,0.000] 1.000 0.000*** MicroJourneyPosition 0.918 *** 0.113 [0.697,1.139] 2.504 0.327*** -0.578*** 0.069 [-0.713,-0.442] 0.561 -0.589***

MicroJourneyNaviSwitch 0.586 *** 0.107 [0.375,0.796] 1.797 -0.060 1.267*** 0.083 [1.104,1.430] 3.550 -0.230***

MicroJourneyInfoSwitch 0.707 *** 0.096 [0.518,0.896] 2.028 0.113 1.145*** 0.077 [0.994,1.295] 3.142 0.207***

Affiliate -0.418 *** 0.094 [-0.602,-0.235] 0.658 0.583*** -0.984*** 0.059 [-1.100,-0.867] 0.374 -0.074

Display Newsletter 0.174 ** 0.086 [0.006,0.342] 1.190 0.325*** 0.522*** 0.044 [0.436,0.609] 1.685 0.314***

PriceComparison -0.130 0.360 [-0.836,0.577] 0.878 0.526* 0.416* 0.222 [-0.019,0.851] 1.516 0.816***

Referrer -0.648 *** 0.128 [-0.898,-0.397] 0.523 0.001 -0.308*** 0.077 [-0.459,-0.157] 0.735 0.182***

Retargeting -0.268 ** 0.122 [-0.508,-0.029] 0.765 0.419*** 0.232*** 0.073 [0.090,0.374] 1.261 0.318*** PaidSearchGeneric -0.373 *** 0.093 [-0.555,-0.191] 0.689 0.336*** 0.210*** 0.057 [0.098,0.322] 1.234 0.417***

PaidSearchBrand 0.354 *** 0.076 [0.204,0.503] 1.425 0.611*** 1.107*** 0.041 [1.027,1.187] 3.025 1.042***

UnpaidSearchGeneric -1.209 *** 0.108 [-1.421,-0.997] 0.298 -0.401*** -0.829*** 0.073 [-0.973,-0.685] 0.436 -0.522***

UnpaidSearchBrand 0.276 *** 0.081 [0.118,0.435] 1.318 0.564*** 0.583*** 0.043 [0.499,0.667] 1.791 0.568*** TypeIn 0.439 *** 0.072 [0.298,0.581] 1.551 0.475*** 0.944*** 0.039 [0.869,1.020] 2.570 0.406***

NaviSwitch -0.219 *** 0.061 [-0.338,-0.1] 0.803 1.440*** -0.628*** 0.030 [-0.686,-0.570] 0.534 1.758***

InfoSwitch -0.385 *** 0.084 [-0.549,-0.221] 0.680 -0.434*** -0.804*** 0.047 [-0.897,-0.711] 0.448 -0.135**

TotalNoChannel 0.956 *** 0.018 [0.921,0.991] 2.601 0.887*** 0.008 [0.871,0.902] 2.428

TotalNoClicks 0.036 *** 0.004 [0.028,0.044] 1.037 0.040*** -0.005*** 0.001 [-0.007,-0.003] 0.995 0.031*** TotalDuration 0.000*** 0.000***

Constant -5.833*** -6.711***

N 81,087 81,085 1,184,582 1,184,575

Observations 123,230 1,398,267

Time at risk 208,352 1,863,964

Log likelihood -16,851.2 -7,434.3 -102,907.7 -40,870.9AIC 33,742.4 14,910.5 205,859.5 81,785.7

BIC 33,936.9 15,105.9 206,126.8 82,049.4

R2 (D) 0.637 0.630

R2 (PH) 0.859 0.794 R2 (McFadden) 0.265 0.301 Note: * p < .10, ** p < .05, *** p < .01; For Model 2b, the micro-journey characteristics are modeled as interaction effects as specified in the Appendix section. In the logit model, display predicts failure perfectly and 2 resp. 7 subjects are dropped. Display was removed; The results for Data Set 2 to Data Set 4 are reported in the Appendix.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 100

3.5.3 Results of Model 3 – The Micro-Journey and Purchase Timing

In Model 2, we analyzed micro-journey characteristics and found that the relative position of

the micro-journey, as well as the navigational and informational switches within the micro-

journey, had a strong effect on purchases. However, these results differed between models.

Therefore, we further detail our analysis in Model 3. We report the results of the sequential Cox

and the sequential logit models in Table 17. First, we look at the relative position of the micro-

journey, the strongest positive predictor in the first transition of Model 3. Regarding direct

purchases, the micro-journey position takes a leading role (Cox Model 3a:

MicroJourneyPosition b = 3.468, p < .01; Logit Model 3a: MicroJourneyPosition b = 3.758, p

< .01). In other words, a user who starts with single clicks, yet continues browsing with a micro-

journey at a later stage, is likely to conduct a direct purchase within the micro-journey. With

the Cox model, we may not show an effect for later purchases, as the predictor becomes

insignificant using Data Set 1 (Cox Model 3b: MicroJourneyPosition b = -0.355, p = .13; Logit

Model 3b: MicroJourneyPosition b = -1.048, p < .01). According to the logit model, a user who

starts browsing with a micro-journey but does not convert directly thereafter is likely to conduct

a purchase at a later stage—however, with a single click. Pre-empting the Cox models’ results

from the additional data sets (see the Appendix), we find indication that some users start with

a micro-journey, yet cease browsing, and then return later to conduct a purchase with single

clicks.

Addressing switches in our model reveals that their effects remain positive and

significant within micro-journeys, yet are negative or insignificant with regard to the overall

journey. For direct purchases, the results from navigational and informational switches are at a

comparable level to each other (Cox Model 3a: MicroJourneyNaviSwitch b = 0.441, p < .01;

MicroJourneyInfoSwitch b = 0.437, p < .01). Analyzing later purchases indicates that

informational switches show a stronger positive effect than do navigational switches (Cox

Model 3b: MicroJourneyNaviSwitch b = 0.515, p < .05; MicroJourneyInfoSwitch b = 0.890, p

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 101

< .01). Informational switches may be interpreted as regression in the purchase process—or as

additional information acquisition processes. Thus, an informational switch within a micro-

journey may indicate that a user takes a step back during a more concentrated browsing session,

potentially to acquire further information via informational channels. In consequence, the user

does not necessarily refrain from buying as such, just from buying directly, yet is likely to

purchase at a later stage through single clicks. As this effect may not hold for informational

switches during the whole journey, the degree of concentration focus, isolated by the micro-

journey, seems to play a pivotal role in combination with switches.

While the number of clicks and the number of channels have a positive effect on direct

purchases (Cox Model 3a: MicroJourneyClicks b = 0.098, p < .01; MicroJourneyChannels b =

1.224, p < .01), the results are contrary to later purchases (Cox Model 3b: MicroJourneyClicks

p = -0.267, b < .01; MicroJourneyChannels b = -0.706, p < .01). Particularly, the number of

channels within the micro-journey show strong bidirectional effects on purchase events, adding

details to our results on user channel preferences. In a micro-journey, a user, who utilizes a

larger number of channels while browsing, is more likely to conduct a direct purchase. In

contrast, users who tend to purchase later show more idiosyncratic, short-term channel

preferences. In summary, a micro-journey with a higher relative position within the journey,

and with a greater number of different channels, indicates users who are prone to convert

directly. In comparison, a micro-journey with a smaller number of clicks and fewer channels

indicates a later purchase anteceded by one or several single clicks. So far, we cannot make a

conclusion about the relative position of the micro-journey for later converters, as the predictor

becomes insignificant. Again, the effect of the journey duration in time is significant but

negligible across all models.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 102

Table 17

Estimation Results: The Effect of the Micro-Journey Characteristics on Direct and Later

Conversions (Model 3)—Essay 2

Data Set 1 (Fashion) Model 3a – Direct Conversion Model 3b – Later Conversion

Variable B SE 95% CI Exp(B) B(Logit) B SE 95% CI Exp(B) B(Logit)

MicroJourneyClicks 0.098*** 0.018 [0.063,0.133] 1.103 0.309*** -0.267 *** 0.071 [-0.405,-0.128] 0.766 -0.081 **

MicroJourneyChannels 1.224*** 0.116 [0.996,1.453] 3.401 0.416*** -0.706 *** 0.143 [-0.986,-0.425] 0.494 -0.115

MicroJourneyDuration 0.000*** 0.000 [0.000,0.001] 1.000 0.000*** 0.001 *** 0.000 [0.001,0.001] 1.001 0.000 ***

MicroJourneyPosition 3.468*** 0.150 [3.174,3.763] 32.073 3.758*** -0.355 0.235 [-0.816,0.105] 0.701 -1.048 ***

MicroJourneyNaviSwitch 0.441*** 0.134 [0.178,0.704] 1.554 0.817*** 0.515 ** 0.223 [0.078,0.951] 1.674 -0.373 ***

MicroJourneyInfoSwitch 0.437*** 0.122 [0.199,0.676] 1.548 0.847*** 0.890 *** 0.186 [0.525,1.255] 2.435 0.000

Affiliate -1.525*** 0.156 [-1.829,-1.220] 0.218 0.422*** 0.161 0.134 [-0.101,0.424] 1.175 0.713 ***

Display

Newsletter -0.873*** 0.155 [-1.176,-0.570] 0.418 0.143 0.589 *** 0.108 [0.378,0.800] 1.802 0.441 ***

PriceComparison -0.855** 0.397 [-1.634,-0.076] 0.425 0.779* -0.976 1.006 [-2.948,0.996] 0.377 0.250

Referrer -1.801*** 0.195 [-2.183,-1.419] 0.165 0.083 -0.045 0.185 [-0.408,0.318] 0.956 -0.153

Retargeting -1.572*** 0.217 [-1.998,-1.146] 0.208 0.166 0.182 0.156 [-0.123,0.487] 1.200 0.501 ***

PaidSearchGeneric -1.490*** 0.151 [-1.786,-1.194] 0.225 0.319*** 0.073 0.139 [-0.200,0.346] 1.076 0.294 ***

PaidSearchBrand -0.917*** 0.144 [-1.199,-0.636] 0.400 0.342*** 0.904 *** 0.098 [0.711,1.097] 2.469 0.748 ***

UnpaidSearchGeneric -2.363*** 0.158 [-2.673,-2.052] 0.094 -0.451*** -0.704 *** 0.179 [-1.055,-0.353] 0.495 -0.378 ***

UnpaidSearchBrand -0.984*** 0.146 [-1.270,-0.699] 0.374 0.271** 0.826 *** 0.100 [0.629,1.023] 2.284 0.756 ***

TypeIn -1.041*** 0.140 [-1.314,-0.767] 0.353 0.021 1.002 *** 0.091 [0.823,1.181] 2.724 0.796 ***

NaviSwitch 0.040 0.114 [-0.184,0.264] 1.041 0.304* -0.267 *** 0.073 [-0.410,-0.123] 0.766 2.143 ***

InfoSwitch -0.419*** 0.121 [-0.656,-0.182] 0.658 -1.490*** -0.189 0.118 [-0.419,0.042] 0.828 -0.148

TotalNoChannel 1.002*** 0.049 [0.905,1.099] 2.724 0.907 *** 0.020 [0.868,0.947] 2.477

TotalNoClicks -0.015 0.015 [-0.045,0.015] 0.985 -0.206*** 0.047 *** 0.004 [0.038,0.056] 1.048 0.054 ***

TotalDuration 0.000** 0.000 ***

Constant -6.166*** -6.981 ***

N 81,087 81,087 80,199 80,199

Observations 123,230 121,474

Time at risk 208,352 204,261

Log likelihood -16,170.2 -8,241.7

Note: * p < .10, ** p < .05, *** p < .01; Display was removed from the model; The results for Data Set 2 to Data Set 4 are reported in the Appendix.

3.6 Robustness of the Estimation Results

In order to determine the robustness of our research and to generalize our implications, we

perform similar analyses for three additional data sets that include a fashion retailer, a luggage

retailer, and a travel company—all pure online players. We include a second fashion retailer to

ensure within-industry robustness, as well as companies from other industries to check for

robustness across industries. The data sets are equivalent in that they include a full set of

customer journeys, with converting and non-converting journeys. Advertiser-specific

differences may arise from their online channel strategies. In addition, we cannot ensure that

marketing campaigns within channels are identical—for example, affiliates between the data

sets may not include the same affiliate networks, or a paid search campaign may differ in its

marketing messages and keywords. As for the first data set, we track each website visit,

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 103

including the channel information and an exact timestamp, and we identify whether a visit leads

to a conversion event (or not). We report the descriptive statistics in Table 12. Except for the

different channel usage, the analyses and model specifications are similar to the sections in the

manuscript. Yet, we report only the results from our main model, the Cox model.

3.6.1 Results of Model 1

In Table 18, we illustrate the results from Model 1. For comparison, the first column repeats

the results from our main data set, whereas the other columns represent the results from the

three additional data sets, including fashion retail (Data Set 2), luggage retail (Data Set 3) and

travel (Data Set 4).

Across all data sets, Model 1b, including the micro-journey predictor, shows a superior

model fit over Model 1a for all four goodness-of-fit criteria. The differences in BIC between

the models 1a and 1b are well above the threshold for strong evidence to define model

superiority (Kass and Raftery 1995; Wasserman 2000). Conducting the likelihood-ratio test, the

improvement in log-likelihood for Model 1b over Model 1a is significant for all data sets (Data

Set 2: χ2 (1) = 675.59, p < .0000; Data Set 3: χ2 (1) = 272.93, p < .0000; Data Set 4: χ2 (1) =

191.69, p < .0000). Furthermore, the micro-journey predictor has a positive and highly

significant effect across all data sets (Data Set 2: MicroJourney b = 0.645, p < .01; Data Set 3:

MicroJourney b = 0.531, p < .01; Data Set 4: MicroJourney b = 0.413, p < .01). The magnitude

of the effect is at a relatively comparable level across data sets, such that we generalize that the

micro-journey is a well-suited and stable predictor for converting users within and even across

industries.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 104

Table 18

Robustness of the Results: The Micro-Journey as Predictor across Data Sets (Model 1)—

Essay 2

Model 1a Model 1b

DS 1 DS 2 DS 3 DS 4 DS 1 DS 2 DS 3 DS 4

Fashion Fashion Luggage Travel Fashion Fashion Luggage Travel

Variable B B B B B B B B

MicroJourney 0.464*** 0.645 *** 0.531*** 0.413***

Affiliate -0.807*** 1.541*** 0.743 *** -0.975*** 1.255*** 0.547***

Display -6.045 *** -6.381***

Newsletter 0.664*** -0.221 *** 0.468*** -0.442***

PriceComparison 0.623*** -0.404*** -0.896 *** 0.376* -0.679*** -1.115***

Referrer -0.131* -0.272*** -0.752*** -0.266*** -0.543 *** -0.923***

Retargeting 0.369*** 0.999 -0.758*** -0.584 *** 0.180** 0.531 -1.022*** -0.798***

PaidSearchGeneric 0.438*** 0.173*** -0.335*** -0.790 *** 0.261*** -0.181 *** -0.600*** -0.998***

PaidSearchBrand 1.211*** 0.685*** 0.689*** 0.777 *** 1.033*** 0.367 *** 0.433*** 0.554***

UnpaidSearchGeneric -0.286*** -0.324*** -0.990*** -0.749 *** -0.531*** -0.725 *** -1.307*** -1.017***

UnpaidSearchBrand 0.759*** 0.128*** -0.313*** 0.229 *** 0.562*** -0.228 *** -0.561*** -0.029

TypeIn 1.029*** 0.260*** 0.144*** 0.910*** 0.041 -0.018

Social 0.197*** -0.192 ***

Other 0.298*** -0.053

NaviSwitch -0.470*** -0.590*** -0.202*** -0.471 *** -0.472*** -0.612 *** -0.220*** -0.447***

InfoSwitch -0.654*** -0.712*** -0.746*** -0.741 *** -0.647*** -0.694 *** -0.761*** -0.725***

TotalNoChannel 0.873*** 1.059*** 1.104*** 1.047 *** 0.875*** 1.068 *** 1.134*** 1.056***

TotalNoClicks -0.004*** 0.016*** 0.050*** 0.013 *** -0.006*** 0.012 *** 0.037*** 0.009***

N 1,184,582 862,114 405,343 600,873 1,184,582 862,114 405,343 600,873

Observations 1,398,267 964,836 461,108 792,345 1,398,267 964,836 461,108 792,345

Time at risk 1,863,964 1,306,432 655,963 1,171,897 1,863,964 1,306,432 655,963 1,171,897

Log likelihood -103,305.6 -171,681.7 -88,058.7 -104,859.8 -103,214.6 -171,343.9 -87,922.3 -104,764.0

AIC 206,641.2 343,393.4 176,145.5 209,745.7 206,461.3 342,719.8 175,874.6 209,556.0

BIC 206,823.5 343,570.1 176,300.1 209,896.3 206,655.7 342,908.3 176,040.2 209,718.1

R2 (D) 0.622 0.436 0.487 0.650 0.628 0.446 0.492 0.653

R2 (PH) 0.779 0.614 0.703 0.902 0.782 0.628 0.712 0.904

Note: * p < .10, ** p < .05, *** p < .01

3.6.2 Results of Model 2

We formulate Model 2 in order to more thoroughly investigate the characteristics of a

micro-journey that affect conversion events. In Table 19,Table 19 we report the detailed results

for models 2a and 2b, and can mostly confirm the results derived from the model with the first

data set. The sample of Model 2a is reduced to journeys with exactly one micro-journey. Thus,

we focus our interpretation regarding the micro-journey characteristics on Model 2a and verify

the robustness of the micro-journey covariate, as such, with Model 2b. Furthermore, Model 2b

may help to confirm the results of Model 2a.

First, we observe Model 2b to find that the predictor of the micro-journey shows

significant and positive results across all data sets. This result confirms the existence of the

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 105

micro-journey even in an extended model, with additional predictors in the form of micro-

journey characteristics and across data sets or advertisers (Data Set 2: MicroJourney b = 1.850,

p < .01; Data Set 3: MicroJourney b = 0.952, p < .01; Data Set 4: MicroJourney b = 1.434, p <

.01). Turning to Model 2a, the relative position of the micro-journey remains the strongest

predictor among the micro-journey predictors across all data sets (Data Set 2:

MicroJourneyPosition b = 1.450, p < .01; Data Set 3: MicroJourneyPosition b = 0.834, p < .01;

Data Set 4: MicroJourneyPosition b = 1.713, p < .01). The predictors of navigational and

informational switches within micro-journeys are associated with positive effects on the time

until conversions, if they become significant (Data Set 2: MicroJourneyNaviSwitch b = 0.336,

p < .05; Data Set 3: MicroJourneyNaviSwitch b = 0.230, p < .10; Data Set 4:

MicroJourneyNaviSwitch b = 0.336, p < .01; MicroJourneyInfoSwitch b = 0.179, p < .05). As

informational switches within micro-journeys are insignificant for Data Set 2 and Data Set 3,

we can only discuss the order of the two effects with respect to Data Set 1 and Data Set 4. While

for Data Set 1 the informational switches show a stronger effect toward conversion events, as

compared to navigational switches, the results are opposite for Data Set 4. Consequently, the

four data sets confirm that category switches within micro-journeys are of high relevance in

predicting conversion events. However, we may not reach conclusions about the order of the

two effects across data sets, as these effects seem to be, at least to some degree, advertiser-

specific. Interestingly, the effects of navigational and informational switches for the full journey

(also outside the micro-journeys) are significant and negative across all data sets, with the

exception of navigational switches in Data Set 2. Apart from this exception, the positive effect

of category switches within micro-journeys thus seems to be associated with a highly intense

user browsing phase—and a limited predictor (negative effect), considering the less intense

browsing states throughout the full journey. The duration of the micro-journey, again, becomes

significant yet remains negligible (Data Set 2: MicroJourneyDuration b = 0.001, p < .01; Data

Set 3: MicroJourneyDuration b = 0.001, p < .01; Data Set 4: MicroJourneyDuration b = 0.001,

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 106

p < .01). While the number of micro-journey clicks is insignificant for Data Set 1, it becomes

significant for all additional data sets; however, the results are mixed (Data Set 2:

MicroJourneyClicks b = 0.043, p < .01; Data Set 3: MicroJourneyClicks b = -0.131, p < .01;

Data Set 4: MicroJourneyClicks b = -0.142, p < .01). While the predictor of the number of

micro-journey clicks has a slight but positive effect on conversions for the fashion retailer, it

becomes negative for the luggage retailer and the travel company. These variations may indicate

industry-specific effects and may also be moderated by the user’s price sensitivity on the

respective product category (Mehta, Rajiv, and Srinivasan 2003). Looking at the descriptive

statistics, data set 4 features some distinctive characteristics: It is the only data set with shorter

micro-journeys in converting click chains compared to non-converting click chains (Table 14),

while the average journey length in clicks is the longest across data sets (Table 12). The

resulting short and intense browsing sessions may work in the favor of users on a browsing path

to buy travel or consumer durables such as luggage. Longer intense browsing sessions may

express indecisiveness. As we can only track users’ web presence for one travel company and

one luggage retailer, visitors leaving the page might purchase at a different website and so are

a loss for our advertiser. The number of different channels within the micro-journey remains

negative across all data sets indicating generalizability (Data Set 2: MicroJourneyChannels b =

-0.566, p < .01; Data Set 3: MicroJourneyChannels b = -0.116, p < .10; Data Set 4:

MicroJourneyChannels b = -0.315, p < .01). In contrast and considering the full journey, the

total number of different channels has a significant and positive effect across all data sets, which

confirms prior research (Klapdor et al. 2015). These results may indicate users’ channel

preferences moderated by the time frame. With the exception of category switches, which, by

definition, occur between two different channels, users show idiosyncratic channel preferences

in very short time intervals (micro-journey), yet, may prefer exposure toward various channels

in longer time intervals (full journey).

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 107

Table 19

Robustness of the Results: The Micro-Journey Characteristics across Data Sets (Model 2)—

Essay 2

Model 2a Model 2b

DS 1 DS 2 DS 3 DS 4 DS 1 DS 2 DS 3 DS 4

Fashion Fashion Luggage Travel Fashion Fashion Luggage Travel

Variable B B B B B B B B

MicroJourney 1.146*** 1.850*** 0.952*** 1.434*** MicroJourneyClicks 0.009 0.043*** -0.131*** -0.142*** 0.054*** 0.082*** -0.055** -0.037 MicroJourneyChannels -0.236*** -0.566*** -0.116* -0.315*** -0.647*** -1.289*** -0.422*** -1.177*** MicroJourneyDuration 0.001*** 0.001*** 0.001*** 0.001*** 0.000*** 0.000*** 0.001*** 0.001*** MicroJourneyPosition 0.918*** 1.450*** 0.834*** 1.713*** -0.578*** -0.440*** -0.477*** 0.259*** MicroJourneyNaviSwitch 0.586*** 0.336** 0.230* 0.336*** 1.267*** 0.714*** 0.162 0.567*** MicroJourneyInfoSwitch 0.707*** 0.190 -0.004 0.179** 1.145*** 0.674*** -0.029 0.663***

Affiliate -0.418*** 1.408*** 1.085*** -0.984*** 1.458*** 0.809***

Display -6.651*** -6.047***

Newsletter 0.174** -0.160 0.522*** -0.086

PriceComparison -0.130 -0.642*** -1.231*** 0.416* -0.539*** -0.828***

Referrer -0.648*** -0.507*** -0.959*** -0.308*** -0.023 -0.758***

Retargeting -0.268** -41.497 -1.158*** -1.327*** 0.232*** 1.111 -0.857*** -0.496***

PaidSearchGeneric -0.373*** -0.518*** -0.621*** -1.357*** 0.210*** 0.337*** -0.430*** -0.660***

PaidSearchBrand 0.354*** -0.206** -0.120 0.316*** 1.107*** 0.906*** 0.639*** 0.930***

UnpaidSearchGeneric -1.209*** -0.573*** -1.255*** -1.501*** -0.829*** -0.226*** -1.059*** -0.582***

UnpaidSearchBrand 0.276*** -0.352*** -0.764*** -0.136* 0.583*** 0.361*** -0.291*** 0.460***

TypeIn 0.439*** -0.359*** -0.593*** 0.944*** 0.541*** 0.162**

Social -0.017 0.273***

Other 0.479*** 0.405***

NaviSwitch -0.219*** -0.013 0.347*** -0.612*** -0.628*** -0.666*** -0.236*** -0.556***

InfoSwitch -0.385*** -0.615*** -0.735*** -0.440*** -0.804*** -0.760*** -0.786*** -0.778***

TotalNoChannel 0.956*** 1.056*** 1.097*** 0.870*** 0.887*** 1.080*** 1.150*** 1.077***

TotalNoClicks 0.036*** 0.056*** 0.125*** 0.143*** -0.005*** 0.010*** 0.036*** 0.000

N 81,087 82,083 67,780 118,441 1,184,582 862,114 405,343 600,873

Observations 123,230 102,856 84,732 161,318 1,398,267 964,836 461,108 792,345

Time at risk 208,352 174,120 144,198 271,318 1,863,964 1,306,432 655,963 1,171,897

Log likelihood -16,851.2 -30,945.0 -22,198.9 -21,335.4 -102,907.7 -171,046.4 -87,800.0 -104,382.4

AIC 33,742.4 61,925.9 44,435.8 42,708.7 205,859.5 342,136.7 175,642.1 208,804.8

BIC 33,936.9 62,097.6 44,613.4 42,898.5 206,126.8 342,395.9 175,873.9 209,036.5

R2 (D) 0.637 0.538 0.591 0.774 0.630 0.451 0.489 0.662

R2 (PH) 0.859 0.750 0.820 0.977 0.794 0.639 0.719 0.911

Note: * p < .10, ** p < .05, *** p < .01; Covariates without result due to scarce observations were removed from results. For Model 2b, the micro-journey characteristics are modeled as interaction effects as specified in the Appendix section.

Model 2b confirms a vast majority of the results for micro-journey characteristics. Both

navigational and informational switches within the micro-journey become positive, and they

become negative and significant with regard to the full journey. Furthermore, the micro-journey

duration measured in time shows a significant but negligible effect. The micro-journey length

measured in clicks has a mixed, though very limited, effect, and the number of different

channels per micro-journey becomes negative—as shown in Model 2a. Except for the relative

position of the micro-journey, the results are relatively stable across data sets, and for models

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 108

2a and 2b. For model 2b, the relative position of the micro-journey becomes negative across all

data sets, which may be explained by the definitional issues, if multiple micro-journeys occur

on one day. Ruling out these definitional issues, Model 2a includes exactly one micro-journey.

We conclude that both the micro-journey position as well as category switches during the

micro-journey are good predictors to better forecast converting users. We point to Model 3 for

further detailing these effects.

3.6.3 Results of Model 3

In Model 3, we utilize the reduced sample as in Model 2a, but distinguish two transitions in a

sequential model that affect the dependent covariate, the time to purchase. While we analyze

the predictors’ effects on direct purchases in the first transition, we focus on later purchases in

the second transition (see Figure 10). In Table 20 we report the estimation results of the Cox

model and refer to Model 3a for direct purchases, and to Model 3b for later purchases. We

further report the results of a sequential logit model in the Appendix section. Considering that

we have four data sets from three different industries, the results show remarkable similarities

increasing robustness and allowing for generalizations.

Looking at the relative position of the micro-journey, we find significant and strong

positive effects on direct conversions in Model 3a for all data sets (Model 3a: Data Set 2:

MicroJourneyPosition b = 3.126, p < .01; Data Set 3: MicroJourneyPosition b = 2.680, p < .01,

Data Set 4: MicroJourneyPosition b = 3.239, p < .01). The results of the effect on later purchases

in Model 3b are mixed. While for the luggage retailer the effect becomes significant and

negative, it becomes significant and positive for the travel company (Model 3b: Data Set 3:

MicroJourneyPosition b = -0.647, p < .01, Data Set 4: MicroJourneyPosition b = 0.634, p <

.01). Overall, we generalize that for direct purchases, the effect of the micro-journey position

is substantially stronger. In other words, a user who begins with a few single clicks and

continues the browsing path later with a micro-journey is more likely to convert directly.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 109

Regarding later purchases, the effect of the micro-journey position seems to be advertiser-

specific and may not be generalized across data sets. Interestingly, for travel, the effect of the

micro-journey position remains significant and positive across Model 2 and Model 3. Thus, a

user who enters a micro-journey in his or her browsing path is highly relevant for direct

purchases, but also is likely to convert later, even if he or she did not yet convert directly.

Potentially, purchase behavior differs between durable goods and, in this case, travel

products—which are more costly and have a stronger social component, so that a travel product

is often searched, discussed, and selected by more than one person. With travel products, later-

situated micro-journeys are a positive indicator of both direct and later purchases. For consumer

durables (luggage), an early-situated micro-journey is a more suitable predictor of conversion

events. Consequently, a user who utilizes a micro-journey early in the search process and does

not purchase directly thereafter is likely to convert with single clicks at a later stage.

Regarding category switches within micro-journeys, some of the results become

insignificant, such that we cannot interpret results between direct and later purchases at a

detailed level. Nevertheless, most results for later purchases are significant and show positive

coefficients for both navigational and informational switches across data sets. The magnitude

of the effects for Data Set 2 and Data Set 4 are at a rather comparable level, contradicting the

results of Data Set 1, which shows a stronger effect for informational switches. These variations

may be driven by differing channel portfolios of the advertisers or may express industry-

specifics. Switches that occur throughout the full journey are mostly significant and are

negatively associated with purchase events. Only three predictors of navigational switches are

significant and positive (Model 3a: Data Set 3: NaviSwitch b = 0.506, p < .01; Model 3b: Data

Set 2: NaviSwitch b = 0.171, p < .05; Data Set 3: NaviSwitch 0.280 b = 0.280, p < .01).

Interestingly, the effects for Model 3b are smaller compared to the positive effects of their

equivalent within the micro-journey (Model 3b: Data Set 2: MicroJourneyNaviSwitch b =

0.517, p < .10; Data Set 3: MicroJourneyNaviSwitch 0.480 b = 0.280, p < .05). Thus, we derive

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 110

two effects. First, informational and navigational switches within micro-journeys are an

appropriate predictor for converting users. However, the order of effects is advertiser-specific

and may base on the channel strategy. Second, both types of switches are better suited to predict

conversion events when they happen within micro-journeys.

Recalling the results from Model 2, the number of micro-journey clicks was negatively

associated with conversions, and the effect of micro-journey channels was limited and differed

across data sets. Utilizing the sequential model, the results are more sorted. With regard to direct

purchases, both predictors become significant and positive across all data sets. The number of

different micro-journey channels, particularly, seems well suited to being a predictor.

Remarkably, these effects show opposite results for later purchases: The effect of the micro-

journey clicks is strongly negative, and the effect of the micro-journey channels is slightly

negative, when significant. Obviously, users who intend to buy directly tend to use various

channels and more clicks within a concentrated browsing session—the micro-journey—

whereas shorter micro-journeys with less channel variety indicate users who are (initially)

gathering information, and who conduct their purchases later. Again, the duration measured in

time has a significant but negligible effect with regard to purchase decisions across data sets.

Journey length happens in clicks, and not necessarily in time units.

Whereas in Model 3a the relative position has by far the strongest effect on (direct)

purchases (Model 3a: Data Set 2: MicroJourneyPosition b = 4.272, p < .01; Data Set 3:

MicroJourneyPosition b = 4.883, p < .01, Data Set 4: MicroJourneyPosition b = 4.631, p < .01),

for Data Set 1 the relative position has either a negative effect on later purchases (Model 3b),

or becomes insignificant for the remaining data sets. With regard to direct purchases (Model

3a), the number of clicks as well as the number of channels have significant and positive effects.

However, the effects become negative and significant for later purchases (Model 3b), with the

exception of one predictor in Data Set 2, which becomes insignificant. Furthermore, the

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 111

magnitude of the effect for the number of micro-journey channels is higher (positive for direct

purchases, negative for later purchases) compared to the number of micro-journey clicks across

all data sets. We generalize that the number of different channels within a micro-journey is the

more relevant predictor that should be considered. Again, the micro-journey duration measured

in time is significant, but negligible regarding its effect on both direct and later purchases, across

all data sets. Our reasoning is strengthened from models 3a and 3b, which include four data sets

in total—and for which different user groups may exist. The first group seems to begin their

browsing course with single clicks and then a pause before finalizing their purchase within a

micro-journey. The second group starts their journey with a more intense phase (micro-

journey), probably to retrieve information, and then returns to finalize their purchase with one

or more single clicks. A micro-journey at the beginning of a journey may indicate

procrastination in the purchase process, and is still relevant as a target for further advertisement

exposure (O’Donoghue and Rabin 1999). Potentially, a segment of these users may make their

purchase decision after the first micro-journey, and then return to the advertiser’s website

deliberately. However, another segment may be reminded to make their purchase decisions

(e.g., by firm-initiated advertising exposures), and thus may be prompted by the advertisers’

marketing measures to finalize their purchase.

When we focus on results related to switches, we find that for switches within a micro-

journey, both navigational and informational switches show positive effects in most data sets

and models. In the event that both predictors (navigational and informational switches within

micro-journeys) become significant, the positive effect of navigational switches is stronger than

the positive effect of informational switches, which is in line with prior research indicating that

navigational switches may be a measure for progress in the purchase decision process.

Interestingly, navigational and informational switches between clicks within the full journey

show, for most estimation results, negative effects—always in Model 3a, and for Data Sets 1

and 4 in Model 3b. We infer that the positive effect of a micro-journey may be overlaying and

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 112

outperforming or may be moderating the effect that is solely related to switches. Former

research has shown that switches, especially navigational switches, have a positive effect on

purchase events (Klapdor et al. 2015); these studies, however, omit micro-journeys in their

investigations. Thus, considering our results, the positive effect of switches may be associated

with the effect of the micro-journey, as they often happen simultaneously—a micro-journey

including a switch.

Table 20

Robustness of the Results: The Effect of the Micro-Journey Characteristics on Direct and

Later Conversions across Data Sets (Model 3)—Essay 2

Model 3 Direct Model 3 Later

DS 1 DS 2 DS 3 DS 4 DS 1 DS 2 DS 3 DS 4

Fashion Fashion Luggage Travel Fashion Fashion Luggage Travel

Variable B B B B B B B B

MicroJourneyClicks 0.098 *** 0.064 *** 0.207 *** 0.101 *** -0.267 *** 0.004 -0.250 *** -0.237 ***

MicroJourneyChannels 1.224 *** 1.709 *** 0.818 *** 0.717 *** -0.706 *** -1.609 *** -0.952 *** -1.029 ***

MicroJourneyDuration 0.000 *** 0.000 *** 0.001 *** 0.000 *** 0.001 *** 0.001 *** 0.001 *** 0.001 ***

MicroJourneyPosition 3.468 *** 3.126 *** 2.680 *** 3.293 *** -0.355 -0.237 -0.647 *** 0.634 ***

MicroJourneyNaviSwitch 0.441 *** 0.220 0.027 0.157 0.515 ** 0.517 * 0.480 ** 0.825 ***

MicroJourneyInfoSwitch 0.437 *** -0.218 -0.494 *** 0.137 0.890 *** 0.588 ** 0.269 0.721 ***

Affiliate -1.525 *** 0.470 *** 0.160 0.161 2.195 *** 1.599 ***

Display -8.892 *** -5.491 ***

Newsletter -0.873 *** -43.679 -43.738 -1.395 *** 0.589 *** -38.787 -41.268 0.677 ***

PriceComparison -0.855 ** -1.702 *** -2.693 *** -0.976 0.210 -0.202

Referrer -1.801 *** -2.458 *** -2.201 *** -0.045 0.350 *** 0.137

Retargeting -1.572 *** -43.871 -2.270 *** -2.765 *** 0.182 -38.676 -0.265 -0.372

PaidSearchGeneric -1.490 *** -2.705 *** -1.582 *** -2.711 *** 0.073 0.515 *** 0.231 ** -0.436 ***

PaidSearchBrand -0.917 *** -2.167 *** -1.126 *** -1.005 *** 0.904 *** 0.610 *** 0.860 *** 1.274 ***

UnpaidSearchGeneric -2.363 *** -2.672 *** -2.361 *** -2.784 *** -0.704 *** 0.416 *** -0.228 * -0.629 ***

UnpaidSearchBrand -0.984 *** -2.306 *** -2.031 *** -1.323 *** 0.826 *** 0.427 *** 0.386 ** 0.705 ***

TypeIn -1.041 *** -2.181 *** -1.452 *** 1.002 *** 0.393 *** 0.309 **

Social -2.028 *** 0.866 ***

Other -1.397 *** 0.744 **

NaviSwitch 0.040 -0.114 0.506 *** -0.613 *** -0.267 *** 0.171 ** 0.280 *** -0.735 ***

InfoSwitch -0.419 *** -0.352 *** -0.375 *** -0.400 *** -0.189 -0.915 *** -1.040 *** -0.464 ***

TotalNoChannel 1.002 *** 1.021 *** 1.318 *** 1.071 *** 0.907 *** 0.993 *** 1.004 *** 0.836 ***

TotalNoClicks -0.015 0.046 *** -0.155 *** -0.079 *** 0.047 *** 0.054 *** 0.144 *** 0.150 ***

N 81,087 82,083 67,780 118,441 80,199 80,125 66,428 116,894

Observations 123,230 102,856 84,732 161,318 121,474 99,848 83,052 159,117

Time at risk 208,352 174,120 144,198 271,318 204,261 164,867 141,296 266,644

Log likelihood -16,170.2 -29,808.7 -21,683.1 -20,584.8 Note: * p < .10, ** p < .05, *** p < .01; Covariates without result due to scarce observations were removed from results.

Regarding all models (models 1 to 3) and data sets, we find that the total number of

different channels within a full journey has a significant and strong positive effect, which is in

accord with prior research (Naik and Raman 2003; Wiesel, Pauwels, and Arts 2011). For

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 113

purchases, however, the effect of the number of total clicks within a full journey becomes

mostly significant, but relatively small. Consequently, controlling for the length of the journey

in clicks takes on a subordinate role in forecasting purchase events. The number of clicks,

solely, is shown not to be important, verifying our assumption that the time component of click

occurrences (i.e., user interaction) is a more relevant predictor, which we model with the micro-

journey.

3.7 General Discussion

Online clickstream data has boosted interest in analyzing online consumers’ path to purchase.

While existing research focuses on singular clicks and associates them with categories (e.g.,

informational and navigational), less is known about how (past) browsing behavior such as

users’ click patterns effect conversions or support in forecasting future conversion events.

Furthermore, although existing research acknowledges the influence of analyzing the

clickstream (Chatterjee, Hoffman, and Novak 2003; Hui, Fader, and Bradlow 2009), less

attention has been given to how this clickstream might be structured in order to represent

consumers’ motivation in online actions that lead to purchases.

The current research takes a new modeling approach to studying browsing patterns. By

combining the analysis of four broad clickstream data sets of different industries, we develop

the micro-journey as influential browsing pattern and document its impact on customer

conversions, while also shedding light on important characteristics of the micro-journey that

are connected to previous research in the field.

Our findings make several contributions to the existing literature. First, they contribute

to the ongoing debate about which clicking actions within the consumer’s online journey are

important and worth focusing on. Building from flow theory, we conceptualize “focused

attention” as the user’s browsing pattern by summarizing intense browsing sessions measured

by successive single clicks that occur within short time intervals. Implementing the time

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 114

component helps us to derive the concept of the micro-journey that ultimately uncovers hidden

purchase intentions of consumers.

Second, our results illustrate that the underlying motive of consumers’ online actions

can be mapped onto their clickstream behavior. As they move through the web, leaving a

complex number of marks, the micro-journey is a starting point to simplify and structure these

traces. Consistent with our theorizing, focused browsing patterns in the form of micro-journeys

better predict online conversions. To account for the complex online environment with its

manifold touchpoints, we detail the micro-journey covariate and amend further predictors that

represent its characteristics. Relying on extant research into browsing behavior, we transfer

effective properties into the context of micro-journeys, thereby connecting our novel concept

to existing categorization approaches. Users with micro-journeys are more likely to convert—

they convert directly after the micro-journey, or, in equal proportions, at a time or a number of

clicks after. Furthermore, we find that relative position within the overall journey, as well as

category switches (e.g., from informational to navigational contacts), are especially suitable to

predict converting customers.

Demonstrating the micro-journey effects across four large-scale individual user-level

data sets underscores their generality and shows systemic differences among industries. For

instance, the effect of the micro-journey position (within the journey) on later purchases may

be moderated by industry-specific influences, such as price sensitivity.23 Furthermore, although

not a focus of our analysis, our modeling approach also adds to the literature by demonstrating

that proportional hazard models are an adequate alternative for predicting user conversions as

it well reflects the chronological nature of clickstream data. While our results derived from the

logit model confirm prior research (Klapdor et al. 2015), the predictor variables of the Cox

23 We discuss generalizations and industry-specific findings in more detail in the Appendix section.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 115

model provide more details with regard to click order. Thus, we conclude that channel switches

(no matter what type), that occur within a micro-journey are positively related to purchase

events. One possible explanation of the Cox models’ results is that more concentrated browsing

sessions may either indicate approaching a purchase transaction (navigational) or reflect

processing of information acquisition (informational). Furthermore, the positive effect of the

micro-journey may outweigh the rather negative tendencies of some switches.

3.7.1 Theoretical Implications

This research links psychological and modeling approaches to study of consumers’

clickstreams. Prior research has used aggregated input measures (such as advertising

impressions or advertising budgets), and has also focused on multiple online channels (e.g.,

Breuer, Brettel, and Engelen 2011). However, none of these studies examines the relationship

between navigational decisions (or clicks) and purchasing. Macro-level collective outcomes

(such as conversion rates) also depend on micro-level individual decisions about what to click

on (Tellis and Franses 2006). Consequently, when trying to understand collective outcomes, it

is important to consider the underlying individual-level psychological processes that drive

online conversions. Along these lines, our research suggests that the micro-journey as a proxy

for a purchase motivation helps to determine which consumers succeed on the path to purchase.

Our findings also suggest that a direct link exists from clicking to buying, as we build

on a directed-buying perspective that is characterized by a customer’s tendency to exhibit very

focused browsing patterns, indicative of the intense goal-directed (i.e., concentrated)

motivation of the customer to purchase a product online. We show that this basic psychological

mechanism antecedes most purchase decisions in an online context, and can be represented by

customers’ clicking behaviors. Consistent with the notion that psychological impulses drive

consumer purchase decisions, the micro-journey acts as a lens for visualizing consumers’

internal states while browsing the web.

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 116

It is also worthwhile to consider these findings in relation to the literature on effective

properties or characteristics of browsing behavior. Just as certain characteristics of browsing

behavior in general may improve the prediction of conversions, certain characteristics of micro-

journeys may cause consumers to be more likely to convert. While there is likely some overlap

in these factors (e.g., the total number of different channels within a full journey has a

significant and strong positive effect on conversion; Naik and Raman 2003; Wiesel, Pauwels,

and Arts 2011), there may also be some important differences. For example, the effect of the

number of total clicks within a full journey on purchases is rather small. Consequently,

controlling for the length of the journey in clicks takes a subordinate role in forecasting purchase

events. Thus, not only the number of clicks is important but also the time component of click

occurrences (i.e., user interaction).

3.7.2 Marketing Implications

These findings also have important marketing implications. Considering the micro-journey as

browsing click pattern should help companies to maximize revenue when placing

advertisements. For instance, identifying users prone to convert may help to improve targeting

as well as bidding decisions in real-time bidding auctions of advertising exchanges. In addition,

it might help online content providers when pricing access to different forms of content (e.g.,

potentially charging more for content that is provided during a micro-journey). Our findings

also shed light on the question of the exact point in time at which micro-journeys indicate a

purchase or, stated differently, the question of whether users show micro-journeys in their

browsing patterns whenever they finalize their purchase decision or they simply acquire

information first, indicated by a micro-journey, and procrastinate about their purchase decision

(O’Donoghue and Rabin 1999). This differentiation is especially relevant as marketers may

utilize micro-journeys, which they can observe while tracking their potential customers, to

adapt their marketing measures toward individual users. In the latter, when users use micro-

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 117

journeys only to gather additional information, marketers should know whether these users are

still relevant for marketing exposure.

3.7.3 Directions for Further Research

Noting that the cookie tracking technologies used in this work focused on individual devices

(e.g., Flosi et al. 2013), future research might examine the effectiveness of micro-journeys on

the basis of actual users. At the single-user level, the effect of the micro-journey may even be

amplified, as the psychological mechanism can be traced back directly to one single person,

rather than to a device that can be used by multiple persons. In addition, a separation of devices

might be a fruitful avenue for further research: Although our data do not allow us to speak to

individual device separation in any detail, the influence of the micro-journeys and their

characteristics may differ when tracked from different devices. For instance, mobile phones are

likely to be used only by one person, whereas desktop computers may have multiple users. In

the same vein, the effect of the micro-journey might be stronger for desktop or laptop usage, as

these devices are regularly accessed when really needed, for instance, for purchasing a product

or service (Watson et al. 2002).

Further research might also examine how the effects observed here are moderated by

situational factors. Interesting candidates in this context would be investigation of how a season

or specific seasonal events affect online browsing and purchasing. Before Christmas,

Thanksgiving, or other “directed-buying events” such as Cyber Monday (i.e., events when

ecommerce spending levels are very high), the focused browsing pattern in the form of the

micro-journey may be even more influential.

Another important factor to consider is the effect of competitors’ websites. We do not

observe journeys (converting) at web shops other than the ones that we track. Thus, we track

neither cross-advertiser effects nor online/offline purchases (Wiesel, Pauwels, and Arts 2011).

Essay 2—Browsing Click Patterns as Micro-Journeys Influence Customer Conversions 118

Nevertheless, as we suggest an underlying psychological mechanism represented by the micro-

journey, future work might focus on the entire activity of consumers on the web.

Finally, these findings also raise broader questions, such as the ability to transfer

psychological mechanisms to models of marketing effectiveness. Although the current results

highlight the robustness of the micro-journey effects, and thereby support our theoretical

derivation, other candidates, such as purchase decision involvement (e.g., Beatty and Smith

1987), might also lead to this specific browsing pattern. The value of better understanding the

psychological mechanisms at work in the minds of consumers as they move toward online

purchases is clear, and deserves further research.

Essay 3—Users Browsing Preferences on the Path to Purchase 121

4 Channels and Categories: User Browsing Preferences on the Path to

Purchase

Ingo Becker, Florian von Wangenheim

Advertisers today employ multiple online channels to draw the users’ attention to their products

or services. In parallel, users, on their path to purchase, may pursue or ignore the various

marketing channels. Such possibilities for browsing, however, create complexities for

interpreting users’ channel selection, and the relationship between browsing traits and purchase

events. In response to handle these complexities, scholars have introduced studies on dedicated,

often channel-specific phenomena, and have analyzed channels grouped into distinct channel

categories. Extending this research, we aim to elucidate users’ actual channel preferences in

multichannel settings, by analyzing the effects of present and accumulated past channel

exposure on purchase propensity. Moreover, by implementing interaction effects between

present and past channel contacts, we explore whether users exhibit homogeneous

(idiosyncratic) or heterogeneous (multi)channel preferences on their path to purchase—and

expand this to channel categories. Applying a proportional hazards model to four unique data

sets reveals generalizable insights into multichannel research, such as clear idiosyncratic

channel preferences of users, or, with regard to channel taxonomies, shows that past exposures

to informational channels predispose subsequent purchases. Our study thus contributes to

marketing effectiveness research, and supports advertisers in optimizing their online marketing

measures.

Essay 3—Users Browsing Preferences on the Path to Purchase 122

4.1 Introduction

Online multichannel marketing refers to the practice of simultaneously offering customers

information, goods, services and support through two or more synchronized channels.

Managing firm–customers interactions in the context of integrated communication strategies

and multichannel management—including, determination of, which channel exposures (e.g.,

paid search, branded search) precede a conversion—has become a cornerstone of online

marketing strategy (Yadav and Pavlou 2014). Thus, developing a deeper understanding of

online channel preferences, the interplay of channels and how channels relate to the customers’

browsing intentions in multichannel settings have become increasingly important for marketing

managers and scholars, also driven by the fact that global business-to-customer (B2C)

ecommerce sales have demonstrated double-digit growth levels surpassing USD 1.7 trillion in

2015 (eMarketer 2014b).

Cross-channel research, focused on direct marketing retailers and brick-and-mortar

stores, suggests that multichannel marketing (comprising, for instance, firm- and customer-

initiated online channels, or offline channels such as print, radio, and TV), constitutes

synergetic cross-channel effects (e.g., spillover effects) that increase marketing effectiveness

(e.g., Edell and Keller 1989; Jagpal 1981; Klapdor et al. 2015; Naik and Raman 2003; Wiesel,

Pauwels, and Arts 2011). While prior offline and hybrid (i.e., online/offline) research shows

that multiple channels may prompt sales synergies (e.g., Chang and Thorson 2004; Edell and

Keller 1989; Naik and Raman 2003), conclusions drawn from pure online multichannel

research are incomplete (Breuer, Brettel, and Engelen 2011). For instance, Jagpal (1981)

propose that simultaneous newspaper and radio advertisements exhibit synergetic effects.

Multiple message sources are perceived as more credible and increase processing motivation

that, in turn, elevates brand recognition and purchase intent (MacInnis and Jaworski 1989; Petty

and Cacioppo 1986). Naik and Raman (2003) analyze television-print synergies and propose

that a marketing medium comprises two relevant effects: Catalyzing sales and stimulating other

Essay 3—Users Browsing Preferences on the Path to Purchase 123

channels. In a survey based study, Chang and Thorson (2004) compare TV and Web synergies

with repetition effects and confirm that the magnitudes of the mentioned effects increase in a

synergetic context. More recently, Wiesel, Pauwels and Arts (2011) analyze the marketing’s

profit impact, and find evidence of multifarious bidirectional cross-channel relationships

between offline and online channels. Although these findings help to understand intermedia

synergies, the research subject is either offline focused or, at least, offline related, or limited to

certain online channels such as search (Rutz and Bucklin 2011), display (Braun and Moe 2013),

or search and display (Kireyev, Pauwels, and Gupta 2013), yet leave channel preferences in

multichannel environments mainly shadowed.

At present, it remains unclear, whether these conclusions from cross-channel research

generalize to a pure online environment, which is typically classified along numerous channels

and categories (see Table 4, Table 31, and the Appendix section). Online customers who use

various channels while browsing (heterogeneous customers), and thus provoke more inter-

channel spillovers, frame the question of whether they are more likely to convert—thereby

becoming more valuable than online customers who stay predominantly within one channel

prior to purchase (homogeneous customers)? Yet the established cross-channel knowledge may

not apply, because in an offline or a hybrid world the media vehicles often differ significantly

in their appearance and sensory stimulus (e.g., radio, print, TV, laptop), while in online settings

the hardware transporting media messages is more constant (e.g., the device’s display).

Interactions between present and past contacts require a minimum of two clicks in a

user’s browsing log, yet, a moderate share of individual customer journeys consists of only one

click. In order to fully capture and approximate the users’ channel preferences, as a first step,

we investigate the role of present and cumulated past channel exposure in commencing

conversion events. Existing research on online marketing effectiveness has has illuminated a

number of phenomena involving singular channels (Ghose and Yang 2009; Jerath, Ma, and

Park 2014; Jerath et al. 2011; Klapdor, von Wangenheim, and Schumann 2014; Rutz and

Essay 3—Users Browsing Preferences on the Path to Purchase 124

Bucklin 2012; Rutz and Trusov 2011; Rutz, Trusov, and Bucklin 2011; Yao and Mela 2011),

and has proven channels affecting one another (e.g., spillover effects), but, mostly in two-

dimensional channel conditions (Kireyev, Pauwels, and Gupta 2013; Rutz and Bucklin 2011;

Xu, Chen, and Whinston 2012; Yang and Ghose 2010). From a more comparative perspective

involving a full set of online channels, less knowledge has been created on present and past

channel exposure, channel spillovers (heterogeneous channel interactions), and channel

carryovers (homogeneous channel interactions), and the actual users’ channel preferences along

the path to purchase. Yet, previous multichannel research on channel taxonomies shows that

the customers’ browsing intentions vary fundamentally between the channel and the channel

category they use (Haan, Wiesel, and Pauwels 2013; Jansen, Booth, and Spink 2008; Klapdor

et al. 2015), preluding the demand to complement extant research by studying the users’ online

channel (and channel category) preferences in a more holistic multichannel setup.

We define a customer browsing preferences as “heterogeneous”, when a customer

shows a clear tendency to utilize more than one online channel (heterogeneous channel

preference) or channel category (heterogeneous category preference) on the browsing path

aiming to conclude in a conversion event. Employing multiple channels, this user group induces

a higher number of channel spillovers, which, in turn may predispose (future) conversion

events. On the contrary, a customer’s browsing preference is denoted “homogeneous”, when a

customer stays primarily with one channel (homogeneous channel preference) or channel

category (homogeneous category preference) prior to the purchase event. Exposure toward

multifarious online channels may be sensed as distraction (Xia and Sudharshan 2002) and, thus,

humper conversions. Furthermore, user may individually favor particular channels. By

conceptualizing online multichannel browsing and purchasing behavior from the customer

angle, our approach provides a holistic view of a customers’ online channel and channel

Essay 3—Users Browsing Preferences on the Path to Purchase 125

category preferences, importantly, while aiming to conclude in a purchase decision.24 To extend

prior research in the multichannel ecommerce context, we address three consecutive research

questions:

1) What does the present online channel exposure (i.e., the user’s click) imply with

regard to the user’s conversion likelihood? Does purchase inclination deviate

between exposures toward particular online channels?

2) What does past channel exposure (recognized in the browsing history) imply

with regard to the user’s conversion likelihood? Does purchase inclination

deviate between (multiple) past exposures toward particular online channels?

3) What users’ actual channel (and channel category) preferences indicate the

formation of an online purchase decision? Is purchase inclination moderated by

carryover (homogeneous interaction) or spillover (heterogeneous interaction)

effects?

Elucidating these question in a holistic multichannel setup allows for adding novel

insights into online marketing effectiveness literature. For instance, the connection between the

customer’s conversion and the customers’ present and past channel exposure(s) becomes most

valuable when compared in a relative context (Danaher and Dagger 2013). If the expected

conversion rate associated with a (customer-initiated) channel like branded paid search is higher

compared to another (firm-initiated) channel like newsletter, marketers may focus their

resources and activities toward this particular channel. At the same time, it is pivotal to

understand the interrelation between marketing channels and how they are moderated by one

another, as customers may exhibit a myriad of browsing histories. Anticipating how these

24 Besides online purchases, the users’ browsing intentions may differ substantially and, along with these

differences, their channel and channel category preferences may deviate as well.

Essay 3—Users Browsing Preferences on the Path to Purchase 126

browsing logs relate to purchase decisions, can be a valuable source for marketers, making these

questions critical from both a theoretical and a managerial point of view.

To elucidate these research questions, we develop a conceptual framework analyzing

the full set of present and past channel clicks. By implementing interaction effects between

present and past exposures, we examine the moderating role of both prior channel exposure and

channel category exposure on browsing (clicking) behavior, and investigate the link to customer

conversions. Building on four large-scale, clickstream data sets from three different industries,

we apply a proportional hazards model, enabling us to derive generalizable empirical insights.

To ensure the highest possible degree of detail, we consider all prior firm-customer interactions,

in contrast to modeling them on a daily level (e.g., Lambrecht and Tucker 2013). Thereby, this

study adds to prior research on online channel effectiveness in a number ways.

First, we contribute novel insights into (online) marketing literature on channel

effectiveness in a multichannel context (Fulgoni and Mörn 2009; Klapdor et al. 2015). For all

data sets, our results show that contrary to cross-channel wisdom (Edell and Keller 1989; Jagpal

1981; Naik and Peters 2009; Naik and Raman 2003; Tellis et al. 2005), in online environments

customers utilizing one preferred channel (or a limited set of channels dominated by one

particular channel) whenever they are prone to convert. These results translate into

homogeneous—rather than heterogeneous—user channel preferences on the path to purchase.

Consequently, monitoring and identifying channel (homogeneous versus heterogeneous)

browsing behavior can route advertisers toward more valuable user segments.

Moreover, we extend knowledge on category approaches in a multichannel setting by

implementing and simultaneously analyzing various well-accepted channel taxonomies and

their interactions (category homogeneous and heterogeneous) into one model (Broder 2002;

Haan, Wiesel, and Pauwels 2013; Rose and Levinson 2004). Combining several theoretical

approaches, our results mirror more diverse channel category user preferences, which adds to

Essay 3—Users Browsing Preferences on the Path to Purchase 127

existing research on channel effectiveness (Yadav and Pavlou 2014). With regard to the contact

origin, homogeneous interactions between present and past firm-initiated channels show

negative, their customer-initiated pendants positive effects on time to purchase, confirming

prior knowledge (Haan, Wiesel, and Pauwels 2013; Wiesel, Pauwels, and Arts 2011). Turning

to the taxonomy of the user’s browsing goal, we find that, independent from the taxonomic

affiliation of the present contacts, their interactions with past informational stock, indicate

positive effects on purchase propensity, complementing prior literature (Jansen, Booth, and

Spink 2008; Klapdor et al. 2015).

In addition, we present novel results on the effectiveness of individual channel clicks in

a more comparative, multichannel online setting (Danaher and Dagger 2013). Across data sets,

customers being more likely to convert rather use customer-initiated contacts such as search or

direct type-ins (i.e. direct website visits), than firm-initiated contacts (Haan, Wiesel, and

Pauwels 2013). With regard to search channels, especially branded search contacts well-reflect

the user’s purchase inclination (Anderl, Schumann, and Kunz 2015; Rutz and Bucklin 2011).

Confirming prior findings, and setting them into a relative context comprising the full set of

online channels employed, we extend research on individual channel effectiveness (Ghose and

Yang 2009; Rutz, Trusov, and Bucklin 2011) as well as related multichannel research in online

(Breuer, Brettel, and Engelen 2011; Li and Kannan 2014), and offline/online environments

(Danaher and Dagger 2013).

Furthermore, these results link to the theory of choice set formation by translating

channel and category exposure along the clickstream into purchase progression or regression

(Campbell 1969; Hauser and Wernerfelt 1990; Howard and Sheth 1969; Howard 1963; Roberts

and Lattin 1991, 1997; Wright and Barbour 1977). Although field data may not perfectly reveal

the customers’ underlying intentions (Shocker et al. 1991), our results well approximate

purchase propensity, as well as purchase reluctance, and, may be interpreted as continuum

Essay 3—Users Browsing Preferences on the Path to Purchase 128

between progression, stagnation and regression along the purchase funnel and purchase

decision making. For instance, channel homogeneous firm-customer interactions may serve as

an agent for progression, multichannel exposure for regression, in the purchase formation

process.

From a broader perspective, these research advancements in multichannel research

generate meaningful insights into user preferences, responding to calls for research that

develops marketing impact models based on individual-level customer path data (Hui, Fader,

and Bradlow 2009; Rust, Lemon, and Zeithaml 2004), that reduces the gap between marketing

theory and practice (Little 2004a; b), and that adds practical cross-industry generalizations and

industry-specific findings (Li and Kannan 2014).

4.2 Conceptual Development

4.2.1 Conceptual Model

We first develop a conceptual model of the relationship between channel effects (Model 1),

category effects (Model 2), and conversions (see Figure 11). In essence, the two main models

include the effects of present and past channel exposures (clicks) and, additionally, the

interaction effects between present and collective past ad exposure (past stock) considering

channel clicks (Model 1) and channel group clicks covering several category approaches

(Model 2). For each of these two main models, we add a base model that excludes the interaction

effects that helps to calibrate the remaining effects.

First, we implement the effect of present firm-customer interactions by channel (Models

1 and 2). Previous research has drawn a multifaceted picture on online channel effectiveness

including singular channels such as search (Ghose and Yang 2009; Rutz, Bucklin, and Sonnier

2012; Rutz, Trusov, and Bucklin 2011) or display marketing (Manchanda et al. 2006; Rutz and

Bucklin 2012), and, less frequently, including two or more channels (Li and Kannan 2014; Xu,

Duan, and Whinston 2014). To expand on that, we model and interpret the complete set of

Essay 3—Users Browsing Preferences on the Path to Purchase 129

channels applied in our data sets. Setting the foundation for subsequent analyses on interaction

effects, we capture the effectiveness of present channel contacts also reflecting one-click

journeys, which do not include interactions, but that account for a substantial share of customer

journeys. The conclusions we draw on the effectiveness of individual online channels—in a

comparative context including the full set of online channels—extend extant research on offline

and online channel effectiveness (Danaher and Dagger 2013).

Next, representing user browsing history, we capture the number of past channel

contacts (past channel stock) separating channel homogeneous variables (Models 1 and 2) and

channel heterogeneous variables (Model 1). Preceded research has illustrated that past channel

exposure may influence future firm-customer touchpoints and journey outcomes (Braun and

Moe 2013; Breuer and Brettel 2012; Klapdor et al. 2015), and, thus, should be taken into

account when analyzing the effectiveness of current marketing measures (Li and Kannan 2014).

Also, the number of previous firm-customer contacts may carry valuable information on

purchase likelihood (Klapdor et al. 2015; Pedrick and Zufryden 1991; Tellis 1988).

In Model 1, with a focus on channel preferences, we further implement interaction

effects between present channel exposure and past homogeneous (and heterogeneous) channel

exposure (past channel stock). While the present channel effects omit the user’s browsing

history, the aggregated past channel effects persist independent from the subsequent (present)

channel exposure. Thus, these interactions are well-suited to extract the effects of present

channel exposure and their antecedents.

In Model 2 we investigate on user preferences for using channel categories. Thus, within

one model, we include four relevant category approaches, and compute—for each approach—

interaction effects between present channel group exposure and past homogeneous (and

heterogeneous) channel group exposure (past category stock). For instance, with regard to the

contacts’ origin, we model present customer-initiated contacts, and add interactions with past

Essay 3—Users Browsing Preferences on the Path to Purchase 130

customer-initiated stock (channel group homogeneous) as well as with past firm-initiated stock

(channel group heterogeneous)—analogous to present firm-initiated contacts. This method

applies to all taxonomies analyzed.25 As categorization approaches consolidate online channels

in channel groups, some overlap may exist between different channel taxonomies (Table 21).

To reduce potential multicollinearity we refrain from implementing present and past channel

group exposure as sole effects, but we, capture these effects by retaining the effects on present

and past channel exposure.

Figure 11

Conceptual Model of Relationships between Channels, Categories, and Conversions—Essay 3

25 The terms “taxonomy” and “category approach” are utilized synonymously. The term “channel group” describes

a dedicated channel group within a taxonomy, for instance, the group of customer-initiated channels within the taxonomy of contact origin.

Essay 3—Users Browsing Preferences on the Path to Purchase 131

4.2.2 Categorization Approaches

In order to make complex data more tractable and to better understand the online purchase

decision process, prior research has introduced various taxonomies to categorize channels

according to certain characteristics into channel groups (e.g., Broder 2002). Adding to our study

on user channel preferences, we investigate whether (sequential) exposure to channels grouped

into categories would lead to likelihood to purchase. Thus, we assign each channel according

to its category affiliation, implementing four different approaches: (1) contact origin, (2)

browsing goal, (3) content integration, and (4) personalization. In Table 21, we illustrate the

channel-category affiliation.

Table 21

Categorization of Online Channels—Essay 3

Channel / Source Browsing goal Contact origin

Degree of content integration

Degree of personalization

Affiliate Informational Firm-initiated Integrated Non-personalized

Display Informational Firm-initiated Separated Non-personalized

Newsletter Navigational Firm-initiated Separated Non-personalized

Price Comparison Informational Customer-initiated Integrated Non-personalized

Referrer Informational Customer-initiated Integrated Non-personalized

Retargeting Informational Firm-initiated Separated Personalized

SEA Generic Informational Customer-initiated Separated Personalized

SEA Brand Navigational Customer-initiated Separated Personalized

SEO Generic Informational Customer-initiated Integrated Personalized

SEO Brand Navigational Customer-initiated Integrated Personalized

Social Media Informational Firm-initiated Separated Non-personalized

Direct Type-in Navigational Customer-initiated Separated Personalized

Other Informational Firm-initiated Separated Non-personalized

Note: The table provides established channel-category linkages, although some links may be controversial. For instance, affiliate and referrer may be customer-initiated in some cases (e.g., for coupon websites); retargeting may be treated as informational or navigational contact as users may themselves choose to re-visit a website (navigational), instead of being recalled only (informational); newsletter can be personalized and non-personalized.

In online environments, because advertisements are not only directed to customers, but

are also initiated by customers, the contact origin may act as a relevant differentiator of

marketing channels (Shankar and Malthouse 2007). Customer-initiated contacts (CICs) are

Essay 3—Users Browsing Preferences on the Path to Purchase 132

generally more effective than firm-initiated contacts (FICs) (Haan, Wiesel, and Pauwels 2013;

Wiesel, Pauwels, and Arts 2011), as they are based on customers’ own actions, and are

perceived to be less intrusive (Shankar and Malthouse 2007). However, firm-initiated media

allow advertisers to intervene in ongoing journeys in order to (re)activate them (Haan, Wiesel,

and Pauwels 2013; Li and Kannan 2014).

Analyzing users’ browsing goals, Broder (2002) differentiates informational and

navigational contacts. The user goal is interpreted as informational, if he or she wants “to learn

something by reading [..] web pages” (Rose and Levinson 2004, p.15), and, as navigational, if

he or she accesses a web page intentionally (Broder 2002). Based on a logit model, Klapdor et

al. (2015) find that a channel sequence (switch) from informational to navigational channels

positively affects purchase propensity, arguing that the user has narrowed down the choice set

on the path to purchase.

Moreover, Haan, Wiesel, and Pauwels (2013) builds on a categorization approach based

on the degree of content-integration. Content-integrated marketing activities include channels

that appear as integral part of a website, for instance, advertisements on price comparison

websites. In contrast, content-separated contacts are only tangentially related to the editorial

content and format of the website such as display advertisements (Haan, Wiesel, and Pauwels

2013). In their study, Haan, Wiesel, and Pauwels (2013) find indications that content-integrated

contacts are more powerful than content-separated contacts in driving purchase funnel

progression.

Finally, technological advancements allow for targeting and, thus, personalization of

online advertising messages (Pavlou and Stewart 2000; Varadarajan and Yadav 2009). While

personalized marketing messages are individualized based on the user’s (prior) browsing

behavior or disclosed characteristics, non-personalized messages—intended for a broad

audience—are identical. Thus, personalized advertisements include retargeted display

Essay 3—Users Browsing Preferences on the Path to Purchase 133

advertisements, as well as, SEO and SEA, as search results originate from individually entered

search terms by the user. Interestingly, in a study on the effectiveness of retargeting, Lambrecht

and Tucker (2013) find that generic retargeted display advertisements are more effective than

their more specific, dynamically retargeted equivalents.

As the categorization approach decollating generic and branded contacts is limited to

customer-initiated channels (Anderl, Schumann, and Kunz 2015), we separate the search

channels in accordance with their brand (generic) affiliation. Table 22 provides the categorical

proportions of clicks by data set. The deviations between data sets reflect the advertisers’

channel preferences.

Table 22

Description of the Data Sets, Including Categories—Essay 3

Data Set 1 Data Set 2 Data Set 3 Data Set 4 Industry Fashion Fashion Luggage Travel Number of clicks 1,635,724 1,122,838 601,417 1,380,190

Thereof customer-initiated clicks 781,847 866,333 592,699 533,059

47.8% 77.2% 98.6% 38.6%

Thereof firm-initiated clicks 853,877 256,505 8,718 847,131

52.2% 22.8% 1.4% 61.4%

Thereof navigational clicks 544,309 292,938 57,865 99,123

33.3% 26.1% 9.6% 7.2%

Thereof informational clicks 1,091,415 829,900 543,552 1,281,067

66.7% 73.9% 90.4% 92.8%

Thereof content-integrated clicks 1,025,834 496,415 97,790 227,719

62.7% 44.2% 16.3% 16.5%

Thereof content-separated clicks 609,890 626,423 503,627 1,152,471

37.3% 55.8% 83.7% 83.5%

Thereof personalized clicks 729,307 742,412 569,664 448,454

44.6% 66.1% 94.7% 32.5%

Thereof non-personalized clicks 906,417 380,426 31,753 931,736

55.4% 33.9% 5.3% 67.5%

4.3 The Four Data Sets

We base our analyses on four sets of clickstream data provided by online advertisers and in

close cooperation with a multichannel tracking provider. Clickstreams are a collection of data

records tracing the path individual users take while browsing the Internet (Bucklin and Sismeiro

Essay 3—Users Browsing Preferences on the Path to Purchase 134

2009). On an individual user level, for each visit to the advertiser’s website our data sets store

the source of the click and a timestamp that is accurate to the second. The source of the click

represents either a user click on an advertising exposure, thus, tracking the channel or a direct

type-in of the advertiser’s URL into the address bar.26 Furthermore, the data records whether a

visit was followed by a conversion event or did not lead to a conversion within a time period of

30 days, which is defined as a non-converting journey. As, in rare cases, conversions may occur

outside this time span, the data sets may contain right-censored data points. Technical data

collection is based on cookie-tracking, and delineates individual devices, approximating

individual users. Although cookie data may be inaccurate in tracking multi-device usage by a

single user or multiple users utilizing a single device and may be subject to cookie deletion

(Flosi, Fulgoni, and Vollman 2013), cookie tracking remains the industry standard to collect

clickstreams (Tucker, 2012).

We selected pure online players as data providers for this study, facilitating exclusion

of online/offline cross-channel effects, as measuring individual-level exposure to multiple

offline media proves highly difficult in practice (Danaher and Dagger 2013). The advertisers

operate in three different industries, thereby supporting generalizations and cross-industry

comparisons. They include three online retailers selling fashion and luggage products, and one

online travel agency. The data sets comprise a minimum of 405,343 individual user journeys

and up to 1,184,582, with a conversion rate oscillating between 0.9% and 2.0%. Depending on

the advertiser, the data sets include from nine to eleven different channels, identified as affiliate,

display, price comparison, newsletter, referrer, retargeting, generic SEA, branded SEA, generic

SEO, branded SEO, social media, direct type-in, and other, which is specified as advertisements

26 In this manuscript, we use the term “channel” and “source” synonymously including both, channel exposures

and direct type-in.

Essay 3—Users Browsing Preferences on the Path to Purchase 135

that may not be designated as one of these sources.27 Search contacts in which the user types in

the brand name of the advertiser, including misspellings, are coded as branded search contacts

following research standard (Jansen, Booth, and Spink 2008). In Table 23, we present a detailed

overview of the data sets.

Table 23

Descriptive Statistics of the Data Sets, Including Channels—Essay 3

Data Set 1 Data Set 2 Data Set 3 Data Set 4 Industry Fashion Fashion Luggage Travel Number of different channels 11 11 10 9Number of different channels analyzed 10 8 9 9Number of journeys 1,184,582 862,114 405,343 600,872

Journey length in clicks 1.38 1.30 1.48 2.30

(1.88) (1.23) (1.28) (5.20)

Number of different channels per journey

1.06 1.06 1.09 1.10

(0.32) (0.27) (0.33) (0.33)

Number of conversions 10,153 16,201 8,117 9,861Journey conversion rate 0.86% 1.88% 2.00% 1.64%Number of clicks 1,635,724 1,122,838 601,417 1,380,190

Thereof Affiliate 754,355 2 1,699 36,487Thereof Display 10 - - 787,743Thereof Newsletter 79,453 12 8 17,204Thereof Price Comparison 2,205 - 15,003 90,302Thereof Referrer 70,394 123,952 15,043 -Thereof Retargeting 20,059 31 7,011 5,697Thereof Paid Search Generic 120,542 119,374 442,617 287,346Thereof Paid Search Branded 50,982 37,493 8,284 54,481Thereof Unpaid Search Generic 123,850 330,081 62,179 73,492Thereof Unpaid Search Branded 75,030 42,380 3,866 27,438Thereof Direct Type-in 338,844 213,053 45,707 -Thereof Social Media - 241,462 - -Thereof Other - 14,998 - -

Note: Standard deviations are in parentheses. The different number of channels by advertiser is derived from the individual channel propensity and selection of each advertiser. Rare channels are removed from the models, as they are underrepresented and insignificant (DS 1: Display; DS 2: Affiliate, Newsletter, Retargeting; DS 3: Newsletter).

In our data sets, Search Engine Advertising (SEA) signifies paid ads on Google’s search

engine, and Search Engine Optimization (SEO) refers to an unpaid, organic search on Google.

27 We provide definitions of the online channels analyzed in the Appendix section.

Essay 3—Users Browsing Preferences on the Path to Purchase 136

Both appear in all four data sets and are separated into different channels for branded and

generic search terms (Jansen, Booth, and Spink 2008), because they reflect different user

browsing states and are subject for interaction effects (Rutz and Bucklin 2011). We provide

detailed definitions of the channels in the Appendix section. The frequency of channel exposure

varies considerably across the four data sets. For example, whereas affiliate accounts for about

46% of all clicks in Data Set 1, its relative share in Data Set 4 is nearly 3% of the clicks. This

variation alleviates endogeneity concerns. To further rule out potential endogeneity, we ran the

models that exclude retargeting and newsletter marketing, as these may interrelate with

previous website visits.

4.4 Model Development

4.4.1 General Model Formulation

The aim of subsequent analyses on customer channel preferences is threefold: First, we focus

on the effects of individual channels on conversion events. Second, we include channel

interactions to examine the effects of homogenous/heterogeneous channel usage along the

journey to purchase events (Model 1). Third, we model category interactions to understand if

and how the exposure to specific channel categories along the browsing path affects conversion

events (Model 2).

To account for right-censoring and to reflect the sequential nature of the path data, we

applied a Cox proportional hazards model (Cox 1972; Seetharaman and Chintagunta 2003).

Proportional hazards models have been widely applied in medical sciences (David Collett

2015), and, more occasionally, in online marketing research, for instance, in the context of

display marketing (Manchanda et al. 2006) and retargeted display marketing (Lambrecht and

Tucker 2013). The dependent variable in proportional hazards models is time T leading up to

the occurrence of an event—here, a binary conversion event: purchase versus no purchase. The

model formula describes the hazard at time t as the product of two quantities—first, the baseline

Essay 3—Users Browsing Preferences on the Path to Purchase 137

hazard, h0(t), which defines the hazard per time unit t at the baseline of the covariates and,

second, the exponential expression e to the linear sum of βiX i, with the sum over p predictor X

covariates, which defines the responsive effect of the predictor covariates on the hazard. As our

tests suggest that the proportional hazards assumption holds across data sets and covariates, we

model the covariates time-independent, which is reflected in the second component, the

exponential expression e in the model formula. It excludes the time component t, defining the

vector of the explanatory X covariates as time-independent. As we are unaware of the particular

form of the underlying hazard, we applied a semiparametric Cox model (Cox 1972), which

derives the hazard as multiplicative replica directly from the data, thus, increasing flexibility

(Seetharaman and Chintagunta 2003). 28 The formula of the Cox model applied for user i h6!t, X$

is defined as,

h6!t, X$ = h9!t$ × exp !1 β?X6?$@

? �,, ( 10 )

with X = (X1, X2, …Xp) predictor or explanatory variables for customer i. Following

prior research to disentangle the data to the most disaggregated and feasible units (Tellis and

Franses 2006), we estimate our models based on individual customer journeys and aggregate

the covariate time T until conversion over days—being in line with prior research (Lambrecht

and Tucker 2013). The vector of covariates in our base Model 1 for user i is specified as follows:

exp !1 βX6?$@

? �= exp�βX6� + βX6# + ⋯ + βX6? + βPastX6� + βPastX6# + ⋯

+ βPastX6? + βChannelX6� + βChannelX6# + ⋯ + βChannelX6?�.

( 11 )

28 We control for a time-dependent specification of the Cox model as well as parametric models including

exponential, Weibull, and Gompertz distribution, all leading to comparable estimation results.

Essay 3—Users Browsing Preferences on the Path to Purchase 138

While the baseline hazard, h0(t), captures the time effect, aggregated over days, the

vector of covariates, captures the effect of channel exposures, past homogeneous channel

exposures and past heterogeneous (multi)channel exposures. In particular, Xij signifies present

clicks on a channel j, for instance, the present affiliate clicks, modeled as continuous covariate.

Furthermore, PastXij, corresponds to the total number of clicks on a particular channel before

each present click, thus, collecting preceding homogeneous channel exposures (idiosyncratic

channel preferences). Similarly, ChannelXij, captures the number of different channels the user

is exposed to before each click, importantly, excluding the channel of the present click. In other

words, the covariate, ChannelXij, captures the heterogeneous (multi)channel occurrences prior

to a click on channel j—apart from channel j—, thus, excluding homogeneous channel

exposures (idiosyncratic channel preferences). An important feature of this definition is that we

model past channel exposures by their actual occurrence including, past exposures both before

and on the same day, which is unique and contrary to prior studies, only including exposures

prior to the corresponding day (e.g., Lambrecht and Tucker 2013). Thereby we capture the

effect of prior exposures more holistically, as our data sets convey multiple (homogeneous and

heterogeneous) contacts within the same day. That is especially relevant when modeling

interaction effects between past and present advertisement exposures in the main models.

4.4.2 Modeling Interaction Effects

In order to better understand user channel preferences, we extend the base model by adding

interaction effects. While the base model includes predictors for present and past channel

exposures, the vector of covariates in the main Model 1 further captures interaction effects

between present channel exposures and accumulated past channel exposures—including both

prior homogeneous and heterogeneous channel contacts:

Essay 3—Users Browsing Preferences on the Path to Purchase 139

exp !1 βX6?$@

? �= exp�βX6� + βX6# + ⋯ + βX6? + βPastX6� + βPastX6# + ⋯

+ βPastX6? + βChannelX6� + βChannelX6# + ⋯ + βChannelX6?

+ βX6� × βPastX6� + βX6# × βPastX6# + ⋯ + βX6? × βPastX6?

+ βX6� × βChannelX6� + βX6# × βChannelX6# + ⋯ + βX6?

× βChannelX6?�.

( 12 )

In this way we measure the effects of idiosyncratic and multichannel user preferences

on their path to purchase. It is worth to note, that we do not explicitly model all possible bilateral

channel interactions, due to the raise of dimensionality (Bellman 1961).

4.4.3 Modeling Category Effects

As a next step, we measure the effect of channel sequences belonging to the same channel group

and to the complementary channel group for each taxonomy (Model 2). Therefore, we model

both present and past channel exposures covering all categories, and further include the

interaction effects between channel groups. The vector of covariates is defined by the following

formula:

exp !1 βX6?$@

? �= exp�βX6� + βX6# + ⋯ + βX6? + βPastX6� + βPastX6# + ⋯

+ βPastX6? + βCatX6� × PastCatX6� + βCatX6# × PastCatX6# + ⋯

+ βCatX6? × CatX6? + βCatX6� × PastCatX6# + βCatX6#

× PastCatX6� + βCatX63 × PastCatX62 + βCatX62 × PastCatX63

+ ⋯ + βCatX6!?��$ × CatX6? + βCatX6? × CatX6!?��$�.

( 13 )

We measure two different types of categorical interaction effects covering homogeneous

and heterogeneous interaction effects within each taxonomy. Homogeneous effects are defined

as interactions between present clicks and the aggregate of preceded clicks, both from the same

Essay 3—Users Browsing Preferences on the Path to Purchase 140

channel group. Accordingly, heterogeneous effects represent interactions between present

clicks from a particular channel group with the sum of preceded clicks of the antagonistic

channel group. For instance, regarding the contact origin taxonomy, a homogenous category

interaction effect describes the effect of interaction between a CIC and the sum of prior CICs,

while a heterogeneous interaction effect includes the effect between a CIC and the sum of prior

FICs—and vice versa. To be precise, we exclude inter category interactions as channel groups

from different taxonomies may include identical channels. Thus, this overlap may involve the

measurement of homogenous and heterogeneous channel group interactions at the same time.

4.5 Estimation Results

We concentrate our interpretation on the main Models 1 and 2, as they capture all effects of the

base models, but also include the interaction effects. Yet, we report the estimation results from

all models including the base models across data sets in the corresponding Tables. The

interpretation of the coefficients is contextual—the data sets do not include information on users

who did not interact with the focal advertiser’s website. Accordingly, a negative coefficient

does not necessarily imply that the corresponding clicking behavior of a user represents a lower

purchase propensity than someone who never visited the website. Furthermore, the magnitude

of a particular effect is not fully comparable between data sets as all data sets are unique insofar

as they may comprise different channels, channel fractions, and different creative content.

4.5.1 Present Channel Effects

Prior studies have substantially investigated the effectiveness of dedicated online channels.

Adding to that, we analyze the (relative) effectiveness of actual channel clicks in a multichannel

context. These findings may serve as guide in calibrating real-time bidding decisions.

The effects of present website visits through affiliate, social media, branded paid and

branded unpaid search, as well as direct type-ins, are consistently positive across data sets. In

contrast, the effects of present clicks on display advertisements and referrer tend to be negative.

Essay 3—Users Browsing Preferences on the Path to Purchase 141

For the remaining channels—newsletter, price comparison, retargeting, and generic paid and

generic unpaid search, the picture is not so clear cut across data sets (Table 24).29

Table 24

Estimation Results: Present Channel Effects (Part 1/3)—Essay 3

DS 1 (Fashion) DS 2 (Fashion) DS 3 (Luggage) DS 4 (Travel) Base Model 1 Base Model 1 Base Model 1 Base Model 1 Variable B B B B B B B B Affiliate 0.033 *** 0.061*** 0.297*** 0.285*** 0.180*** 0.200***

Display -3.014*** -3.426***

Newsletter 0.092 *** 0.090*** -0.090** -0.091**

PriceComparison 0.082 -1.651*** 0.156*** 0.152*** 0.015 0.036**

Referrer -0.024 -0.122 -0.063 -0.196** 0.126 -1.060**

Retargeting 0.063 0.148*** 0.157** 0.089 -0.171* -0.182*

SEAGeneric 0.188 *** 0.299*** 0.094*** -0.033 -0.059*** -0.057*** -0.087*** -0.029SEABranded 0.263 *** 0.244*** 0.352*** 0.462*** 0.259*** 0.205*** 0.193*** 0.199***

SEOGeneric 0.009 -0.047 0.187*** 0.215*** 0.051 0.074* -0.325*** -0.353***

SEOBranded 0.151 *** 0.133*** 0.239*** 0.252*** 0.153*** 0.141* 0.044*** 0.104***

Social 0.097*** 0.092***

TypeIn 0.162 *** 0.126*** 0.121*** 0.061* 0.070 0.003

Other 0.131*** 0.225***

…

N 1,184,582 1,184,582 862,114 862,114 405,343 405,343 600,873 600,873 Observations 1,398,267 1,398,267 964,836 964,836 461,108 461,108 792,345 792,345 Time at Risk 1,863,964 1,863,964 1,306,432 1,306,432 655,963 655,963 1,171,897 1,171,897 Log Likelihood -109,770.5 -109,369.2 -179,987.9 -179,763.3 -91,447.8 -91,235.8 -108,482.5 -108,236.3 AIC 219,600.9 218,838.4 360,023.8 359,606.6 182,949.6 182,561.7 217,019.0 216,562.6 BIC 219,965.4 219,446.0 360,306.5 360,077.8 183,247.7 183,058.5 217,331.8 217,083.8 R2 (D) 0.379 0.401 0.185 0.188 0.391 0.391 0.576 0.581 R2 (PH) 0.354 0.391 0.153 0.169 0.406 0.429 0.803 0.812 Note: * p < .10, ** p < .05, *** p < .01; This table mirrors only a part of the full table. For a complete assay, the full table is available in the Appendix section. The base model excludes interaction effects. Model 1 includes interaction effects.

Affiliates serve as a gateway to a website. Because users already anticipate where they

are being transferred, for users who have a definite idea about their purchase intent, there is

potential to encourage use of an affiliate contact on their path to purchase (Data Set 1: Affiliate

b = 0.061, p < .01; Data Set 3: Affiliate b = 0.285, p < .01; Data Set 4: Affiliate b = 0.200, p <

.01).30 In social media, for instance, the user who is redirected to an advertiser’s offering must

29 The large number of covariates and the resulting space necessity and its restrictions, we report only the relevant

parts of the estimation results and provide the full tables additionally in the Appendix section (Table 30 and Table 31). Thereby, we maximize readability without missing any details.

30 We report the results of main models in the continuous text body. The results of the base models can be reviewed in the corresponding tables.

Essay 3—Users Browsing Preferences on the Path to Purchase 142

leave the current activity. However, if a user decides to leave the current location in order to

follow an advertisement, it must have raised strong interest or indicated a useful response to a

need, indicating that social media clicks may be a relevant source for anticipating conversion

events (Data Set 2: Social b = 0.092, p < .01). Direct type-ins, in contrast, solely involve users

who are already aware of the existence of a corresponding website and, at least in part, know

what it offers. Thus, more directed users may apply this channel (Data Set 1: TypeIn b = 0.126,

p < .01; Data Set 2: TypeIn b = 0.061, p < .1). The same may hold for branded search including

paid search (e.g., Data Set 1: SEAbranded b = 0.244, p < .01; Data Set 2: SEAbranded b =

0.462, p < .01) and unpaid search, at some lower magnitude (e.g., Data Set 1: SEObranded b =

0.133, p < .01; Data Set 2: SEObranded b = 0.252, p < .01). Interestingly, other rules seem to

apply for generic search. While in our data sets generic search clicks show mostly positive

effects for the online retailers (e.g., Data Set 1: SEAgeneric b = 0.299, p < .01; Data Set 2:

SEOgeneric b = 0.215, p < .01), they reveal negative effects for the travel website (Data Set 4:

SEOgeneric b = -0.353, p < .01). Preceding research on paid search claims that generic paid

search has higher apparent cost compared to branded paid search, thereby creating spillovers to

subsequent branded search (Jansen, Sobel, and Zhang 2011; Rutz and Bucklin 2011). While

our results indicate congruence with the first claim, we do not specifically focus on interactions

between generic and branded search, which may further enhance the advertising value of

generic search, even for the travel company. Notwithstanding these effects, branded search is

more directed (navigational) while generic search is broader (informational), also indicating a

different user state within the purchase deliberation process (Broder 2002; Jansen, Sobel, and

Zhang 2011; Klapdor et al. 2015; Roberts and Lattin 1997). Newsletter indicates positive effects

for retail, more precisely fashion retail, yet negative effects for travel (Data Set 1: Newsletter b

= 0.090, p < .01; Data Set 4: Newsletter b = -0.091, p < .05). This deviation may be explained

by the dissimilar nature of the two product categories. While fashion newsletters may attract

users to take advantage of an often time-limited offering (e.g., sale, discount campaign), the

Essay 3—Users Browsing Preferences on the Path to Purchase 143

acceptance of a travel offering requires more complex preconditions—for example, leaving

work for a holiday period, or group decision making. Therefore, users reacting to newsletters

may simply be curious when viewing the offering or its price, not knowing their actual ability

to accept a special offer. A similar explanation may apply for retargeted display. While

retargeted display may reawaken interest in retail products, travel products are subject to more

external influencers such as time constraints or group decision making (Data Set 1: Retargeting

b = 0.148, p < .01; Data Set 3: Retargeting b = 0.157, p < .05; Data Set 4: Newsletter b = -0.182,

p < .1). Moreover, the coefficients for price comparison websites show deviating signs (Data

Set 1: PriceComparison b = -1.651, p < .01; Data Set 2: PriceComparison b = 0.156, p < .01;

Data Set 4: PriceComparison b = 0.036, p < .05). Accordingly, price competitiveness and degree

of product commoditization may influence user choice (Mehta, Rajiv, and Srinivasan 2003).

With regard to negative effects, display advertising, being initiated by the advertiser (Haan,

Wiesel, and Pauwels 2013), may increase user awareness or interest (e.g., Chatterjee, Hoffman,

and Novak 2003; Ilfeld and Winer 2002; Sherman and Deighton 2001), or may affect changes

in future purchase intent (e.g., Braun and Moe 2013; Fulgoni and Mörn 2009). In the relative

context of our study, however, and with regard to immediate effects, our results suggest that

display advertising is negatively associated with purchase propensity (Data Set 4: Display b =

-3.426, p < .01). Still, display exhibits positive effects, on search, for instance, leading to an

overall increase of ROI (Kireyev, Pauwels, and Gupta 2013). They may be worth investing in

for the longer term, as their effects may last for a comparably extensive period (Breuer, Brettel,

and Engelen 2011). Referral campaigns comprise the second consistently negative channel

effect across data sets (Data Set 2: Referrer b = -0.196, p < .05; Data Set 3: Referrer b = -1.06,

p < .05). The relationship between the referrer and the referred user is closer than in affiliate

marketing, in the sense that the referrer website is not necessarily compensated in a

remuneration model. Thus, because it is not part of the referrer’s business model, the desire to

refer a customer to a web shop may be less pronounced.

Essay 3—Users Browsing Preferences on the Path to Purchase 144

4.5.2 Past Channel Effects

Previous research claims that more user-advertiser interactions (clicks) indicate a higher

purchase propensity (e.g., Klapdor et al. 2015). In this study, we concentrate on channel specific

findings, as the repeated use of particular channels in the browsing history may leave a more

multifaceted picture, detailing out preceded more general, channel unspecific findings.

While the past usage of retargeting seems to positively affect conversions, branded paid

and branded unpaid search, as well as direct type-ins, are negatively associated with

conversions. The remaining channels show results that are more diverse across data set (Table

25).

Table 25

Estimation Results: Past Channel Effects (Part 2/3)—Essay 3

DS 1 (Fashion) DS 2 (Fashion) DS 3 (Luggage) DS 4 (Travel) Base Model 1 Base Model 1 Base Model 1 Base Model 1 Variable B B B B B B B B … PastAffiliate 0.034*** 0.083*** -0.361*** -0.485 *** -0.067*** -0.449***

PastDisplay -0.269* -0.471*

PastNewsletter -0.002 -0.069*** 0.116*** 0.007PastPriceComparison -0.158 -1.554*** 0.049* -0.145 ** 0.046*** 0.172***

PastReferrer 0.128*** 0.061 0.052*** -0.091* 0.059 -1.137 ***

PastRetargeting 0.026*** 0.006 -0.058 -0.135 0.067*** 0.059*

PastSEAGeneric -0.028** 0.015 -0.049*** -0.235*** 0.140*** 0.110 *** 0.052*** 0.139***

PastSEABranded -0.043*** -0.156*** -0.052*** -0.065*** -0.131*** -0.590 *** -0.012** -0.181***

PastSEOGeneric -0.004 -0.053** 0.040*** 0.051*** 0.037** -0.026 0.069*** -0.005PastSEOBranded -0.026*** -0.058*** -0.029*** -0.093*** -0.116*** -0.514 *** -0.008 -0.173***

PastSocial 0.022*** 0.003

PastTypeIn 0.010*** -0.083*** -0.006*** -0.092*** 0.008 -0.123 **

PastOther 0.185*** 0.417***

… N 1,184,582 1,184,582 862,114 862,114 405,343 405,343 600,873 600,873Observations 1,398,2671,398,267 964,836 964,836 461,108 461,108 792,345 792,345Time at Risk 1,863,9641,863,964 1,306,432 1,306,432655,963 655,963 1,171,897 1,171,897Log Likelihood -109,770.5 -109,369.2-179,987.9 -179,763.3 -91,447.8-91,235.8 -108,482.5-108,236.3AIC 219,600.9 218,838.4 360,023.8 359,606.6182,949.6182,561.7 217,019.0 216,562.6BIC 219,965.4 219,446.0 360,306.5 360,077.8183,247.7183,058.5 217,331.8 217,083.8R2 (D) 0.379 0.401 0.185 0.188 0.391 0.391 0.576 0.581R2 (PH) 0.354 0.391 0.153 0.169 0.406 0.429 0.803 0.812Note: * p < .10, ** p < .05, *** p < .01; This table mirrors only a part of the full table. For a complete assay, the full table is available in the Appendix section. The base model excludes interaction effects. Model 1 includes interaction effects.

Essay 3—Users Browsing Preferences on the Path to Purchase 145

On the subject of retargeting, the coefficients reveal a slightly positive effect (Data Set

4: PastRetargeting b = 0.059, p < .1), which is especially interesting for marketers concerned

about frequency capping in the context of retargeted marketing measures. Thus, higher

repetition rates of retargeted banners may re-activate users on their path to purchase. In contrast,

for search, especially branded search (paid and unpaid), a repeated use of branded keywords

indicates that users are either indecisive, or, over time, are losing their purchase interest in a

specific product offered by a web shop (e.g., Data Set 1: PastSEAbranded b = -0.156, p < .01;

Data Set 2: PastSEAbranded b = -0.065, p < .01; Data Set 3: PastSEAbranded b = -0.590, p <

.01; Data Set 4: PastSEAbranded b = -0.181, p < .01). Direct type-ins show similar browsing

and purchasing patterns (Data Set 1: TypeIn b = -0.083, p < .01; Data Set 2: TypeIn b = -0.092,

p < .01; Data Set 3: TypeIn b = -0.123, p < .05). Both branded search and direct type-in are

perceived to be navigational channels with a higher degree of contextual matching, indicating

that the user already has a (pre-)defined idea of what to expect (Broder 2002; Broder et al.

2007). As present contact via the corresponding channels positively influence the time to

conversion, and numerous past equivalent contacts are negative, the user, obviously, decides to

conclude with a purchase event in the short term (after the contact point), or, alternatively, may

be unlikely to purchase at all. The results for generic paid and generic unpaid search are mixed

and depend on the data set (e.g., Data Set 2: PastSEAgeneric b = -0.235, p < .01; Data Set 3:

PastSEAgeneric b = 0.110, p < .01; Data Set 4: PastSEAgeneric b = 0.139, p < .01). These

contacts indicate the informational nature of the browsing stage (Rose and Levinson 2004),

which, in some cases, may indicate information acquisition and, in consequence, may still result

in purchase transactions (e.g., Data Set 2: PastSEOgeneric b = 0.051, p < .01). Although the

negative effects slightly outweigh the positive, affiliate, referrer, price comparison and

newsletter should be specifically analyzed for each advertiser. Hereby, repeated contacts seem

to indicate interest, though do not necessarily direct toward purchase events. For instance, a

customer who frequently uses an affiliate seems indecisive; otherwise, a purchase transaction

Essay 3—Users Browsing Preferences on the Path to Purchase 146

will instead follow on short response (Data Set 1: PastAffiliate b = 0.083, p < .01; Data Set 3:

PastAffiliate b = -0.485, p < .01; Data Set 4: PastAffiliate b = -0.449, p < .01; e.g., Data Set 3:

Affiliate b = 0.285, p < .01).

Across data sets, the results on past repeated contacts are more ambiguous, as the results

on present contacts limit generalized recommendations and increase the necessity for analyzing

these effects on a case-by-case level. Their joint consideration may help in defining rules for

frequency capping in real-time bidding.

4.5.3 Homogenous and Heterogeneous Channel Interactions

Prior offline and hybrid (online/offline) research, has proved that multiple channels prompt

sales synergies (e.g., Chang and Thorson 2004; Edell and Keller 1989; Naik and Raman 2003).

Extending this effect to an online context, we analyze whether users exhibit dedicated

homogeneous or heterogeneous channel preferences. Therefore, we implement homogeneous

and heterogeneous (multichannel) exposure, and interpret the corresponding interaction effects.

Our study results reveal a surprisingly clear picture across all data sets (Table 26). The

effects of inter-channel, homogeneous spillovers become mostly significant and positive (e.g.,

Data Set 1: SEAbranded × PastSEAbranded b = 0.135, p < .01; Data Set 2: SEAgeneric ×

PastSEAgeneric b = 0.347, p < .01; Data Set 3: TypeIn × PastTypeIn b = 0.138, p < .01; Data

Set 4: Affiliate × PastAffiliate b = 0.374, p < .01). The heterogeneous channel interaction effects

with prior multichannel exposure becoming mostly significant and negative (e.g., Data Set 1:

SEAbranded × PastChannelsNoSEAbranded b = -0.284, p < .01; Data Set 2: SEAgeneric ×

PastChannelsNoSEAgeneric b = -0.403, p < .01; Data Set 3: TypeIn × PastChannelsNoTypeIn

b = -0.431, p < .01; Data Set 4: Affiliate × PastChannelsNoAffiliate b = -0.060, p < .05). In

other words, users seem to have preferred online channels on their path to purchase. Multiple

channel exposures, manifested in heterogeneous channel utilization, are perceived as

interruptive, and thus are associated with a lower relative purchase propensity (Xia and

Essay 3—Users Browsing Preferences on the Path to Purchase 147

Sudharshan 2002). In consequence, as users favor a limited set of channels, online marketing

may be understood as marketing channel itself like, for instance, television or print, contrasting

to a multi-facetted agglomeration of individually perceived marketing channels.

Notwithstanding, within online marketing there are multiple different vehicles, that may still

exert influences among one another (e.g., Rutz and Bucklin 2011).

Table 26

Estimation Results: Channel Interactions (Part 3/3)—Essay 3

DS 1 (Fashion) DS 2 (Fashion) DS 3 (Luggage) DS 4 (Travel)

Base Model 1 Base Model 1 Base Model 1 Base Model 1

Variable B B B B B B B B

…

Affiliate × PastAffiliate -0.069*** 0.368 ** 0.374***

Display × PastDisplay 0.333

Newsletter × PastNewsletter 0.080*** 0.225***

PriceComparison × PastPriceComparison 1.990*** 0.226 *** -0.129***

Referrer × PastReferrer 0.071 0.163*** 1.313 ***

Retargeting × PastRetargeting 0.013 0.146 0.036

SEAGeneric × PastSEAGeneric -0.040 0.347*** 0.063 *** -0.086***

SEABranded × PastSEABranded 0.135*** 0.020 0.549 *** 0.178***

SEOGeneric × PastSEOGeneric 0.076** -0.017* 0.106 *** 0.129***

SEOBranded × PastSEOBranded 0.051*** 0.067*** 0.502 *** 0.183***

Social × PastSocial 0.028***

TypeIn × PastTypeIn 0.095*** 0.088*** 0.138 ***

Other × PastOther -0.300***

Affiliate × PastChannelNoAffiliate 0.267*** -0.492 *** -0.060*

Display × PastChannelNoDisplay 0.670***

Newsletter × PastChannelNoNewsletter -0.357*** -0.479***

PriceComparison × PastChannelNoPriceComp. -0.043 -0.531 *** -0.265***

Referrer × PastChannelNoReferrer 0.068 -0.096** -0.295 ***

Retargeting × PastChannelNoRetargeting -0.185*** -0.414 *** -0.148

SEAGeneric × PastChannelNoSEAGeneric -0.142*** -0.403*** -0.611 *** -0.070*

SEABranded × PastChannelNoSEABranded -0.284*** -0.431*** -0.575 *** -0.324***

SEOGeneric × PastChannelNoSEOGeneric -0.015 -0.183*** -0.502 *** -0.321***

SEOBranded × PastChannelNoSEOBranded -0.198*** -0.277*** -0.571 *** -0.482***

Social × PastChannelNoSocial -0.319***

TypeIn × PastChannelNoTypeIn -0.116*** -0.251*** -0.431 ***

Other × PastChannelNoOther -0.207**

N 1,184,582 1,184,582 862,114 862,114 405,343 405,343 600,873 600,873

Observations 1,398,267 1,398,267 964,836 964,836 461,108 461,108 792,345 792,345

Time at Risk 1,863,964 1,863,964 1,306,432 1,306,432 655,963 655,963 1,171,897 1,171,897

Log Likelihood -109,770.5 -109,369.2 -179,987.9-179,763.3 -91,447.8 -91,235.8 -108,482.5 -108,236.3

AIC 219,600.9 218,838.4 360,023.8 359,606.6 182,949.6 182,561.7 217,019.0 216,562.6

BIC 219,965.4 219,446.0 360,306.5 360,077.8 183,247.7 183,058.5 217,331.8 217,083.8

R2 (D) 0.379 0.401 0.185 0.188 0.391 0.391 0.576 0.581

R2 (PH) 0.354 0.391 0.153 0.169 0.406 0.429 0.803 0.812Note: * p < .10, ** p < .05, *** p < .01; This table mirrors only a part of the full table. For a complete assay, the full table is available in the Appendix section. The base model excludes interaction effects. Model 1 includes interaction effects.

Although our findings suggest idiosyncratic channel preferences, on a more detailed

level punctual exceptions may apply. With regard to homogenous channel interactions, the

Essay 3—Users Browsing Preferences on the Path to Purchase 148

coefficients for affiliate (Data Set 1: Affiliate × PastAffiliate b = -0.069, p < .01), price

comparison (Data Set 4: PriceComparison × PastPriceComparison b = -0.129, p < .01), generic

paid search (Data Set 4: SEAgeneric × PastSEAgeneric b = -0.086, p < .01), and generic unpaid

search (Data Set 4: SEOgeneric × PastSEOgeneric b = -0.017, p < .1) become significant and

show a (slight) negative sign. Interestingly, from the perspective of present and past click

exposure equivalents, all these negative effects are associated with significant and positive

effects (e.g., Data Set 1: Affiliate b = 0.061, p < .01; Data Set 1: PastAffiliate b = 0.083, p <

.01). Obviously, these channels may exhibit negative repetitive effects, but the effects are

statistically positive indicators if they are treated as decoupled from their channel pendants

(present exposure), or if their channel descendant is left undefined (past exposure). These

results demonstrate that findings in marketing effectiveness research may not necessarily be

generalized from one single data set. With regard to heterogeneous channel preferences, our

four data sets show two interesting exemptions. First, the interaction between present affiliate

clicks and the number of prior channels shows a significant and positive coefficient (Data Set

1: Affiliate × PastChannelsNoAffiliate b = 0.267, p < .01). Obviously, in some cases the affiliate

seems to be fertilized by use of diverse channels. Given that affiliates may include review or

coupon websites, users who include affiliates may be more advanced in Internet channel usage

(Lambrecht and Tucker 2013). They may, therefore, use multiple channels to seek specific

deals, culminating in an affiliate visit before concluding a purchase event. Second, with regard

to the travel company, display advertising is stimulated by prior multichannel use (Data Set 4:

Display × PastChannelsNoDisplay b = 0.670, p < .01). From our previous analysis, we know

that present and cumulated past display exposure measured in isolation shows negative effects.

In contrast to other channels, a display may not exhibit a self-amplifying effect through repeated

use. Instead, display users who (actively) visit a corresponding website via multi-faceted paths

appear to be more responsive to firm-initiated media such as display marketing (Shankar and

Malthouse 2007). Remarkable is the direction of the inter-channel stimulation, as prior research

Essay 3—Users Browsing Preferences on the Path to Purchase 149

suggests that display may fuel indirect effects such as brand awareness and direct purchase

intent (e.g., Hollis 2005; Qiu and Malthouse 2009), or has shown synergetic effects, such that

display may raise search conversion and related metrics (Kireyev, Pauwels, and Gupta 2013).

While we may not refute the existence of these effects, we demonstrate counter-directional

effects, from multiple channel exposures, toward display. As both exceptional effects premise

one unique data set each, we may not generalize these findings, instead recommending an

analytical evaluation on a case-by-case foundation.

4.5.4 Category Interactions

Prior research has connected relevant attributes with channels to approximate stages within the

purchase deliberation process, and to shed light onto underlying user intentions. Hence, adding

to channel perceptions, as a next step we investigate category interaction effects (Model 2). The

first observation indicates that neither homogeneous nor heterogeneous category interactions

consistently point toward purchase events—or non-purchase events—, leaving a more

ambiguous, yet interesting picture compared to our channel-focused model (Table 27).

Contact Origin. While category homogenous interactions between present customer-initiated

channels and prior customer-initiated channels consistently show a positive effect on the time

to purchase across all data sets (e.g., Data Set 1: CIC × PastCIC b = 0.058, p < .01), the

interaction effect of the firm-initiated equivalents remain negative across data sets (e.g., Data

Set 1: FIC × PastFIC b = -0.029, p < .01). In line with preceding research, (repeated) CICs are

more effective with regard to purchase propensity than are FICs, suggesting that they are less

intrusive and more relevant (Wiesel, Pauwels, and Arts 2011). Observing the heterogeneous

category interactions, a present CIC following FICs indicate a negative tendency (Data Set 4:

CIC × PastFIC b = -0.189, p < .01). Interestingly, the counterpart interaction effects show

significant and positive effects (e.g., Data Set 4: FIC × PastCIC b = 0.110, p < .01). Considering

homogenous and heterogeneous category interactions, (multiple) CICs have a positive

Essay 3—Users Browsing Preferences on the Path to Purchase 150

connection, and (multiple) FICs have a negative connection with purchase intent. Importantly,

even if CICs are followed by a FIC, the effect remains positive, making users with CICs

especially receptive for firm-initiated or pushed marketing media (e.g., Data Set 2: FIC ×

PastCIC b = 0.051, p < .1; Data Set 3: FIC × PastCIC b = 0.041, p < .05). The opposite holds

true for FICs, such that even a subsequent CIC may be ineffective (Data Set 4: CIC × PastFIC

b = -0.189, p < .01). Consequently, switches from FICs toward CICs may not imply progression

in all instances.

Table 27

Estimation Results: Category Interactions (Table Part)—Essay 3

DS 1 (Fashion) DS 2 (Fashion) DS 3 (Luggage) DS 4 (Travel)

Base Model 2 Base Model 2 Base Model 2 Base Model 2

Variable B B B B B B B B

…

CIC × PastCIC 0.058*** 0.110*** 0.366*** 0.161***

FIC × PastFIC -0.029*** -0.010 -0.382* -0.128**

Navi × PastNavi -0.035*** -0.050*** -0.365*** -0.125***

Info × PastInfo 0.050*** 0.001 -0.054*** 0.082***

Integrated × PastIntegrated -0.003 -0.051*** 0.035 -0.145***

Separated × PastSeparated -0.035*** -0.030*** -0.154*** -0.061***

Personal × PastPersonal -0.040** -0.050 0.173*** -0.021

Nonpersonal × PastNonpersonal 0.007 -0.019 -0.417*** 0.161***

CIC × PastFIC -0.007 -0.024 -0.220 -0.189***

FIC × PastCIC -0.011 0.051* 0.041** 0.110***

Navi × PastInfo 0.050*** -0.026** -0.022 0.090***

Info × PastNavi -0.005 -0.014 -0.152*** -0.057***

Integrated × PastSeparated -0.008* -0.015* -0.057*** -0.013

Separated × PastIntegrated 0.012** -0.027*** -0.048 -0.064***

Personal × PastNonpersonal 0.006 0.048** -0.194*** 0.210***

Nonpersonal × PastPersonal 0.018*** 0.000 0.073*** -0.029

N 1,184,582 1,184,582 862,114 862,114 405,343 405,343 600,873 600,873

Observations 1,398,267 1,398,267 964,836 964,836 461,108 461,108 792,345 792,345

Time at Risk 1,863,964 1,863,964 1,306,432 1,306,432 655,963 655,963 1,171,897 1,171,897

Log Likelihood -111,550.8 -111,550.8-180,997.7 -180,910.0 -92,739.4 -92,578.7 -110,269.2 -110,131.1

AIC 223,141.6 223,042.7 362,027.3 361,883.9 185,514.9 185,225.4 220,574.4 220,330.2

BIC 223,384.6 223,480.1 362,215.8 362,260.9 185,713.6 185,600.8 220,782.9 220,724.1

R2 (D) 0.198 0.211 0.122 0.135 0.322 0.270 0.518 0.526

R2 (PH) 0.171 0.178 0.081 0.088 0.249 0.270 0.727 0.734Note: * p < .10, ** p < .05, *** p < .01; This table mirrors only a part of the full table. For a complete assay, the full table is available in the Appendix section. The base model excludes interaction effects. Model 2 includes interaction effects.

Browsing goal. Homogenous category interactions between navigational contacts and

navigational stock are significant, and are consistently negative with regard to time to purchase

(e.g., Data Set 3: Navi × PastNavi b = -0.365, p < .01). The results of their informational

Essay 3—Users Browsing Preferences on the Path to Purchase 151

pendants are more diverse, yet with a slight positive prevalence (e.g., Data Set 1: Info × PastInfo

b = 0.050, p < .01; Data Set 3: Info × PastInfo b = -0.054, p < .01; Data Set 4: Info × PastInfo

b = 0.082, p < .01). Moreover, heterogeneous category interactions between navigational

contacts and past informational stock are mixed with a slight positive overweight (e.g., Data

Set 1: Navi × PastInfo b = 0.050, p < .01) and, if significant, are consistently negative for

informational contacts and past navigational stock (Data Set 3: Info × PastNavi b = -0.152, p <

.01; Data Set 4: Info × PastNavi b = -0.057, p < .01). In summary, interactions between present

contacts with past navigational stock are negatively associated with time to purchase.

Apparently, users who primarily browse with navigational contacts are less likely to conclude

with a purchase event. The same applies for switches toward informational contacts that express

backward movement in the purchase funnel (Klapdor et al. 2015). In contrast, interactions with

past informational stock show, in many cases, positive effects—independent from the

taxonomic affiliation of the present contacts, which enhance prior findings, based on a logit

model (Klapdor et al. 2015). In other words, more navigational clicks are ineffective, yet more

informational clicks are well suited to uncover users who are prone to convert. Potentially, users

with a frank interest in purchasing tend to acquire information by means of (multiple)

informational contacts; users with a plentitude of navigational contacts forge ahead in the

purchase funnel, either purchasing immediately or not purchasing at all.

Degree of Content Integration. The results, including homogeneous interactions between

content-integrated and content-separated channels, as well as their heterogeneous complements,

show remarkable similarities across all data sets: Most coefficients are highly significant and,

apart from one exception (Data Set 1: Integrated × PastSeparated b = 0.012, p < .05), show a

negative effect on time to purchase (e.g., Data Set 2: Integrated × PastSeparated b = -0.027, p

< .01; Data Set 4: Integrated × PastSeparated b = -0.064, p < .05). Either way, in our context

the respective taxonomy seems unsuited for identifying users who are prone to convert. Haan,

Wiesel and Pauwels (2013) show that for CICs, content-integrated user-advertiser contacts are

Essay 3—Users Browsing Preferences on the Path to Purchase 152

more effective in catalyzing purchase funnel progression and sales than their content-separated

equivalents. With regard to present contacts, the channel-specific results are diverse and, to

some degree, mirror previous findings (Haan, Wiesel, and Pauwels 2013). However, they may

not apply for the interaction effects.

Degree of Personalization. Overall, the results are multifaceted. Homogeneous interactions

between present clicks and past click stock exhibit positive as well as negative effects on time

to purchase vaguely at equal proportions. Notably, for one of the fashion retailers, the

interaction of present personal contacts and past personal contacts shows a negative effect (Data

Set 1: Personal × PastPersonal b = -0.040, p < .05). Yet, looking at the luggage retailer’s data

set, this effect becomes positive (Data Set 3: Personal × PastPersonal b = 0.173, p < .01), tough,

the opposing interaction between present and past non-personal touchpoints now becomes

significant and negative (Data Set 3: NonPersonal × PastNonPersonal b = -0.417, p < .01).

Obviously, the directional manifestation of the effects seem industry (company) specific. Based

on our data sets, online retail may benefit from personal contacts during the browsing history,

travel companies, in contrast, should rather harness non-personal contacts (Data Set 4:

NonPersonal × PastNonPersonal b = 0.161, p < .01). Heterogeneous interaction effects

corroborate these findings to some degree, with the past stock element of their homogenous

pendants setting the direction and some selected more positive connotation toward purchase

propensity. For instance, interactions with past non-personal stock remain positive for travel

(Data Set 4: Personal × PastNonPersonal b = 0.021, p < .01), and show mixed, though partly

positive effects for the online retail (Data Set 2: Personal × PastNonPersonal b = 0.048, p < .05;

Data Set 3: Personal × PastNonPersonal b = -0.194, p < .01). Interactions between present non-

personal clicks and past personal contact stock show a significant and positive effect for two

online retailers and become insignificant for the travel company (Data Set 1: NonPersonal ×

PastPersonal b = 0.018, p < .01; Data Set 3: NonPersonal × PastPersonal b = 0.073, p < .01).

Collectively, it is difficult to derive valid and universal conclusions for this taxonomy from our

Essay 3—Users Browsing Preferences on the Path to Purchase 153

analyses, as the results vary vastly across data sets and effects. According to Tucker (2014) the

effectiveness of personalized advertising is a function of privacy controls. In a study focused

on retargeted banner advertisements, for instance, personalized display advertisements are

found less purchase effective as generic retargeted display advertisements (Lambrecht and

Tucker 2013). Adding to this, our research reveals that industry-specific features seem to apply,

thus, data sets should be analyzed individually to conclude on the effectiveness of personalized

respectively non-personalized advertisements.

4.6 Discussion

In this study, we developed a model to analyze online journey data aiming to investigate on

users’ channel preferences in a multichannel online environment with a focus on channel and

channel category interactions. Thereby we add to prior research on online channel effectiveness

in a single channel setting (e.g., Ghose and Yang 2008, 2010; Manchanda, Dubé, Goh, and P.

Chintagunta 2006; Rutz, Bucklin, and Sonnier 2012; Rutz, Trusov, and Bucklin 2011), on

channel interplay in a two-dimensional setting (e.g., Fulgoni and Mörn 2009; Kireyev, Pauwels,

and Gupta 2013; Rutz and Bucklin 2011; Yang and Ghose 2010), and on channel taxonomies

defining channel groups or investigating their role in the purchase deliberation process (e.g.,

Broder 2002; Haan, Wiesel, and Pauwels 2013; Li and Kannan 2014; Rose and Levinson 2004).

Furthermore, our research also links back to the progression (regression) in the choice set

formation process in the purchase funnel (Hauser and Wernerfelt 1990; Shocker et al. 1991).

Based on four single-sourced individual-level data sets from three different industries, we

implement a wide range of channels and taxonomies into two primary models allowing for

analyzing their role in predicting user conversions more holistically. The models’ estimation

results reveal valuable insights into user (homogeneous and heterogeneous) channel and

category preferences along their path to purchase, contribute to online marketing effectiveness

research, and link to the theory of choice sets paralleling the purchase funnel in multifaceted

ways.

Essay 3—Users Browsing Preferences on the Path to Purchase 154

First, we contribute novel knowledge on online marketing effectiveness and user

channel preferences and browsing behavior in multichannel settings (Agichtein et al. 2006;

Fulgoni and Mörn 2009; Klapdor et al. 2015). We link user channel preferences and purchase

decision-making by analyzing present channel preferences, past channel stock and, foremost,

by introducing interactions among present and past channel exposures. Importantly, we apply

these interactions not only between present channel exposure and their exposures on the past

days, but include also past exposures on the same day. Amending to former multichannel

research, we reveal that, in a pure online environment, users commonly apply preferred

channels on their path to purchase. These results indicate homogenous/idiosyncratic channel

preferences, rather than heterogeneous/multichannel preferences (Edell and Keller 1989; Jagpal

1981; Naik and Raman 2003; Tellis et al. 2005), and extend established knowledge on

synergetic effects between online channels (Kireyev, Pauwels, and Gupta 2013; Li and Kannan

2014). Across all data sets and, thus, industries, multiple homogeneous channel clicks are

associated with an increase of purchase propensity and indicate progression in the user’s path

to purchase. Monitoring and anticipating homogeneous and heterogeneous browsing routes can

support advertisers to identify more valuable users.

Second, we generate novel insights into research on category approaches in a

multichannel setting by implementing all well-accepted and relevant channel taxonomies into

one model, jointly analyzing homogeneous and heterogeneous channel group interactions

within each taxonomy (Broder 2002; Haan, Wiesel, and Pauwels 2013; Rose and Levinson

2004). Thereby we combine several existing theoretical approaches and extend research on

channel effectiveness (Yadav and Pavlou 2014). The results propose a more diverse picture

concerning data sets, industries, and directions with regard to the user’s progression (regression)

in the purchase funnel. For instance, turning to the taxonomy of contact origin, homogeneous

interactions between present and past customer-initiated channels reveal consistently positive,

their firm-initiated equivalents consistently negative effects on time to purchase, endorsing

Essay 3—Users Browsing Preferences on the Path to Purchase 155

prior findings (Haan, Wiesel, and Pauwels 2013; Wiesel, Pauwels, and Arts 2011). Looking at

the browsing goal classification, our results confirm existing research and shed light into novel,

counterintuitive phenomena. Expectedly, switches toward navigational contacts are positively

associated with time to purchase—their informational counterparts negatively (Jansen, Booth,

and Spink 2008; Klapdor et al. 2015). Adding more detail, we find that, interactions with past

navigational stock indicate negative effects, while interactions with past informational stock, in

most cases, indicate positive effects on purchase propensity. Remarkably these results are

independent from the taxonomic nature of the present contacts. Regarding the degree of

personalization, surprisingly, past personal stock is especially effective for online retailers,

however, negative for travel. Instead, travel rather benefits from non-personal prior contacts.

These advancements in research on channel taxonomies generate meaningful insights into user

preferences represented by interaction effects along the user journey, responding to the research

requests to introduce marketing impact models building on individual-level customer path data

(Hui, Fader, and Bradlow 2009; Rust, Lemon, and Zeithaml 2004) that bridge the gap between

theory and practical relevance (Little 2004a; b). By encompassing all channels and established

channel taxonomies we further corresponds to the claim that channels should not be analyzed

in isolation which may induce ineffective conclusions (Li and Kannan 2014).

Third, in a more comparative setting characterized by a complete set of channels (also

considering various industries), we are the first to interpret and translate individual channel

clicks into purchase propensity, adding to preceded research on channel effectiveness in less

comprehensive environments (Braun and Moe 2013; Goldfarb and Tucker 2011a; Rutz and

Bucklin 2011) or in consolidated settings based on channel taxonomies (Haan, Wiesel, and

Pauwels 2013; Jansen, Booth, and Spink 2008). Our results suggest that users more prone to

convert utilize customer-initiated contacts including search channels or direct type-in rather

than firm-initiated contacts, confirming prior findings, yet on an individual channel level (Haan,

Wiesel, and Pauwels 2013). Due to the analysis of individual channels, we can show that

Essay 3—Users Browsing Preferences on the Path to Purchase 156

branded customer-initiated channels such as branded search well-reflect users prone to convert,

which in line with previous research on channel taxonomies (Anderl, Schumann, and Kunz

2015). Setting channel effectiveness in a comprehensive and competitive context, our findings

support and extent prior research on channel effectiveness and category approaches.

Fourth, our research links back to the theory of choice set formation paralleling the

users’ purchase funnel, as we accordingly comprise channel and category exposure and its

effects on purchase propensity along the full journey (Campbell 1969; Hauser and Wernerfelt

1990; Howard and Sheth 1969; Howard 1963; Roberts and Lattin 1991, 1997; Wright and

Barbour 1977). Albeit field data may not fully unveil the users’ set affiliations (Shocker et al.

1991), the estimation results, foremost the interaction effects, exhibit purchase propensity, and

purchase reluctance, and, thus, may be interpreted as progression, stagnation or regression in

the purchase funnel and linked back to the choice set formation process. Idiosyncratic channel

customer-advertiser interactions therefore may serve as an agent for progression in a purchase

decision process—multifaceted channel exposure for regression accordingly.

Fifth, our research also links theoretical approaches on marketing effectiveness and

clickstream based research to practice, a perpetually lament challenge in marketing research

(Little 1970, 1979; Lodish 2001) and repeatedly claimed by scholars (Yadav and Pavlou 2014).

It is also worthwhile to consider the richness of data and the research scope in relation to prior

literature on online marketing effectiveness. With the exception of a few recent studies (Klapdor

et al. 2015; Li and Kannan 2014), research embracing the full availability of online channels

and channel taxonomies are exceptionally scarce. Studies further implementing multiple

industries are almost non-existent. Opposed to prior research (e.g., Abhishek, Fader, and

Hosanagar 2012; Breuer, Brettel, and Engelen 2011; Kireyev, Pauwels, and Gupta 2013; Xu,

Duan, and Whinston 2014), we are fortunate to build on four real-world data sets that comprise

the complete set of channels the corresponding advertiser applies, enabling us to shed light into

Essay 3—Users Browsing Preferences on the Path to Purchase 157

multifaceted channel and channel taxonomy research and allowing for decollating practical

generalizations and industry-specifics.

Moreover, our findings demonstrate valuable implications for practitioners and supports

advertisers in shaping their online marketing activities. Based on each individual user’s

browsing history, marketers may leverage our methodology and results to develop rule sets for

targeting, retargeting, and hence real-time bidding. Notably, our insights are not only relevant

to target already known customers, but also apply to unknown users recorded by the advertiser

while browsing the web. Adequate retargeting and budget allocation toward well-selected

channels have demonstrated to improve marketing effectiveness and marketing ROI

(Lambrecht and Tucker 2013; Tucker 2012). Following their argumentation, advertisers

anticipating our findings may benefit in a related manner. Except for display marketing, users

with idiosyncratic click preferences in their browsing history appear more promising for

targeted creatives of the corresponding channel. Display marketing, on the contrary, should be

broadcasted to users with more diverse channel exposure on their browsing path. Regarding

category exposure, click stock of informational nature (browsing goal taxonomy) indicate users

progressing in their purchase process. Likewise, our results suggest that click stock of CICs

(even anteceding present FICs) indicate a higher purchase inclination, making users with

multiple CICs especially accessible via pushed marketing media. Beyond these exemplary

illustrations, our results offer a myriad of insights to improve the effectiveness of online

marketing measures.

4.7 Limitations and Research Directions

Given that our models are estimated on data collected using cookie tracking technology, some

shortcoming may not be completely ruled out including cookie deletion and the usage of the

same device by multiple users, or the usage of multiple devices by one single user (Flosi,

Fulgoni, and Vollman 2013). Though cookie tracking continues to be industry standard (Tucker

Essay 3—Users Browsing Preferences on the Path to Purchase 158

2012), novel tracking technologies such as digital finger prints may alleviate these data related

concerns. Furthermore, our data sets comprise clickstreams omitting impressions such that we

do not track users that have solely viewed an online advertisements. Although display click

rates decreased from 2% in 1995 (Cho and Cheon 2004) to no more than 0.1% about 15 years

later (Chapman 2011; Fulgoni and Mörn 2009), the mere display impression has proven indirect

effects on users such as changes in brand awareness, brand attitudes and purchase intentions as

function of ad exposure (Briggs and Hollis 1997; Cho, Lee, and Tharp 2001; Drèze and

Hussherr 2003). It is conceivable that, to some degree, equivalent effects may apply for other

online channels that are also subject for mere impressions, for instance, email or search engine

marketing. Furthermore, omitting impression data may entail some pre-selection bias such that

certain users may avoid clicking on (particular) channels and, in consequence, never appear in

our data records. To decrease these effects, marketing research may involve impressions as well

as clicks and apply, for instance, a two-staged model. However, impression tracking may not

suffice, as the technically tracked impression does not warrant users actually seeing and reading

the creative’s content. Hereby, controlled experiments may elucidate these effects, however,

may involve other shortcomings.

Moreover, the use of real-world field data measures the pure customer-advertiser

interactions. The users’ motivation and (potential) psychological mechanisms remain masked.

Thus, the data enables us to measure correlation and to approximate explanations, which,

however, is incongruent with causality. Therefore, we focus our interpretation on the purchase

inclination associated with channel and channel category exposure. A substantial field

experiment, ideally combined with a survey based study, may diminish these concerns opening

an avenue for future research.

Regarding data processing, some channel-category classifications are controversial. For

instance, retargeted display advertisements are classified as informational in the browsing goal

Essay 3—Users Browsing Preferences on the Path to Purchase 159

taxonomy, though, some users may be well-aware of the particular offering displayed on the

creative and, thus, their reaction may rather express the navigational nature of their browsing

state. Another noteworthy example in this context is email marketing, as it may be personalized

and non-personalized. As the advertisers in our data sets purely emit non-personalized email

newsletters, we assigned it accordingly, however, in future studies a clear advertiser-specific

differentiation may apply. Along these lines, a bulletproof affiliation of channels in general is

complex, as channel use may differ and the user’s underlying intention is subject for estimation

if not inquired.

Moreover, the implementation of the above mentioned channels may cause activity bias

due to potential pre-selection. Retargeted displays are by definition placed to users who have

already exhibited at least one website visit. Email advertisements are only dispatched to users

that have already found their way into the advertiser’s customer list, potentially by a newsletter

sign-up or due to prior purchase orders. As not all employed data sets convey the affected

channels (e.g., Data Set 2) and leading to stable results at the same time, we may partly alleviate

these concerns. Further, we performed our analyses across all data excluding these channels—

again with comparable results.

We aimed to shed light into channel and channel category preferences along the full

customer journey with a clear focus on corresponding homogenous and heterogeneous

browsing preferences. However, it would be very interesting to analyze the interplay of all

applied online channels among one another comprehensively in one model. Given that our data

sets comprise between eight and ten online channels, the possible number combinations of

interactions would raise dramatically (Bellman 1961). Albeit the size of the data sets applied,

the number of the consequential variables may not be handled adequately with the model

chosen. Finally, it is a trade-off between the data set size and the number of channels considered.

That said, a deeper understanding on the interplay of channels in multichannel environments

Essay 3—Users Browsing Preferences on the Path to Purchase 160

open manifold interesting opportunities for further research. In a broader context, one related

and distinct research topic may also include the effectiveness of the frequency of channel

exposure—also referred to as frequency capping. Specifying the maximum number of

exposures (by channel) from which the positive advertising effects turns to a negative effect is

essential for pushed marketing media, such as retargeted display marketing.

Conclusion 163

5 Conclusion

This dissertation examines three relevant and current challenges in the fields of online

marketing effectiveness, customer online behavior and online multichannel marketing based on

clickstream data. Essay 1 proposes an innovative methodology of attribution models

contributing novel insights into online channel effectiveness and channel spillover and

carryover effects. Driven by flow theory, Essay 2 introduces a novel, time-related concept,

namely the micro-journey, that supports in better explaining customer purchasing behavior and

enhances forecasting customer conversions. Essay 3 analyzes the customers’ present and past

channel preferences on their path to purchase, focusing on homogeneous and heterogeneous

channel and channel category click browsing histories and the underlying purchase intent.

5.1 Implications

In Essay 1, we present an attribution framework based on first- and higher-order Markov walks

to examine the contribution of individual online channels and to generate novel insights into

carryover and spillover effects between online channels in a multichannel environment.

Utilizing four data sets from three different industries, and comparing our results to existing

heuristics (e.g., one-click heuristics), our study is the first to delineate cross-industry

generalizations from industry-specific findings (Abhishek, Fader, and Hosanagar 2012; Li and

Kannan 2014). Thereby, our results expand multichannel research on attribution modeling (e.g.,

Li and Kannan 2014; Xu, Duan, and Whinston 2014), research on online channel efficiency

(Neslin and Shankar 2009), and provide valuable directions for online marketers on channel

contributions, channel budgets allocations, and calibrate targeting strategies.

We bring forward generalizable insights into the effectiveness of individual online

channels in a multichannel setting, by applying our graph-based framework on four empirical

data sets and comparing the results to two prominent attribution heuristics (first- and last-click

wins) and to two logit models, adding to extant attribution literature (Abhishek, Fader, and

Conclusion 164

Hosanagar 2012; Li and Kannan 2014; Xu, Duan, and Whinston 2014). Our findings suggest

that firm-initiated channels are consistently undervalued, whereas direct type-ins and SEA, both

customer-initiated channels, are overestimated by the heuristic approaches across industries.

For other customer-initiated channels, especially SEO, the results are industry-specific, thus,

we conjecture that additional factors such as brand characteristics seem to affect customers’

click behavior (Jerath, Ma, and Park 2014).

Furthermore, the higher-order models shed light into the interplay of channels (i.e.,

carryover and spillover effects). Across data sets, our results consistently ascribe a substantial

contribution to homogeneous channel sequences across data sets which may reflect extant

channel preferences for some users (Godfrey, Seiders, and Voss 2011; Li and Kannan 2014).

With regard to the channels’ contact origin, we observe spillover effects within, as well as

between, channel groups. Customer-initiated channels show substantial removal effects if they

are followed by other customer-initiated channels, whereas spillover effects between firm-

initiated channels are by and large negligible (Anderl, Schumann, and Kunz 2015).31

From a modeling perspective, our framework constitutes a novel, objective, versatile

and efficient alternative to existing attribution modeling techniques (Abhishek, Fader, and

Hosanagar 2012; Berman 2015; Haan, Wiesel, and Pauwels 2013; Kireyev, Pauwels, and Gupta

2013; Li and Kannan 2014; Xu, Duan, and Whinston 2014) by representing multichannel online

customer path data as Markov walks in directed graphs. While first- and higher-order models

measure channel contributions in multichannel settings, higher-order models are particularly

well-suited to compute the value contribution of channel sequences. Implemented into practice

31 In Essay 3, we deeper investigate homogeneous and heterogeneous channel (and channel category) browsing

preferences.

Conclusion 165

and supporting previous findings (Abhishek, Fader, and Hosanagar 2012; Berman 2015; Li and

Kannan 2014), our framework has proven high practical and theoretical value.

Essay 2 encompasses a novel modeling approach to studying click-time related

browsing pattern. Building on four large clickstream data sets, we conceptualize the micro-

journey as relevant browsing behavior and document its impact on customer conversions,

further including important characteristics of the micro-journey that are connected to and

established in online marketing literature. Thereby, our findings contribute to the ongoing

debate on which clicking actions within the consumer’s online browsing history are important

in anticipating purchase intent (Chatterjee, Hoffman, and Novak 2003; Hui, Fader, and Bradlow

2009).

Routing from flow theory, we conceptualize “concentration/attention focus”, in an

online context (Novak, Hoffman, and Yung 2000), as potential user browsing state by

anticipating intense browsing sessions measured by single clicks that occur in succession within

short time intervals. We are the first to model the time between successive customer-advertiser

interactions (clicks) to form the concept of the micro-journey that, in consequence, helps to

uncovers hidden purchase intentions of consumers. Our results reveal that implementing the

micro-journey improves the models’ predictive capabilities and supports in identifying users

more prone to convert. In consequence, they indicate that underlying browsing motives are

linked to consumer’s online actions mirrored in manifold marks in their clickstream logs. The

micro-journey may serve as a starting point to simplify and structure these traces. Relying on

extant research (e.g., Broder 2002; Jansen, Booth, and Spink 2008; Klapdor et al. 2015; Wiesel,

Pauwels, and Arts 2011), we transfer effective properties into the context of micro-journeys,

thereby connecting our novel concept to prior findings such as insights from categorization

approaches (taxonomy of the browsing goal). Users with micro-journeys in their browsing

history are more prone to convert—they convert directly following the micro-journey, or, in

Conclusion 166

virtually equal proportions, a number of clicks thereafter. Moreover, the relative position of the

micro-journey within the overall journey, as well as category switches, importantly, within the

micro-journey, (informational to navigational contacts—and vice versa), are especially suitable

to predict converting customers (Klapdor et al. 2015). Adapting the micro-journey on four

large-scale individual user-level data sets allows to generalize the concept’s effectiveness and

further reveals systemic differences of its characteristics among industries (Li and Kannan

2014). For instance, the effect of the micro-journey position (within the journey) on later

purchases may be moderated by industry-specific influences, such as price sensitivity.

From a more methodological perspective, this study links psychological and clickstream

approaches to understand costumer (purchase) motivation. As macro-level collective outcomes

(e.g., conversion rates) are subject to micro-level behavior and decision making of individuals

(e.g., click behavior), it is advisable to consider the most disaggregated input unit possible to

approximate capturing the underlying individual psychological processes (Tellis and Franses

2006). Building on one of the disaggregated input units possible (individual-user level), the

concept of the micro-journey, serves as a proxy for purchase motivation adding to research on

(multichannel) marketing effectiveness focused on other concepts, and mostly building on

aggregated data (e.g., Breuer, Brettel, and Engelen 2011).

In Essay 3, we investigate the users channel preferences (homogeneous or

heterogeneous) on their path to purchase. Implementing the full set of online channels and

several, relevant channel taxonomies allow for analyzing the role of past and present channel

clicks, as well as the interplay between online channels and channel groups in predicting user

conversions. Our results reveal novel insights into user channel and category preferences along

their path to purchase, contribute to online marketing effectiveness research (Fulgoni and Mörn

2009; Manchanda et al. 2006), multichannel research (Haan, Wiesel, and Pauwels 2013;

Conclusion 167

Leeflang et al. 2014), and link to the theory of choice sets paralleling the purchase funnel in

multifaceted ways (e.g., Shocker et al. 1991).

We link user channel preferences and purchase decision-making by evaluating present

channel exposure, past channel stock and, furthermore, by modeling their interactions. Our

results reveal that, in pure online environment, users consistently show idiosyncratic channel

preferences, adding to prior knowledge on synergetic effects between online channels (Fulgoni

and Mörn 2009; Klapdor et al. 2015), to literature on online (and offline) channel synergies

(Edell and Keller 1989; Jagpal 1981; Naik and Raman 2003; Tellis et al. 2005), as well as add

a novel channel constellation that indicates progression in the choice set formation process

(Hauser and Wernerfelt 1990; Shocker et al. 1991).

Likewise, on categorization approaches, we derive novel insights by conjoining well-

established taxonomies and their interactions (Broder 2002; Broder et al. 2007; Haan, Wiesel,

and Pauwels 2013; Rose and Levinson 2004). Combining several existing theoretical

approaches, our results infer a more diverse picture across data sets (i.e., industries), and

provide directions with regard to the user’s progression (regression) extend research on channel

effectiveness (Shocker et al. 1991; Yadav and Pavlou 2014). For instance, independent from

the taxonomic nature of the present contact, we conjecture that interactions with more extensive

past navigational stock indicate negative effects, and, in contrast, interactions with past

informational stock indicate positive effects on purchase behavior. These advancements on

channel category preferences generate meaningful insights dissolved in the users’ browsing log,

contributing to current research requests to further investigate individual-level customer path

data (Hui, Fader, and Bradlow 2009; Rust, Lemon, and Zeithaml 2004) and to bridge the gap

between theory and practice (Little 2004a; b).

From a broader perspective, we developed marketing impact models to translate online

multichannel path data into purchase inclination. Thereby we expand prior research on online

Conclusion 168

effectiveness in single channel settings (e.g., Ghose and Yang 2008, 2010; Manchanda, Dubé,

Goh, and P. Chintagunta 2006; Rutz, Bucklin, and Sonnier 2012; Rutz, Trusov, and Bucklin

2011), on channel interrelations in a two-dimensional settings (e.g., Fulgoni and Mörn 2009;

Kireyev, Pauwels, and Gupta 2013; Rutz and Bucklin 2011; Yang and Ghose 2010), and on

channel category approaches investigating the role of channel groups in the purchase

deliberation process (Broder 2002; Campbell 1969; Haan, Wiesel, and Pauwels 2013; Hauser

and Wernerfelt 1990; Howard and Sheth 1969; Howard 1963; Li and Kannan 2014; Roberts

and Lattin 1991, 1997; Rose and Levinson 2004; Wright and Barbour 1977). In opposition to

prior multichannel research (e.g., Abhishek, Fader, and Hosanagar 2012; Breuer, Brettel, and

Engelen 2011; Kireyev, Pauwels, and Gupta 2013; Xu, Duan, and Whinston 2014), we are

fortunate to build on four empirical data sets, enabling us to move toward multifaceted channel

and channel taxonomy research and to differentiate industry-specific findings from cross-

industry generalizations (Li and Kannan 2014). Our findings support the claim that online

channels should not be analyzed in isolation, which may encourage ineffective conclusions, and

correspond to the request to deeper examine cross-industry effects (Li and Kannan 2014).

Moreover, our research demonstrates valuable implications on various explicit problems

that online marketers confront. Our findings can support in selecting and recalibrating the

allocation of marketing budget toward individual channels. Applying our attribution framework

(Essay 1), helps to contrast online channel budgets and virtual channel contributions and may

facilitate adaptions toward optimal budget allocation (Raman et al. 2012). Identifying the users’

channel preferences (Essay 3), collectively and individually, may further complement online

channel strategies and optimize budget decisions. Moreover, our results help to better predict

conversions, conversion timing, and to unveil individual customer more prone to convert. That

said, they may serve as valid input source to sharpen rule sets in state-of-art applications such

as real-time bidding auctions and targeting applications. For instance, micro-journeys detected

in the users’ browsing paths and anticipating their increased purchase propensity (direct and

Conclusion 169

later) may act as an indicator for targeting selected users (Essay 2). Thoughtful targeting

activities and data-driven attribution techniques have proven to be advantageous raise

marketing effectiveness and marketing ROI (Lambrecht and Tucker 2013; Tucker 2012), thus,

advertisers incorporating our findings may optimize their marketing activities. Importantly, our

findings are not limited to already known users—they also apply to unknown users tracked in

the advertisers’ browsing logs

From a practitioners’ perspective, our research helps to bridge the gap between

marketing theory and practice. Being a perpetually lamented challenge in marketing research

(Little 1970, 1979), researchers to date continue to recall for closing the still widening gap

between the interests and priorities of academic marketers and the needs of marketing

executives challenged by uncertain, complex and dynamic marketing environments (Jaworski

2011; Lehmann, Mcalister, and Staelin 2011; Lilien 2011; Little 2004b; Reibstein, Day, and

Wind 2009). Yet, existing marketing research does not fully capture the increasing richness and

complexity of firms' online marketing activities (Yadav and Pavlou 2014). Our research is based

on flexible and interpretable frameworks generating results that go beyond theoretical

relevance, as they offer a plentitude of insights to improve the effectiveness of online marketing

measures. For instance, we implemented our attribution framework as a real-life system at a

German multichannel tracking provider in a real industry environment (Essay 1). Thus far,

several clients, operating in the fashion, sports equipment, and telecommunications industry,

have applied our attribution model, serving as proof for its high usability and positive impact

on marketing effectiveness.

5.2 Outlook

The studies in this dissertation have several limitations and open promising avenues for future

research, which we identify in the following chapter. Although we utilized four large-scale,

Conclusion 170

real-world data sets that include online user click behavior on highly detailed level, some

limitations may relate to these data sets.

The customer journeys in these data sets were relatively short on average which is also

indicated by a substantial share of one-click journeys. However, advanced attribution

techniques (Essay 1), as well as click patterns as micro-journeys (Essay 2), or as interaction

effects (Essay 3), are especially relevant for journeys consisting of more than one click. Given

the real-world nature of our data sets, and the aspiration to reflect the advertisers’ marketspace

as realistic as possible, the consideration of the full data sets including, of course, one-click

journeys, represents the most adequate approach. Nonetheless, we also controlled for reduced

samples across all studies, which manifested our results. For instance, the deviance between

our graph-based attribution model and the two heuristic approaches, namely first click and last

click wins, was more pronounced with reduced samples (Essay 1). In Essay 2, we control for

the number of journey clicks and, omitting one-click journeys, the magnitude of the micro-

journey’s effect mirrored in the corresponding coefficients became even stronger. Still, it is

noteworthy that all studies are rather designed for multi-click customer data.

Furthermore, our data sets include exact click information, yet omitting view exposures,

which may affect the results of certain online channels to a different degree. That said, display

marketing clearly emphasizes graphical elements and is (mostly) initiated by the firm, rather

than by the user (retargeted display may be interpreted as firm- and partly customer-initiated

channel). While click rates on display advertisements have dropped to less than 0.1% (Chapman

2011; Fulgoni and Mörn 2009), they have proven brand-related effects and indirect effects on

purchase intent (Briggs and Hollis 1997; Cho, Lee, and Tharp 2001; Drèze and Hussherr 2003).

Although not being thoroughly analyzed so far, similar effects may still be valid for other online

channels such as newsletter marketing or search results. Additionally, excluding views may

imply preselecting users more prone to click (on certain channels) and ignore users who refuse

Conclusion 171

clicking and therefore never appear in our data records. To rule out these effects, the

consideration of view data is advisable, for instance, by applying two-staged models. Also the

analysis of the indirect effects of channel exposure on brand awareness and purchase

inclination—apart from display marketing—remains largely uncovered, and would be an

interesting task for future research. All our approaches would be well-suited to handle views as

well as clicks, however, we were not able to obtain customer path data including clicks. The

tracking of impression data also suffers from the uncertainty whether a user actually notices the

advertisement and reads the creatives’ message. Controlled field experiments may be a useful

tool to shed light into these effects.

Given that our data sets are collected using cookie tracking techniques, some

assumptions may be inaccurate. Cookies may be deleted by the user, which may cause a break

of this particular user’s journey. Moreover, the usage of the same device by multiple users, or

the usage of multiple devices by one single user, is not reflected. Albeit cookie tracking remains

industry standard (Tucker 2012), novel tracking techniques such as digital finger prints may

alleviate this matter. In addition, our data does not include information of the device applied

(e.g., desktop, laptop, tablet, or mobile phone). Analyzing if and how the device affects the

impact of marketing messages, seems to be a fruitful avenue for further research, as utilization

and habits may differ between devices. While mobile devices are most likely held by one person

only, it is less obvious how many individuals access a desktop computer. In the same vein,

desktops or laptops may be, for convenience reasons (e.g., display size), the first choice to

perform more complex tasks such as online shopping.

Besides limitations associated with cookie tracking technologies, we may not observe

journeys at competitors’ web shops. Due to regulation, we may not include cross-website

conversions. Therefore, journeys marked as non-converting in our data sets, may factually still

result in a conversion event, however, at an untraced online web shop or at offline retailers or

Conclusion 172

travel agencies (Wiesel, Pauwels, and Arts 2011). From the perspective of the analyzed web

shop, these conversions still may be interpreted as non-converting, since they do not generate

revenues for the corresponding advertiser. From a broader perspective, these users may be seen

as lost, as they exhibit real purchase intent. Future work may focus on the users’ activities more

holistically covering the web or integrating online and offline purchase paths. The latter, of

course, raises novel challenges in tracking technologies.

Moreover, to evaluate the effectiveness of online marketing channels, companies may

also consider revenues and profits from conversions, and, potentially, the customer lifetime

value (CLV). In a study, Chan, Wu, and Xie (2011) show that the CLV differs by the online

marketing channels though which a customer is acquired. Our graph-based approach is well

suited for these monetary extensions, however, will require additional data (Essay 1). With

regard to Essay 2 and Essay 3, the models must be adapted to handle, for instance, such

continuous dependent variables—which, in general, seems feasible.

Finally, real-world clickstream data purely measures the customer-advertiser

interactions (i.e., correlations between marketing measures and conversions), evoking the

perpetual debate on correlation and causality, in an online context (Luo 2009). As the users’

real thoughts, underlying motivation, and psychological mechanisms remain hidden, a strict

causal interpretation of customer journey data is inappropriate. For some measured correlations

alternate explanations may apply including activity bias (Lewis, Rao, and Reiley 2011), pre-

selection of users expected to be more prone to convert via targeting (Lambrecht and Tucker

2013), or addressing of the established customer base via newsletter (Tezinde, Smith, and

Murphy 2002). Although we ran the analyses excluding the above mentioned channels leading

to similar results, we may not exhaustively alleviate all underlying mechanisms. Elucidating

the causal relationship between browsing (i.e., clicking) behavior and online purchase decision

making would require substantial field experiments with randomized exposure, ideally

Conclusion 173

paralleled by a survey based study. Such experiments are complex and costly to implement in

practice, especially in a multichannel setting as certain channels (e.g., SEO or direct type-in)

are beyond the control of the experimental setup. Nonetheless, such experimental studies would

be a valuable extension to online marketing research.

References 175

6 References

Abhishek, Vibhanshu, Peter S. Fader, and Kartik Hosanagar (2012), “Media Exposure

through the Funnel: A Model of Multi-Stage Attribution,” Working Paper, (accessed

October 5, 2013), [available at http://ssrn.com/abstract=2158421].

——— and Kartik Hosanagar (2012), “Optimal Bidding in Multi-Item Multi-Slot Sponsored

Search Auctions,” in Proceedings of the 13th ACM Conference on Electronic Commerce

- EC ’12.

Agarwal, Ashish, Kartik Hosanagar, and Michael D. Smith (2011), “Location, Location,

Location: An Analysis of Profitability of Position in Online Advertising Markets,”

Journal of Marketing Research, 48 (6), 1057–73.

Agichtein, Eugene, Eric Brill, Susan Dumais, and Robert Ragno (2006), “Learning User

Interaction Models for Predicting Web Search Result Preferences,” in Proceedings of the

29th Annual International ACM SIGIR Conference on Research and Development in

Information Retrieval - SIGIR ’06, New York, New York, USA: ACM.

Akaike, Hirotugu (1974), “A New Look at the Statistical Model Identification,” IEEE

Transactionson Automatic Control, 19 (6), 716–23.

Alba, Joseph W. and Amitava Chattopadhyay (1985), “Effects of Context and Part-Category

Cues on Recall of Competing Brands,” Journal of Marketing Research, 22 (3), 340–49.

Anderl, Eva, Jan Hendrik Schumann, and Werner Kunz (2015), “Helping Firms Reduce

Complexity in Multichannel Online Data: A New Taxonomy-Based Approach for

Customer Journeys,” Journal of Retailing, forthcoming.

Animesh, Animesh, Siva Viswanathan, and Ritu Agarwal (2011), “Competing ‘Creatively’ in

Sponsored Search Markets: The Effect of Rank, Differentiation Strategy, and

Competition on Performance,” Information Systems Research, 22 (1), 153–69.

References 176

Ansari, Asim and Carl F. Mela (2003), “E-Customization,” Journal of Marketing Research,

40 (2), 131–45.

Archak, Nikolay, Vahab S. Mirrokni, and S. Muthukrishnan (2010), “Mining Advertiser-

Specific User Behavior Using Adfactors,” in Proceedings of the 19th International

Conference on World Wide Web, ACM, 31–40.

Baye, Michael R., J. Rupert J. Gatti, Paul Kattuman, and John Morgan (2009), “Clicks,

Discontinuities, and Firm Demand Online,” Journal of Economics and Management

Strategy.

Beales, Howard (2011), “The Value of Behavioral Targeting,” Research study. Network

Advertising Initiative.

Beatty, Sharon E. and Scott M. Smith (1987), “External Search Effort: An Investigation

Across Several Product Categories,” Journal of Consumer Research, 14 (1), 83–95.

Bellman, Richard E. (1961), Adaptive Control Processes: A Guided Tour, Princeton, NJ:

Princeton University Press.

Berendt, Bettina, Bamshad Mobasher, Myra Spiliopoulou, and Jim Wiltshire (2001),

“Measuring the Accuracy of Sessionizers for Web Usage Analysis,” in Proceedings of

the Workshop on Web Mining, First SIAM International Conference on Data Mining,

Chicago (IL), 7–14.

Berman, Ron (2015), “Beyond the Last Touch: Attribution in Online Advertising,” Working

Paper, (accessed November 24, 2015), [available at http://ron-

berman.com/papers/attribution.pdf].

Berthon, Pierre, Leyland F. Pitt, and Richard T. Watson (1996), “The World Wide Web as an

Advertising Medium: Toward an Understanding of Conversion Efficiency,” Journal of

Advertising Research, 36 (1), 43–54.

References 177

Birnbaum, Michael H. (1999), “Testing Critical Properties of Decision Making on the

Internet,” Psychological Science, 10 (September), 399–407.

Blackwell, Roger D., Paul W. Miniard, and James F. Engel (2005), Consumer Behavior,

Madison: Thomson/South-Western (10th edition).

Bonfrer, André and Xavier Drèze (2009), “Real-Time Evaluation of E-mail Campaign

Performance,” Marketing Science, 28 (2), 251–63.

Bosnjak, Michael, Mirta Galesic, and Tracy Tuten (2007), “Personality Determinants of

Online Shopping: Explaining Online Purchase Intentions Using a Hierarchical

Approach,” Journal of Business Research, 60, 597–605.

Bowman, Douglas and Das Narayandas (2001), “Managing Customer-Initiated Contacts with

Manufacturers: The Impact on Share of Category Requirements and Word-of-Mouth

Behavior,” Journal of Marketing Research, 38 (3), 281–97.

Bradley, Andrew P. (1997), “The Use of the Area Under the ROC Curve in the Evaluation of

Machine Learning Algorithms,” Pattern Recognition, 30 (7), 1145–59.

Braun, Michael and Wendy W. Moe (2013), “Online Display Advertising: Modeling the

Effects of Multiple Creatives and Individual Impression Histories,” Marketing Science,

32 (5), 753–67.

Breuer, Ralph and Malte Brettel (2012), “Short- and Long-Term Effects of Online

Advertising: Differences Between New and Existing Customers,” Journal of Interactive

Marketing, 26 (3), 155–66.

———, ———, and Andreas Engelen (2011), “Incorporating Long-Term Effects in

Determining the Effectiveness of Different Types of Online Advertising,” Marketing

Letters, 22 (4), 327–40.

References 178

Briggs, Rex and Nigel Hollis (1997), “Advertising on the Web: Is There Response Before

Click-Through?,” Journal of Advertising Research, 37 (2), 33–45.

Broder, Andrei (2002), “A Taxonomy of Web Search,” ACM SIGIR Forum, 36 (2), 3–10.

———, Marcus Fontoura, Vanja Josifovski, and Lance Riedel (2007), “A Semantic Approach

to Contextual Advertising,” in Proceedings of the 30th Annual International ACM SIGIR

Conference on Research and Development in Information Retrieval, ACM Press.

Bronnenberg, Bart J. (1998), “Advertising frequency decisions in a discrete Markov process

under a budget constraint,” Journal of Marketing Research, 35 (3), 399–406.

Brown, Juanita J. and Albert R. Wildt (1992), “Consideration Set Measurement,” Journal of

the Academy of Marketing Science, 20 (3), 235–43.

Bucklin, Randolph E., James M. Lattin, Asim Ansari, David Bell, Eloise Coupey, Sunil

Gupta, John D. C. Little, Carl Mela, Alan Montgomery, and Joel Steckel (2002), “Choice

and the Internet: From Clickstream to Research Stream,” Marketing Letters, 13 (3), 245–

58.

——— and Catarina Sismeiro (2003), “A Model of Web Site Browsing Behavior Estimated

on Clickstream Data,” Journal of Marketing Research, 40 (3), 249–67.

——— and ——— (2009), “Click Here for Internet Insight: Advances in Clickstream Data

Analysis in Marketing,” Journal of Interactive Marketing, 23 (1), 35–48.

Buis, Maarten L. (2011), “The Consequences of Unobserved Heterogeneity in a Sequential

Logit Model,” Research in Social Stratification and Mobility, 29 (3), 247–62.

Burnham, Kenneth P. and David R. Anderson (2002), Model Selection and Multimodel

Inference: A Practical Information-Theoretic Approach, (2nd edition). New York, NY:

Springer.

References 179

Campbell, Brian Milton (1969), The Existence of Evoked Set and Determinats of its

Magnitude in Brand Choice Behaviour, PhD dissertation. Columbia University. Ann

Arbor, MI: UMI Dissertation Services.

Catledge, Lara D. and James E. Pitkow (1995), “Characterizing Browsing Strategies in the

World-Wide Web,” Computer Networks and ISDN Systems, 27 (6), 1065–73.

Chan, Tat Y., Chunhua Wu, and Ying Xie (2011), “Measuring the Lifetime Value of

Customers Acquired from Google Search Advertising,” Marketing Science, 30 (5), 837–

50.

Chang, Yuhmiin and Esther Thorson (2004), “Television and Web Advertising Synergies,”

Journal of Advertising, 33 (2), 75–84.

Chapman, Mike (2011), “What Clicks Worldwide? Rich Media Can Help Lift Click-Through

- But Only a Little,” Adweek, (accessed July 1, 2015), [available at

http://www.adweek.com/news/advertising-branding/what-clicks-worldwide-132085].

Chatterjee, Patrali, Donna L. Hoffman, and Thomas P. Novak (2003), “Modeling the

Clickstream: Implications for Web-Based Advertising Efforts,” Marketing Science, 22

(4), 520–41.

Che, Hai and P. B. (Seethu) Seetharaman (2009), “‘Speed of Replacement’: Modeling Brand

Loyalty Using Last-Move Data,” Journal of Marketing Research, 46 (4), 494–505.

Chen, Jianqing, De Liu, and Andrew B. Whinston (2009), “Auctioning Keywords in Online

Search,” Journal of Marketing, 73 (July), 125–41.

Chierichetti, Flavio, Ravi Kumar, Prabhakar Raghavan, and Tamás Sarlós (2012), “Are Web

Users Really Markovian?,” Proceedings of the 21st international conference on World

Wide Web - WWW ’12, 609.

References 180

Cho, C. and Hongsik John Cheon (2004), “Why Do People Avoid Advertising On The

Internet?,” Journal of Advertising, 33 (4), 89–97.

Cho, Chang-Hoan, Jung-Gyo Lee, and Marye Tharp (2001), “Different Forced-Exposure

Levels to Banner Advertisements,” Journal of Advertising Research.

Constantinides, Efthymios and Peter Geurts (2005), “The Impact of Web Experience on

Virtual Buying Behaviour: An Empirical Study,” Journal of Customer Behaviour, 4 (3),

307–35.

Cooley, Robert, Bamshad Mobasher, and Jaideep Srivastava (1999), “Data Preparation for

Mining World Wide Web Browsing Patterns,” Knowledge and Information Systems, 1

(1), 5–32.

Cormen, Thomas H., Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein (2009),

Introduction to Algorithms, 3rd Edition, US: Cambridge, MA: Massachusetts Institute of

Technology.

Court, David, Dave Elzinga, Susan Mulder, and Ole Jørgen Vetvik (2009), “The Consumer

Decision Journey,” McKinsey Quarterly, 3, 96–107.

Cox, David Roxbee (1972), “Regression Models and Life-Tables,” Journal of the Royal

Statistical Society, Ser. B, 34 (2), 187–220.

Csikszentmihalyi, Mihaly (1975), “Play and Intrinsic Rewards,” Journal of Humanistic

Psychology, 15 (3), 41–63.

——— (1977), Beyond Boredom and Anxiety, San Francisco, CA: Jossey-Bass.

——— and Isabella Selega Csikszentmihalyi (1988), Optimal Experience: Psychological

Studies of Flow in Consciousness, Cambridge (UK): Cambridge University Press.

References 181

D’angelo, Frank (2009), “Happy Birthday, Digital Advertising!,” AdvertisingAge, (accessed

June 11, 2015), [available at available at

http://adage.com/digitalnext/post?article_id=139964].

Dahlen, Micael (2001), “Banner Advertisements Through a New Lens,” Journal of

Advertising Research, 41 (4), 23.

Dalessandro, Brian, Claudia Perlich, Ori Stitelman, and Foster Provost (2012), “Causally

Motivated Attribution for Online Advertising,” Working Paper.

Danaher, Peter J. (2007), “Modeling Page Views across Multiple Websites with an

Application to Internet Reach and Frequency Prediction,” Marketing Science, 26 (3),

422–37.

——— and Tracey S. Dagger (2013), “Comparing the Relative Effectiveness of Advertising

Channels: A Case Study of a Multimedia Blitz Campaign,” Journal of Marketing

Research, 50 (4), 517–34.

———, Guy W. Mullarkey, and Skander Essegaier (2006), “Factors Affecting Web Site Visit

Duration: A Cross-Domain Analysis,” Journal of Marketing Research, 43 (2), 182–94.

Dar, Eyal Even, Yishay Mansour, Vahab S. Mirrokni, S. Muthukrishnan, and Uri Nadav

(2009), “Bid Optimization for Broad Match Ad Auctions,” in WWW, World Wide Web

Conference.

David Collett (2015), Modelling Survival Data in Medical Research, 3rd edition, US:

Chapman and Hall/CRC.

Day, George S. (2011), “Closing the Marketing Capabilities Gap,” Journal of Marketing, 75

(4), 183–95.

Dou, Wenyu, Kai H. Lim, Chenting Su, Nan Zhou, and Nan Cui (2010), “Brand Positioning

Strategy Using Search Engine Marketing,” MIS Quarterly, 34 (2), 261–79.

References 182

Drèze, Xavier and François-Xavier Hussherr (2003), “Internet Advertising: Is Anybody

Watching?,” Journal of Interactive Marketing, 17 (4), 8–23.

——— and Fred Zufryden (1998), “Is Internet Advertising Ready for Prime Time?,” Journal

of Advertising Research, 38 (3), 7–18.

Duffy, Dennis L. (2005), “Affiliate Marketing and its Impact on E-Commerce,” Journal of

Consumer Marketing, 22 (3), 161–63.

Econsultancy (2012), “Marketing Attribution: Valuing the Customer Journey,” In Association

with Google Analytics. London, UK.

Edell, Julie A. and Kevin Lane Keller (1989), “The Information Processing of Coordinated

Media Campaigns,” Journal of Marketing Research, 26 (2), 149–63.

Edelman, Ben and Wesley Brandi (2015), “Risk, Information, and Incentives in Online

Affiliate Marketing,” Journal of Marketing Research, 52 (1), 1–12.

Edelman, David C. (2010), “Branding in the Digital Age: You’re Spending Your Money in

All the Wrong Places,” Harvard Business Review, 88 (12), 62–69.

eMarketer (2013), “Ecommerce Sales Topped $1 Trillion for First Time in 2012,” (accessed

July 26, 2015), [available at http://www.emarketer.com/Article/Ecommerce-Sales-

Topped-1-Trillion-First-Time-2012/1009649].

——— (2014a), “Global Ad Spending Growth to Double This Year,” (accessed July 25,

2015), [available at http://www.emarketer.com/Article/Global-Ad-Spending-Growth-

Double-This-Year/1010997].

——— (2014b), “Global B2C Ecommerce Sales to Hit $1.5 Trillion This Year Driven by

Growth in Emerging Markets,” (accessed July 26, 2015), [available at

http://www.emarketer.com/Article/Global-B2C-Ecommerce-Sales-Hit-15-Trillion-This-

Year-Driven-by-Growth-Emerging-Markets/1010575].

References 183

Feldman, Jon, S. Muthukrishnan, Martin Pál, and Cliff Stein (2007), “Budget Optimization in

Search-Based Advertising Auctions,” Proceedings of the 8th ACM conference on

Electronic Commerce - EC ’07.

Flosi, Stephanie, Gian M. Fulgoni, and Andrea Vollman (2013), “If an Advertisement Runs

Online and No One Sees It, Is It Still an Ad? Empirical Generalizations in Digital

Advertising,” Journal of Advertising Research, 53 (2), 192–99.

Fulgoni, Gian M. and Marie Pauline Mörn (2009), “Whither the Click? How Online

Advertising Works,” Journal of Advertising Research, 49 (2), 134–42.

Gallagher, Katherine, K. Dale Foster, and Jeffrey Parsons (2001), “The Medium Is Not the

Message: Advertising Effectiveness and Content Evaluation in Print and on the Web,”

Journal of Advertising Research, 41 (4), 57–70.

Ghani, Jawaid A. and Satish P. Deshpande (1994), “Task Characteristics and the Experience

of Optimal Flow in Human-Computer Interaction,” The Journal of Psychology, 128 (4),

381–91.

Ghose, Anindya and Sha Yang (2008), “Comparing Performance Metrics in Organic Search

with Sponsored Search Advertising Categories and Subject Descriptors,” in Proceedings

of the 2nd International Workshop on Data Mining and Audience Intelligence for

Advertising, ADKDD’08, 18–26.

——— and ——— (2009), “An Empirical Analysis of Search Engine Advertising:

Sponsored Search in Electronic Markets,” Management Science, 55 (10), 1605–22.

——— and ——— (2010), “Modeling Cross-Category Purchases in Sponsored Search

Advertising,” Working Paper, (accessed May 2, 2015), [available at

http://ssrn.com/abstract=1312864].

References 184

Godfrey, Andrea, Kathleen Seiders, and Glenn B. Voss (2011), “Enough Is Enough! The Fine

Line in Executing Multichannel Relational Communication,” Journal of Marketing, 75

(July), 94–109.

Goldfarb, Avi and Catherine Tucker (2011a), “Search Engine Advertising: Channel

Substitution When Pricing Ads to Context,” Management Science, 57 (3), 458–70.

——— and ——— (2011b), “Online Display Advertising: Targeting and Obtrusiveness,”

Marketing Science, 30 (3), 389–404.

Google (2013), “The Customer Journey to Online Purchase,” (accessed June 16, 2015),

[available at https://www.thinkwithgoogle.com/tools/customer-journey-to-online-

purchase.html#!/the-us/arts-and-entertainment/large/generic-paid-search].

Guo, Zhiling (2012), “Optimal Decision Making for Online Referral Marketing,” Decision

Support Systems, 52 (2), 373–83.

Ha, Louisa (2008), “Online Advertising Research in Advertising Journals: A Review,”

Journal of Current Issues and Research in Advertising, 30 (1), 31–48.

Haan, Evert de, Thorsten Wiesel, and Koen Pauwels (2013), “Which Advertising Forms

Make a Difference in Online Path to Purchase?,” Marketing Science Institute Working

Paper Series No. 13, Boston, MA.

Häubl, Gerald and Valerie Trifts (2000), “Consumer Decision Making in Online Shopping

Environments: The Effects of Interactive Decision Aids,” Marketing Science, 19 (1), 4–

21.

Hauser, John R. and Birger Wernerfelt (1990), “An Evaluation Cost Model of Consideration

Sets,” Journal of Consumer Research, 16 (4), 393–408.

He, Haibo and Edwardo A. Garcia (2009), “Learning from imbalanced data,” IEEE

Transactions on Knowledge and Data Engineering, 21 (9), 1263–84.

References 185

Hoffman, Donna L. and Thomas P. Novak (1996), “Marketing in Hypermedia Computer-

Mediated Environments: Conceptual Foundations,” Journal of Marketing, 60 (3), 50–68.

Hollis, Nigel (2005), “Ten Years of Learning on How Online Advertising Builds Brands,”

Journal of Advertising Research, 45 (2), 255–68.

Homburg, Christian, Viviana V. Steiner, and Dirk Totzek (2009), “Managing Dynamics in a

Customer,” Journal of Marketing, 73 (5), 70–89.

Hosmer, David W., Stanley Lemeshow, and Susanne May (2008), Applied Survival Analysis:

Regression Modeling of Time-to-Event Data, 2nd edition, Hoboken, NJ: Wiley.

Howard, John A. (1963), Consumer Behavior: Application of Theory, New York: McGraw-

Hill Book Company.

——— and Jagdish N. Sheth (1969), The Theory of Buyer Behavior, New York: John Wiley

& Sons, Inc.: John Wiley.

Huberman, Bernardo A., Peter L. T. Pirolli, James E. Pitkow, and Rajan M. Lukose (1998),

“Strong Regularities in World Wide Web Surfing,” Science, 280 (5360), 95–97.

Hui, Sam K., Peter S. Fader, and Eric T. Bradlow (2009), “Path Data in Marketing: An

Integrative Framework and Prospectus for Model Building,” Marketing Science, 28 (2),

320–35.

IBM Institute for Business Value (2011), “From Stretched to Strengthened – Insights from the

Global Chief Marketing Officer Study,” IBM CMO C-suite Studies.

Ilfeld, Johanna S. and Russell S. Winer (2002), “Generating Website Traffic,” Journal of

Advertising Research, 42 (5), 49–61.

Jagpal, Harsharanjeet (1981), “Measuring Joint Advertising Effects in Multiproduct Firms,”

Journal of Advertising Research, 21 (1), 65–69.

References 186

Jansen, Bernard J., Danielle L. Booth, and Amanda Spink (2008), “Determining the

Informational, Navigational, and Transactional Intent of Web Queries,” Information

Processing & Management, 44 (3), 1251–66.

——— and Simone Schuster (2011), “Bidding on the Buying Funnel for Sponsored Search

and Keyword Advertising,” Journal of Electronic Commerce Research, 12 (1), 1–18.

———, Kate Sobel, and Mimi Zhang (2011), “The Brand Effect of Key Phrases and

Advertisements in Sponsored Search,” International Journal of Electronic Commerce,

16 (1), 77–106.

——— and Amanda Spink (2009), “Investigating Customer Click Through Behaviour with

Integrated Sponsored and Nonsponsored Results,” International Journal of Internet

Marketing and Advertising, 5 (1-2), 74–94.

Jaworski, Bernard J. (2011), “On Managerial Relevance,” Journal of Marketing, 75 (4), 211–

24.

Jerath, Kinshuk, Liye Ma, and Young-Hoon Park (2014), “Consumer Click Behavior at a

Search Engine: The Role of Keyword Popularity,” Journal of Marketing Research, 51

(4), 480–86.

———, ———, ———, and Kannan Srinivasan (2011), “A ‘Position Paradox’ in Sponsored

Search Auctions,” Marketing Science, 30 (4), 612–27.

Johnson, Eric J. (2001), “Digitizing Consumer Research,” Journal of Consumer Research, 28

(2), 331–36.

Jordan, Patrick, Mohammad Mahdian, Sergei Vassilvitskii, and Erik Vee (2011), “The

Multiple Attribution Problem in Pay-Per-Conversion Advertising,” in Proceedings of the

34th International ACM SIGIR Conference on Research and Development in

Information Retrieval.

References 187

Kamakura, Wagner, Praveen K. Kopalle, and Donald R. Lehmann (2014), “Empirical

Generalizations in Retailing,” Journal of Retailing, 90 (2), 121–24.

Kass, Robert E. and Adrian E. Raftery (1995), “Bayes Factors,” Journal of the American

Statistical Association, 90 (430), 773–95.

Katona, Zsolt and Miklos Sarvary (2010), “The Race for Sponsored Links: Bidding Patterns

for Search Advertising,” Marketing Science, 29 (2), 199–215.

Kim, Juran and Sally J. McMillan (2008), “Evaluation of Internet Advertising Research. A

Bibliometric Analysis of Citations from Key Sources,” Journal of Advertising, 37 (1),

99–112.

Kireyev, Pavel, Koen Pauwels, and Sunil Gupta (2013), “Do Display Ads Influence Search?

Attribution and Dynamics in Online Advertising,” HBS Working Paper No. 13-070.

Boston, MA, (accessed March 12, 2015), [available at

http://www.hbs.edu/faculty/Publication Files/13-070.pdf].

Klapdor, Sebastian, Eva Anderl, Jan Schumann, and Florian Von Wangenheim (2015),

“Using Multichannel Behavior to Predict Online Conversions,” Journal of Advertising

Research, 55 (4), 433–42.

———, Florian von Wangenheim, and Jan Schumann (2014), “Finding the Right Words: The

Influence of Linguistic , Content- , and User-related Keyword Characteristics on

Performance of Paid Search Campaigns,” Journal of Interactive Marketing, 28 (4), 285–

301.

Koufaris, Marios (2002), “Applying the Technology Acceptance Model and Flow Theory to

Online Consumer Behavior,” Information Systems Research, 13 (2), 205–23.

References 188

Kumar, V., Vikram Bhaskaran, Rohan Mirchandani, and Milap Shah (2013), “Creating a

Measurable Social Media Marketing Strategy: Increasing the Value and ROI of

Intangibles and Tangibles for Hokey Pokey,” Marketing Science, 32 (2), 194–212.

Kvålseth, Tarald O. (1985), “Cautionary Note about R2,” The American Statistician, 39 (4),

279–85.

Lambrecht, Anja and Catherine Tucker (2013), “When Does Retargeting Work? Information

Specificity in Online Advertising,” Journal of Marketing Research, 50 (5), 561–76.

LeClerc, France, Bern H. Schmitt, and Laurette Dubé (1995), “Waiting Time and Decision

Making: Is Time like Money?,” Journal of Consumer Research, 22 (1), 110–19.

Leeflang, Peter S. H., Peter C. Verhoef, Peter Dahlström, and Tjark Freundt (2014),

“Challenges and Solutions for Marketing in a Digital Era,” European Management

Journal, 32 (1), 1–12.

——— and Dick R. Wittink (2000), “Building models for marketing decisions : Past , present

and future,” International Journal of Research in Marketing, 17 (2/3), 105–26.

Lehmann, Donald R., Leigh Mcalister, and Richard Staelin (2011), “Sophistication in

Research in Marketing,” Journal of Marketing, 75 (4), 155–65.

Lewis, Randall A., Justin M. Rao, and David H. Reiley (2011), “Here, There, and

Everywhere: Correlated Online Behaviors Can Lead to Overestimates of the Effects of

Advertising,” in Proceedings of the 20th International Conference on World Wide Web,

WWW 2011, Hyderabad, India: ACM, 157–66.

——— and David H. Reiley (2014), “Online Ads and Offline Sales: Measuring the Effect of

Retail Advertising via a Controlled Experiment on Yahoo!,” Quantitative Marketing and

Economics, 12 (3), 235–66.

References 189

Li, Hongshuang Alice and P.K. Kannan (2014), “Attributing Conversions in a Multichannel

Online Marketing Environment: An Empirical Model and a Field Experiment,” Journal

of Marketing Research, 51 (1), 40–56.

Libai, Barak, Eyal Biyalogorsky, and Eitan Gerstner (2003), “Setting Referral Fees in

Affiliate Marketing,” Journal of Service Research, 5 (4), 303–15.

Lilien, Gary L. (2011), “Bridging the Academic – Practitioner Devide in Marketing Decision

Models,” Journal of Marketing, 75 (4), 196–210.

Lin, Mingfeng, Henry C. Lucas Jr., and Galit Shmueli (2013), “Too Big to Fail: Large

Samples and the p-Value Problem,” Information Systems Research, 24 (4), 906–17.

Little, John D. C. (1970), “Models and Managers: The Concept of a Decision Calculus,”

Management Science, 16 (8), B466–85.

——— (1979), “Decision Support Systems for Marketing Managers,” Journal of Marketing,

43 (3), 9–26.

——— (2004a), “Models and Managers: The Concept of a Decision Calculus,” Management

Science, 50 (12), 1841–53.

——— (2004b), “Comments on ‘Models and Managers: The Concept of a Decision

Calculus,’” Management Science, 50 (12 Supplement), 1854–60.

Lodish, Leonard M. (2001), “Building Marketing Models that Make Money,” Interfaces, 31

(3), 45–55.

Luo, Xueming (2009), “Quantifying the Long-Term Impact of Negative Word of Mouth on

Cash Flows and Stock Prices,” Marketing Science, 28 (1), 148–65.

MacInnis, Deborah J. and Bernard J. Jaworski (1989), “Information Processing from

Advertisements: Toward an Integrative Framework,” Journal of Marketing, 53 (4), 1–23.

References 190

Manchanda, Puneet, Jean-Pierre Dubé, Khim Yong Goh, and Pradeep K. Chintagunta (2006),

“The Effect of Banner Advertising on Internet Purchasing,” Journal of Marketing

Research, 43 (1), 98–108.

Mandel, Naomi and Eric J. Johnson (2002), “When Web Pages Influence Choice: Effects of

Visual Primes on Experts and Novices,” Journal of Consumer Research, 29 (2), 235–45.

Marketing Science Institute [MSI] (2014), “2014-2016 Research Priorities,” (accessed May 3,

2015), [available at http://www.msi.org/research/2014–2201.

McFadden, Daniel (1974), “Conditional Logit Analysis of Qualitative Choice Behavior,” in

Frontiers of Econometric, P. Zarembka, edition New York: Academic Press, Inc., 105–

42.

McGraw, Kenneth O., Mark D. Tew, and John E. Williams (2000), “The Integrity of Web-

Delivered Experiments: Can You Trust the Data?,” Psychological Science, 11 (6), 502–

6.

McKinsey and Company, Inc. (2010), “McKinsey CMSO Forum,” Internal Conference

Contribution (Survey of CMOs and CSOs of European Companies, N = 126), Paris.

Mehta, Nitin, Surendra Rajiv, and Kannan Srinivasan (2003), “Price Uncertainty and

Consumer Search: A Structural Model of Consideration Set Formation,” Marketing

Science, 22 (1), 58–84.

Menard, Scott (2000), “Coefficients of Determination for Multiple Logistic Regression

Analysis,” The American Statistician, 54 (1), 17–24.

Moe, Wendy W. (2003), “Buying, Searching, or Browsing: Differentiating Between Online

Shoppers Using In-Store Navigational Clickstream,” Journal of Consumer Psychology,

13 (1), 29–39.

References 191

——— (2006), “An Empirical Two-Stage Choice Model with Varying Decision Rules

Applied to Internet Clickstream Data,” Journal of Marketing Research, 43 (4), 680–92.

——— and Peter S. Fader (2004a), “Capturing Evolving Visit Behavior in Clickstream

Data,” Journal of Interactive Marketing, 18 (1), 5–19.

——— and ——— (2004b), “Dynamic Conversion Behavior at E-Commerce Sites,”

Management Science, 50 (3), 326–35.

Moffett, Tina (2014), “The Forrester Wave: Cross-Channel Attribution Providers.”

Montgomery, Alan L., Shibo Li, Kannan Srinivasan, and John C. Liechty (2004), “Modeling

Online Browsing and Path Analysis Using Clickstream Data,” Marketing Science, 23 (4),

579–95.

Morimoto, Mariko and Susan Chang (2006), “Consumers’ Attitudes toward Unsolicited

Commercial E-mail and Postal Direct Mail Marketing Methods: Intrusiveness, Perceived

Loss of Control, and Irritation,” Journal of Interactive Advertising, 7 (1), 1–11.

Mulpuru, Sucharita (2011), “The Purchase Path Of Online Buyers,” Forrester Report.

Murray, Keith B. (1991), “A Test of Services Marketing Theory: Consumer Information

Acquisition Activities,” Journal of Marketing, 55 (1), 10–25.

Muthukrishnan, S., Martin Pál, and Zoya Svitkina (2007), “Stochastic Models for Budget

Optimization in Search-Based Advertising,” in Proceedings of the 16th international

Conference on World Wide Web - WWW ’07, Banff, Canada.

Nabout, Nadia Abou, Markus Lilienthal, and Bernd Skiera (2014), “Empirical Generalizations

in Search Engine Advertising,” Journal of Retailing, 90 (2), 206–16.

Nagelkerke, N. J. D. (1991), “A Note on a General Definition of the Coefficient of

Determination,” Biometrika, 78 (3), 691–92.

References 192

Naik, Prasad A. and Kay Peters (2009), “A Hierarchical Marketing Communications Model

of Online and Offline Media Synergies,” Journal of Interactive Marketing, 23 (4), 288–

99.

——— and Kalyan Raman (2003), “Understanding the Impact of Synergy in Multimedia

Communications,” Journal of Marketing Research, 40 (4), 375–88.

Neslin, Scott A., Sunil Gupta, Wagner Kamakura, Junxiang Lu, and Charlotte H. Mason

(2006), “Defection Detection: Measuring and Understanding the Predictive Accuracy of

Customer Churn Models,” Journal of Marketing Research, 43 (2), 204–11.

——— and Venkatesh Shankar (2009), “Key Issues in Multichannel Customer Management:

Current Knowledge and Future Directions,” Journal of Interactive Marketing, 23 (1),

70–81.

Nottorf, F. (2014), “Modeling the Clickstream Across Multiple Online Advertising Channels

Using a Binary Logit with Bayesian Mixture of Normals,” Electronic Commerce

Research and Applications, 13 (1), 45–55.

Novak, Thomas P., Donna L. Hoffman, and Yiu-Fai Yung (2000), “Measuring the Customer

Experience in Online Environments: A Structural Modeling Approach,” Marketing

Science, 19 (1), 22–42.

O’Donoghue, Ted and Matthew Rabin (1999), “Doing It Now or Later,” American Economic

Review, 89 (1), 103–24.

O’Quigley, John, Ronghui Xu, and Janez Stare (2005), “Explained Randomness in

Proportional Hazards Models,” Statistics in Medicine, 24 (3), 479–89.

Osur, Ari (2012), “The Forrester Wave: Interactive Attribution Vendors,” Forrester White

Paper. Cambridge, MA.

References 193

Padmanabhan, Balaji, Zhiqiang Zheng, and Steven O. Kimbrough (2001), “Personalization

From Incomplete Data: What You Don’t Know Can Hurt, Conference on Knowledge

Discovery in Data,” in Proceedings of the Seventh ACM SIGKDD International

Conference on Knowledge Discovery and Data Mining, 154–63.

Papatla, Purushottam and Amit Bhatnagar (2002), “Choosing the Right Mix of On-line

Affiliates: How Do You Select the Best?,” Journal of Advertising, 31 (3), 69–81.

Parkinson, Thomas L. and Michael Reilly (1979), “An Information Processing Approach to

Evoked Set Formation,” in Advances in Consumer Research Volume 06, eds. William L.

Wilkie, Ann Abor, MI: Association for Consumer Research, 227–32.

Pavlou, Paul A. and David W. Stewart (2000), “Measuring the Effects and Effectiveness of

Interactive Advertising: A Research Agenda,” Journal of Interactive Advertising, 1 (1),

62–78.

Payne, John W., James R. Bettman, and Eric J. Johnson (1988), “Adaptive Strategy Selection

in Decision Making,” Journal of Experimental Psychology: Learning, Memory, and

Cognition, 14 (3), 534–52.

Pedrick, James H. and Fred S. Zufryden (1991), “Evaluating the Impact of Advertising Media

Plans: A Model of Consumer Purchase Dynamics Using Single-Source Data,” Marketing

Science, 10 (2), 111–30.

Petty, Richard E. and John T. Cacioppo (1986), Communication and Persuasion: Central and

Peripheral Routes to Attitude Change, New York: Springer.

Pfeifer, Phillip E. and Robert L. Carraway (2000), “Modeling Customer Relationships as

Markov Chains,” Journal of Interactive Marketing, 14 (2), 43–55.

Van den Poel, Dirk and Wouter Buckinx (2005), “Predicting Online-Purchasing Behaviour,”

European Journal of Operational Research, 166 (2), 557–75.

References 194

PriceWaterhouseCoopers (2014), “IAB internet Advertising Revenue Report 2013.”

Provost, Foster, Brian Dalessandro, Rod Hook, Xiaohan Zhang, and Alan Murray (2009),

“Audience Selection for On-line Brand Advertising: Privacy-Friendly Social Network

Targeting,” in Proceedings of the 15th ACM SIGKDD International Conference on

Knowledge Discovery and Data Mining - KDD ’09.

Putsis, William P. Jr. and Narasimhan Srinivasan (1994), “Buying or Just Browsing? The

Duration of Purchase Deliberation,” Journal of Marketing Research, 31 (3), 393–402.

Qiu, Dingxi and Edward C. Malthouse (2009), “Quantifying the Indirect Effects of a

Marketing Contact,” Expert Systems with Applications, 36 (3), 6446–52.

Raman, Kalyan, Murali K. Mantrala, Shrihari Sridhar, and Yihui (Elina) Tang (2012),

“Optimal Resource Allocation with Time-Varying Marketing Effectiveness, Margins and

Costs,” Journal of Interactive Marketing, 26 (1), 43–52.

Reibstein, David J., George Day, and Jerry Wind (2009), “Guest Editorial: Is Marketing

Academia Losing its Way?,” Journal of Marketing, 73 (4), 1–3.

Roberts, John H. and James M. Lattin (1991), “Development and Testing of a Model of

Consideration Set Composition,” Journal of Marketing Research, 28 (4), 429–40.

——— and ——— (1997), “Consideration: Review of Research and Prospects for Future

Insights,” Journal of Marketing Research, 34 (3), 406–10.

Rose, Daniel E. and Danny Levinson (2004), “Understanding User Goals in Web Search,” in

Proceedings of the 13th International Conference on World Wide Web, WWW ’04, New

York, NY, USA, 13–19.

Royston, Patrick (2006), “Explained Variation for Survival Models,” The Stata Journal, 6 (1),

83–96.

References 195

——— and Willi Sauerbrei (2004), “A New Measure of Prognostic Separation in Survival

Data,” Statistics in Medicine, 23, 723–48.

Rusmevichientong, Paat and David P. Williamson (2006), “An Adaptive Algorithm for

Selecting Profitable Keywords for Search-Based Advertising Services,” in Proceedings

of the 7th ACM conference on Electronic Commerce - EC ’06, 260–69.

Rust, Roland T., Tim Ambler, Gregory S. Carpenter, V. Kumar, and Rajendra K. Srivastava

(2004), “Measuring Marketing Productivity: Current Knowledge and Future Directions,”

Journal of Marketing, 68 (4), 76–89.

———, Katherine N. Lemon, and Valarie A. Zeithaml (2004), “Return on Marketing: Using

Customer Equity to Focus Marketing Strategy,” Journal of Marketing, 68 (January),

109–27.

Rutz, Oliver J. and Randolph E. Bucklin (2007), “A Model of Individual Keyword

Performance in Paid Search Advertising,” Working Paper, (accessed June 3, 2015),

[available at

http://164.67.163.139/Documents/areas/fac/marketing/bucklin_keyword.pdf].

——— and ——— (2011), “From Generic to Branded: A Model of Spillover in Paid Search

Advertising,” Journal of Marketing Research, 48 (1), 87–102.

——— and ——— (2012), “Does Banner Advertising Affect Browsing for Brands?

Clickstream Choice Model Says Yes, for Some,” Quantitative Marketing and

Economics, 10 (2), 231–57.

———, ———, and Garrett P. Sonnier (2012), “A Latent Instrumental Variables Approach

to Modeling Keyword Conversion in Paid Search Advertising,” Journal of Marketing

Research, 49 (3), 306–19.

References 196

——— and M. Trusov (2011), “Zooming In on Paid Search Ads - A Consumer-Level Model

Calibrated on Aggregated Data,” Marketing Science, 30 (5), 789–800.

———, Michael Trusov, and Randolph E. Bucklin (2011), “Modeling Indirect Effects of Paid

Search Advertising: Which Keywords Lead to More Future Visits?,” Marketing Science,

30 (4), 646–65.

Scherer, Anne, Nancy V. Wünderlich, and Florian von Wangenheim (2015), “The Value of

Self-Service: Long-Term Effects of Technology-Based Self-Service Usage on Customer

Retention,” MIS Quarterly, 39 (1), 177–200.

Schwarz, Gideon E. (1978), “Estimating the Dimension of a Model,” The Annals of Statistics,

6 (2), 461–64.

Seetharaman, P. B. (Seethu) and Pradeep K. Chintagunta (2003), “The Proportional Hazard

Model for Purchase Timing: A Comparison of Alternative Specifications,” Journal of

Business & Economic Statistics, 21 (3), 368–82.

Shankar, Venkatesh and Edward C. Malthouse (2007), “The Growth of Interactions and

Dialogs in Interactive Marketing,” Journal of Interactive Marketing, 21 (2), 2–4.

Shao, Xuhui and Lexin Li (2011), “Data-Driven Multi-Touch Attribution Models,” in

Proceedings of the 17th ACM SIGKDD International Conference on Knowledge

Discovery and Data Mining - KDD ’11, New York, NY: ACM, 258–64.

Sherman, Lee and John Deighton (2001), “Banner Advertising: Measuring Effectiveness and

Optimizing Placement,” Journal of Interactive Marketing, 15 (2), 60–64.

Shocker, Allan D., Moshe Ben-Akiva, Bruno Boccara, and Prakash Nedungadi (1991),

“Consideration Set Influences on Consumer Decision-Making and Choice: Issues,

Models, and Suggestions,” Marketing Letters, 2 (3), 181–97.

References 197

Sismeiro, Catarina and Randolph E. Bucklin (2004), “Modeling Purchase Behavior at an E-

Commerce Web Site: A Task Completion Approach,” Journal of Marketing Research,

41 (3), 306–23.

Sood, Ashish, Gareth M. James, and Gerard J. Tellis (2009), “Functional Regression: A New

Model for Predicting Market Penetration of New Products,” Marketing Science, 28 (1),

36–51.

Spiggle, Susan and Murphy A. Sewall (1987), “A Choice Sets Model of Rretail Selection,”

Journal of Marketing, 51 (2), 97–111.

Spiliopoulou, Myra and Lukas C. Faulstich (1999), “WUM: A Tool for Web Utilization

Analysis,” Lecture Notes in Computer Science: The World Wide Web and Databases,

1590, 184–203.

Styan, George P. H. and Harry Smith (1964), “Markov Chains Applied to Marketing,”

Journal of Marketing Research, 1 (1), 50–55.

Tellis, Gerard J. (1988), “Advertising Exposure, Loyalty, and Brand Purchase: A Two-Stage

Model of Choice,” Journal of Marketing Research, 25 (2), 134–44.

———, Rajesh K. Chandy, Deborah MacInnis, and Pattana Thaivanich (2005), “Modeling

the Microeffects of Television Advertising: Which Ad Works, When, Where, for How

Long, and Why?,” Marketing Science, 24 (3), 351–66.

——— and Philip Hans Franses (2006), “Optimal Data Interval for Estimating Advertising

Response,” Marketing Science, 25 (3), 217–29.

Tezinde, Tito, Brett Smith, and Jamie Murphy (2002), “Getting Permission: Exploring Factors

Affecting Permission Marketing,” Journal of Interactive Marketing, 16 (4), 28–36.

References 198

The CMO Club & Visual IQ, Inc. (2014), “Building Bridges to the Promised Land: Big Data,

Attribution & Omni-Channel. A CMO Perspective,” (accessed January 29, 2015),

[available at https://thecmoclub.com/wp-content/uploads/2014/12/VisualIQ-Guide.pdf].

Trusov, Michael, Anand V. Bodapati, and Randolph E. Bucklin (2010), “Determining

Influential Users in Internet Social Networks,” Journal of Marketing Research, 47 (4),

643–58.

———, Randolph E. Bucklin, and Koen Pauwels (2009), “Effects of Word-of-Mouth Versus

Traditional Marketing: Findings from an Internet Social Networking Site,” Journal of

Marketing, 73 (5), 90–102.

Tucker, Catherine (2012), “The Implications of Improved Attribution and Measurability for

Online Advertising Markets,” Competition in the Online Environment.

——— (2014), “Social Networks, Personalized Advertising, and Privacy Controls,” NET

Institute Working Paper No. 10-07; MIT Sloan Research Paper No. 4851-10.

Vakratsas, Demetrios and Tim Ambler (1999), “How Advertising Works: What Do We

Really Know?,” The Journal of Marketing, 63 (1), 26–43.

Varadarajan, Rajan and Manjit S. Yadav (2009), “Marketing Strategy in an Internet-Enabled

Environment: A Retrospective on the First Ten Years of JIM and a Prospective on the

Next Ten Years,” Journal of Interactive Marketing, 23 (1), 11–22.

Wasserman, Larry (2000), “Bayesian Model Selection and Model Averaging,” Journal of

Mathematical Psychology, 44 (1), 92–107.

Watson, Richard T., Leyland F. Pitt, Pierre Berthon, and George M. Zinkhan (2002), “U-

Commerce: Expanding the Universe of Marketing,” Journal of the Academy of

Marketing Science, 30 (4), 333–47.

References 199

Webster, Jane, Linda Klebe Trevino, and Lisa Ryan (1993), “The Dimensionality and

Correlates of Flow in Human-Computer Interaction,” Computers in Human Behavior, 9

(4), 411–26.

Wiesel, Thorsten, Koen Pauwels, and Joep Arts (2011), “Marketing’s Profit Impact:

Quantifying Online and Off-line Funnel Progression,” Marketing Science, 30 (4), 604–

11.

Wright, Peter L. and Frederick R. Barbour (1977), “Phased Decision Strategies: Sequels to

Initial Screening,” in Multiple Criteria Decision Making: North Holland TIMS Studies in

the Management Science, M. Starr and M. Zeleny. eds. Amsterdam: North-Holland

Publishing Company, 91–109.

Wu, Jianan and Arvind Rangaswamy (2003), “A Fuzzy Set Model of Search and

Consideration with an Application to an Online Market,” Marketing Science, 22 (3),

411–34.

Wübben, Markus and Florian von Wangenheim (2008), “Instant Customer Base Analysis:

Managerial Heuristics Often ‘Get It Right,’” Journal of Marketing, 72 (3), 82–93.

Xia, Lan and D. Sudharshan (2002), “Effects of Interruptions on Consumer Online Decision

Processes,” Journal of Consumer Psychology, 12 (3), 265–80.

Xu, Lizhen, Jianqing Chen, and Andrew Whinston (2012), “Effects of the Presence of

Organic Listing in Search Advertising,” Information Systems Research, 23 (4), 1284–

1302.

———, Jason A. Duan, and Andrew Whinston (2014), “Path to Purchase: A Mutually

Exciting Point Process Model for Online Advertising and Conversion,” Management

Science, 60 (6), 1392–1412.

References 200

Yadav, Manjit S. and Paul A. Pavlou (2014), “Marketing in Computer-Mediated

Environments: Research Synthesis and New Directions,” Journal of Marketing, 78 (1),

20–40.

Yang, Sha and Anindya Ghose (2010), “Analyzing the Relationship Between Organic and

Sponsored Search Advertising: Positive, Negative, or Zero Interdependence?,”

Marketing Science, 29 (4), 602–23.

Yao, Song and Carl F. Mela (2011), “A Dynamic Model of Sponsored Search Advertising,”

Marketing Science, 30 (3), 447–68.

Zhu, June and Bernard Tan (2007), “Effectiveness of Blog Advertising: Impact of

Communicator Expertise, Advertising Intent, and Product Involvement. In S. Rivard & J.

Webster (Eds.),” in Proceedings of the 28th International Conference on Information

Systems, 121.

Appendix 203

7 Appendix

Table 28

Correlation Matrix by Data Set—Essay 1

Data Set 1 (Travel) SEA SEO Price

Comparison Display Newsletter Retargeting Affiliate Other

SEA 1.00 SEO .00*** 1.00 Price Comparison -.02*** -.01*** 1.00 Display -.08*** -.04*** -.01*** 1.00 Newsletter -.05*** -.01*** .00*** -.02*** 1.00 Retargeting .01*** .01*** .00 -.01*** .01*** 1.00 Affiliate -.05*** -.01*** .00** -.02*** .00** .00 1.00 Other -.06*** -.03*** -.01*** -.02*** -.02*** -.01*** -.02 *** 1.00Note: * p < .05, ** p < .01, *** p < .001.

Data Set 2 (Fashion) Type-in SEA SEO Price

Comparison Display Newsletter Affiliate Referrer

Type-in 1.00 SEA .01*** 1.00 SEO .00*** .10*** 1.00 Price Comparison .00 .00*** .01*** 1.00 Display .05*** .05*** .08*** .01 *** 1.00 Newsletter .07*** .05*** .04*** .00 .15 *** 1.00 Affiliate -.14*** -.09*** -.10*** -.01 *** -.01 *** -.05*** 1.00 Referrer -.04*** -.01*** .00 .00 ** .01 *** -.01*** -.12*** 1.00Note: * p < .05, ** p < .01, *** p < .001.

Data Set 3 (Fashion) Type-in SEA SEO Newsletter Retargeting Social Media Referrer Type-in 1.00 SEA -.06*** 1.00 SEO -.11*** -.07*** 1.00 Newsletter -.03*** -.03*** -.05*** 1.00 Retargeting .00 .00 .00 .00 1.00 Social Media -.06*** -.08*** -.15*** -.03*** .00 1.00 Referrer -.09*** -.10*** -.17*** -.04*** .00 -.11*** 1.00Note: * p < .05, ** p < .01, *** p < .001.

Data Set 4 (Luggage) Type-in SEA SEO Price

Comparison Retargeting Affiliate Referrer

Type-in 1.00 SEA -.01*** 1.00 SEO .00* -.07*** 1.00 Price Comparison .00* -.05*** .02*** 1.00 Retargeting .00 .06*** .00 .01*** 1.00 Affiliate .06*** .02*** .02*** .01*** .00 1.00 Referrer .00 -.14*** -.04*** .00** .00 .01*** 1.00Note: * p < .05, ** p < .01, *** p < .001.

Appendix 204

Table 29

Estimation Results for Logit Model 2—Essay 1

Data Set 1 (Travel) Data Set 2 (Fashion) Data Set 3 (Fashion) Data Set 4 (Luggage)

B Exp(B) ME B Exp(B) ME B Exp(B) ME B Exp(B) ME

t=1

Type-in 3,27 26,32 0,01 5,64 282,46 0,72 9,30 10895,44 0,98**

SEA 10,27 28827,82 0,97 4,38 80,19 0,16 5,56 259,87 0,56 6,70 808,79 0,90

SEO 4,54 93,85 0,04 5,01 149,17 0,23 7,14 1257,64 0,92 6,63 757,92 0,89

Price Comp. 8,86 7073,34 0,91 4,54 94,00 0,18 6,62 750,40 0,90

Display 8,66 5782,05 0,85 4,30 73,72 0,12

Newsletter 8,42 4525,44 0,88 5,56 260,55 0,27 6,19 490,26 0,68

Retargeting 9,21 9955,45 0,29 5,41 224,66 0,29 6,77 873,87 0,07

Social Media 5,45 232,40 0,46

Affiliate 9,15 9402,58 0,86 4,91 135,52 0,17 6,46 637,29 0,81

Referrer 5,61 273,54 0,18 6,11 450,80 0,58 7,82 2498,47 0,94

Other 7,67 2143,53 0,72

t=2

Type-in 1,55*** 4,71 0,01*** 1,87*** 6,52 0,05 *** 1,70*** 5,49 0,06***

SEA 2,02*** 7,51 0,01*** 1,74*** 5,71 0,01*** 1,69*** 5,43 0,04 *** 1,10*** 3,01 0,03***

SEO -1,79*** 0,17 0,00*** 1,87*** 6,48 0,01*** -6,89 0,00 -0,01 *** 1,05*** 2,85 0,02***

Price Comp. 1,68*** 5,37 0,01*** 1,97*** 7,16 0,01** 1,06*** 2,88 0,03***

Display 0,86*** 2,36 0,00*** 1,55*** 4,69 0,01***

Newsletter 1,24*** 3,47 0,00** 1,61*** 5,01 0,01*** 1,35*** 3,85 0,03 ***

Retargeting 1,11*** 3,04 0,00*** 1,21*** 3,36 0,02 *** 1,05*** 2,86 0,02***

Social Media 1,53*** 4,64 0,03 ***

Affiliate 0,68*** 1,98 0,00*** 1,45*** 4,26 0,01*** 0,96*** 2,61 0,02***

Referrer 1,76*** 5,84 0,01*** 1,97*** 7,20 0,06 *** 1,56*** 4,76 0,05***

Other 0,96*** 2,62 0,00***

t=3

Type-in 0,62*** 1,86 0,00*** 0,56** 1,76 0,01 * 0,99*** 2,69 0,02**

SEA 0,73*** 2,08 0,00*** 0,63*** 1,87 0,00*** 0,96*** 2,62 0,02 *** 0,55*** 1,73 0,01**

SEO -1,67*** 0,19 0,00*** 0,78*** 2,18 0,00*** -0,25 0,78 0,00 0,58** 1,78 0,01*

Price Comp. 0,87*** 2,39 0,00*** -0,13 0,88 0,00 0,40* 1,49 0,01

Display 0,62*** 1,87 0,00*** 0,58*** 1,79 0,00***

Newsletter 0,40 1,49 0,00 0,74*** 2,10 0,00*** 0,69*** 1,99 0,01 ***

Retargeting 0,47*** 1,60 0,00*** 0,74*** 2,10 0,01 *** 0,48*** 1,62 0,01***

Social Media 0,92*** 2,52 0,02 ***

Affiliate 0,63*** 1,88 0,00*** 0,64*** 1,89 0,00*** 0,42*** 1,52 0,01***

Referrer 0,67*** 1,95 0,00*** 1,05*** 2,87 0,02 *** 0,73*** 2,07 0,01***

Other 0,59*** 1,81 0,00**

t=4

Type-in 0,96*** 2,61 0,00*** 1,73*** 5,63 0,05 *** 1,77*** 5,90 0,06**

SEA 1,06*** 2,88 0,00*** 1,40*** 4,07 0,01*** 1,29*** 3,61 0,03 *** 0,67*** 1,96 0,01**

SEO -2,27*** 0,10 0,00*** 1,19*** 3,29 0,01*** 1,10*** 3,02 0,02 *** 0,89*** 2,43 0,02**

Price Comp. 1,27*** 3,58 0,00*** 1,57*** 4,79 0,01 0,78*** 2,18 0,02*

Display 1,32*** 3,76 0,00*** 1,02*** 2,77 0,00***

Newsletter 1,73*** 5,63 0,01** 1,27*** 3,54 0,01*** 1,21*** 3,37 0,02 ***

Retargeting 1,23*** 3,41 0,00*** 0,54*** 1,72 0,01***

Social Media 1,44*** 4,20 0,03 ***

Affiliate 1,59*** 4,90 0,01*** 1,09*** 2,98 0,00*** 0,56*** 1,75 0,01***

Referrer 1,06*** 2,88 0,00*** 1,52*** 4,56 0,04 *** 0,59*** 1,81 0,01***

Other 1,05*** 2,86 0,00*** 0,00 0,00 0,00

Intercept -13,57 0,00 -10,57 0,00 -10,57 0,00 -11,43 0,00

Observations 600,978 1,184,583 862,112 405,339

Note: * p < .05, ** p < .01, *** p < .001.

Appendix 205

Table 30

Estimation Results: Full Table on Channel Effects (Part 1 – 3)—Essay 3

DS 1 (Fashion) DS 2 (Fashion) DS 3 (Luggage) DS 4 (Travel) Base Model 1 Base Model 1 Base Model 1 Base Model 1 Variable B B B B B B B B Affiliate 0.033 *** 0.061*** 0.297*** 0.285*** 0.180 *** 0.200 ***

Display -3.014 *** -3.426 ***

Newsletter 0.092 *** 0.090*** -0.090 ** -0.091 **

PriceComparison 0.082 -1.651*** 0.156*** 0.152*** 0.015 0.036 **

Referrer -0.024 -0.122 -0.063 -0.196** 0.126 -1.060**

Retargeting 0.063 0.148*** 0.157** 0.089 -0.171 * -0.182 *

SEAGeneric 0.188 *** 0.299*** 0.094*** -0.033 -0.059*** -0.057*** -0.087 *** -0.029 SEABranded 0.263 *** 0.244*** 0.352*** 0.462*** 0.259*** 0.205*** 0.193 *** 0.199 ***

SEOGeneric 0.009 -0.047 0.187*** 0.215*** 0.051 0.074* -0.325 *** -0.353 ***

SEOBranded 0.151 *** 0.133*** 0.239*** 0.252*** 0.153*** 0.141* 0.044 *** 0.104 ***

Social 0.097*** 0.092***

TypeIn 0.162 *** 0.126*** 0.121*** 0.061* 0.070 0.003

Other 0.131*** 0.225***

PastAffiliate 0.034 *** 0.083*** -0.361*** -0.485*** -0.067 *** -0.449 ***

PastDisplay -0.269 * -0.471 *

PastNewsletter -0.002 -0.069*** 0.116 *** 0.007 PastPriceComparison -0.158 -1.554*** 0.049* -0.145** 0.046 *** 0.172 ***

PastReferrer 0.128 *** 0.061 0.052*** -0.091* 0.059 -1.137***

PastRetargeting 0.026 *** 0.006 -0.058 -0.135 0.067 *** 0.059 *

PastSEAGeneric -0.028 ** 0.015 -0.049*** -0.235*** 0.140*** 0.110*** 0.052 *** 0.139 ***

PastSEABranded -0.043 *** -0.156*** -0.052*** -0.065*** -0.131*** -0.590*** -0.012 ** -0.181 ***

PastSEOGeneric -0.004 -0.053** 0.040*** 0.051*** 0.037** -0.026 0.069 *** -0.005 PastSEOBranded -0.026 *** -0.058*** -0.029*** -0.093*** -0.116*** -0.514*** -0.008 -0.173 ***

PastSocial 0.022*** 0.003

PastTypeIn 0.010 *** -0.083*** -0.006*** -0.092*** 0.008 -0.123**

PastOther 0.185*** 0.417***

PastChannelNoAffiliate 0.982 *** 1.342*** -1.061*** -0.716*** -0.941 *** -0.832 ***

PastChannelNoDisplay 1.994 *** 1.955 ***

PastChannelNoNewsletter -0.044 0.025 -0.053 -0.082 PastChannelNoPriceComparison -0.424 * -1.914*** 0.615*** 0.862*** 0.468 *** 0.598 ***

PastChannelNoReferrer 0.421 *** 0.735*** 0.289*** 0.277*** 0.790*** -0.092

PastChannelNoRetargeting 0.226 *** 0.429*** 0.559*** 0.909*** 0.082 0.176 PastChannelNoSEAGeneric 0.208 *** 0.569*** -0.073** -0.156*** 0.207*** 0.498*** 0.393 *** 0.551 ***

PastChannelNoSEABranded -0.611 *** -0.509*** -0.167*** -0.064* -0.666*** -0.553*** -0.993 *** -0.928 ***

PastChannelNoSEOGeneric 0.554 *** 0.783*** 0.542*** 0.664*** 0.928*** 1.268*** -0.043 0.003 PastChannelNoSEOBranded -0.410 *** -0.241*** 0.068** 0.109*** 0.062 0.074 -0.439 *** -0.464 ***

PastChannelNoSocial 0.099*** 0.155***

PastChannelNoTypeIn -0.482 *** -0.303*** -0.086*** -0.081** -0.542*** -0.296***

PastChannelNoOther -0.082 0.167**

Affiliate × PastAffiliate -0.069*** 0.368** 0.374 ***

Display × PastDisplay 0.333 Newsletter × PastNewsletter 0.080*** 0.225 ***

PriceComparison × PastPriceComparison 1.990*** 0.226*** -0.129 ***

Referrer × PastReferrer 0.071 0.163*** 1.313***

Retargeting × PastRetargeting 0.013 0.146 0.036 SEAGeneric × PastSEAGeneric -0.040 0.347*** 0.063*** -0.086 ***

SEABranded × PastSEABranded 0.135*** 0.020 0.549*** 0.178 ***

SEOGeneric × PastSEOGeneric 0.076** -0.017* 0.106*** 0.129 ***

SEOBranded × PastSEOBranded 0.051*** 0.067*** 0.502*** 0.183 ***

Social × PastSocial 0.028***

TypeIn × PastTypeIn 0.095*** 0.088*** 0.138***

Other × PastOther -0.300***

Affiliate × PastChannelNoAffiliate 0.267*** -0.492*** -0.060 *

Display × PastChannelNoDisplay 0.670 ***

Newsletter × PastChannelNoNewsletter -0.357*** -0.479 ***

PriceComparison × PastChannelNoPriceComparison -0.043 -0.531*** -0.265 ***

Referrer × PastChannelNoReferrer 0.068 -0.096** -0.295***

Retargeting × PastChannelNoRetargeting -0.185*** -0.414*** -0.148 SEAGeneric × PastChannelNoSEAGeneric -0.142*** -0.403*** -0.611*** -0.070 *

SEABranded × PastChannelNoSEABranded -0.284*** -0.431*** -0.575*** -0.324 ***

SEOGeneric × PastChannelNoSEOGeneric -0.015 -0.183*** -0.502*** -0.321 ***

SEOBranded × PastChannelNoSEOBranded -0.198*** -0.277*** -0.571*** -0.482 ***

Social × PastChannelNoSocial -0.319***

TypeIn × PastChannelNoTypeIn -0.116*** -0.251*** -0.431***

Other × PastChannelNoOther -0.207**

N 1,184,582 1,184,582 862,114 862,114 405,343405,343 600,873 600,873Observations 1,398,267 1,398,267 964,836 964,836 461,108461,108 792,345 792,345Time at Risk 1,863,964 1,863,964 1,306,432 1,306,432 655,963655,963 1,171,897 1,171,897Log Likelihood -109,770.5 -109,369.2 -179,987.9 -179,763.3 -91,447.8 -91,235.8 -108,482.5 -108,236.3AIC 219,600.9 218,838.4 360,023.8 359,606.6 182,949.6182,561.7 217,019.0 216,562.6BIC 219,965.4 219,446.0 360,306.5 360,077.8 183,247.7183,058.5 217,331.8 217,083.8R2 (D) 0.379 0.401 0.185 0.188 0.391 0.391 0.576 0.581R2 (PH) 0.354 0.391 0.153 0.169 0.406 0.429 0.803 0.812Note: * p < .05, ** p < .01, *** p < .001. The base model excludes interaction effects. Model 1 includes interactions effects.

Appendix 206

Table 31

Estimation Results: Full Table on Category Effects—Essay 3

DS 1 (Fashion) DS 2 (Fashion) DS 3 (Luggage) DS 4 (Travel) Base Model 2 Base Model 2 Base Model 2 Base Model 2 Variable B B B B B B B B Affiliate -0.778*** -0.788*** 0.897*** 0.864 *** 0.299*** 0.372***

Display -5.452*** -5.413***

Newsletter 0.074*** 0.088*** -0.215*** -0.204***

PriceComparison 0.181 0.111 0.101** 0.085 * -0.619*** -0.605***

Referrer -0.414*** -0.487*** -0.211*** -0.218*** -0.414*** -0.409 ***

Retargeting 0.034 0.097** 0.108* 0.310 *** -0.456*** -0.437***

SEAGeneric 0.012 -0.002 0.165*** 0.162*** -0.158*** -0.165 *** -0.790*** -0.770***

SEABranded 0.348*** 0.352*** 0.502*** 0.523*** 0.715*** 0.738 *** 0.413*** 0.429***

SEOGeneric -0.262*** -0.279*** 0.004 0.006 -0.224*** -0.367 *** -0.535*** -0.523***

SEOBranded 0.144*** 0.152*** 0.292*** 0.307*** 0.222*** 0.218 *** 0.188*** 0.237***

Social 0.081*** 0.099***

TypeIn 0.250*** 0.248*** 0.233*** 0.238*** 0.588*** 0.552 ***

Other 0.138*** 0.142***

PastAffiliate 0.041*** 0.010 -0.062 0.050 -0.131*** -0.179***

PastDisplay -1.160*** -1.212***

PastNewsletter 0.008*** 0.095*** 0.201*** 0.360***

PastPriceComparison 0.199** 0.108 0.070*** 0.114 0.048*** -0.213***

PastReferrer 0.103*** -0.039 0.076*** 0.026 0.234*** 0.103

PastRetargeting 0.024*** 0.056** 0.061 0.350 0.128*** 0.261***

PastSEAGeneric 0.062*** -0.001 0.086*** 0.058** 0.170*** -0.172 *** 0.080*** -0.093***

PastSEABranded -0.028*** 0.023* -0.023*** -0.011 -0.144*** -0.186 *** -0.015*** 0.029PastSEOGeneric 0.018 -0.072*** 0.064*** 0.060*** 0.084*** -0.408 *** 0.155*** 0.034PastSEOBranded 0.000 0.011 -0.023*** 0.016 0.145*** -0.076 -0.003 0.122***

PastSocial 0.036*** 0.084***

PastTypeIn -0.003 0.050*** -0.011*** 0.009 0.005 -0.016

PastOther 0.219*** 0.272***

CIC × PastCIC 0.058*** 0.110*** 0.366 *** 0.161***

FIC × PastFIC -0.029*** -0.010 -0.382 * -0.128**

Navi × PastNavi -0.035*** -0.050*** -0.365 *** -0.125***

Info × PastInfo 0.050*** 0.001 -0.054 *** 0.082***

Integrated × PastIntegrated -0.003 -0.051*** 0.035 -0.145***

Separated × PastSeparated -0.035*** -0.030*** -0.154 *** -0.061***

Personal × PastPersonal -0.040** -0.050 0.173 *** -0.021Nonpersonal × PastNonpersonal 0.007 -0.019 -0.417 *** 0.161***

CIC × PastFIC -0.007 -0.024 -0.220 -0.189***

FIC × PastCIC -0.011 0.051* 0.041 ** 0.110***

Navi × PastInfo 0.050*** -0.026** -0.022 0.090***

Info × PastNavi -0.005 -0.014 -0.152 *** -0.057***

Integrated × PastSeparated -0.008* -0.015* -0.057 *** -0.013Separated × PastIntegrated 0.012** -0.027*** -0.048 -0.064***

Personal × PastNonpersonal 0.006 0.048** -0.194 ** * 0.210***

Nonpersonal × PastPersonal 0.018*** 0.000 0.073 *** -0.029

N 1,184,582 1,184,582 862,114 862,114405,343 405,343 600,873 600,873Observations 1,398,267 1,398,267 964,836 964,836461,108 461,108 792,345 792,345Time at Risk 1,863,964 1,863,964 1,306,432 1,306,432655,963 655,963 1,171,897 1,171,897Log Likelihood -111,550.8 -111,550.8 -180,997.7 -180,910.0 -92,739.4 -92,578.7 -110,269.2 -110,131.1AIC 223,141.6 223,042.7 362,027.3 361,883.9185,514.9 185,225.4 220,574.4 220,330.2BIC 223,384.6 223,480.1 362,215.8 362,260.9185,713.6 185,600.8 220,782.9 220,724.1R2 (D) 0.198 0.211 0.122 0.135 0.322 0.270 0.518 0.526R2 (PH) 0.171 0.178 0.081 0.088 0.249 0.270 0.727 0.734Note: * p < .05, ** p < .01, *** p < .001. The base model excludes interaction effects. Model 2 includes interaction effects.

Appendix 207

7.1 Definition of Online Marketing Channels (Essay 1 – 3)

In this dissertation, we implement a variety of online marketing channels. To ensure clarity, we

provide brief definitions of the channels applied.

Affiliate. Affiliate marketing is a form of commission based marketing in which a business (e.g.,

retailer) rewards the affiliate (e.g., a product review website) for referring a user toward the

businesses website. The commission may be paid out just by transferring the user or if the user

conducts a pre-defined transaction, such as a purchase (Duffy 2005; Edelman and Brandi 2015;

Libai, Biyalogorsky, and Gerstner 2003; Papatla and Bhatnagar 2002). In our data sets, affiliate

may include affiliate networks, such as Zanox, and coupon websites.

Display. Display advertising respectively banner advertising entails embedding a graphical

object with the advertising message into a website. This form of advertisement follows two

potential goals. First, to raise website traffic as users are redirected to the advertiser’s website

when they click on the display ad (Fulgoni and Mörn 2009; Hollis 2005). Second, to increase

brand awareness if the user refuses to click, however, is exposed to the display advertisement

(Drèze and Hussherr 2003).

Newsletter. Newsletter marketing, also known as email marketing, encompasses sending

marketing messages toward potential customers using email (Breuer, Brettel, and Engelen

2011). According to our definition, newsletter marketing includes promotional messages send

to users that have agreed to receive the newsletter (Tezinde, Smith, and Murphy 2002). The

data sets do not track unsolicited email messages known as spam (Morimoto and Chang 2006).

Price comparison. Price comparison websites or comparison shopping agents are vertical

search engines that allow users to compare product by price and features. Price comparison

websites aggregate product listings fee-based from a multitude of businesses and direct users

toward their website instead of selling by themselves (Breuer, Brettel, and Engelen 2011).

Appendix 208

Referrer. Online referral marketing or referrer is a marketing practice that rewards customers

(or websites) who successfully refer new customers to the ecommerce’s website (Guo 2012).

In contrast to affiliate marketing, the relationship between the referrer website and the referred

customer is closer, and not necessarily subject to a remuneration model.

Retargeted display. Retargeting or retargeted display is a subclass of display advertising that is

personalized toward the user based on his or her browsing history (Lambrecht and Tucker

2013). It aims to re-engage users who have visited an advertiser’s website, however, did not

conclude in a purchase event. Retargeted ad messages may include generic messages such as

the advertiser’s brand, or specific messages on the products the user has browsed before

(Lambrecht and Tucker 2013).

Search engine advertising (SEA). To date almost all general search engines such as Google

offer two results sections on many keywords entered by the user: Unpaid listings (SEO) and

paid listings (SEA) (Ghose and Yang 2009; Nabout, Lilienthal, and Skiera 2014). Search engine

advertising (SEA) or paid listings are typically shown at the top or the right side on the results

page and are designed by advertisers. The advertisers place a bid to be display alongside with

unpaid search engine results and pay a certain amount to the search engine for a user actually

clicking these advertisements (Yang and Ghose 2010). They are personalized in the sense that

the advertiser selects relevant keywords to trigger their ad (Nabout, Lilienthal, and Skiera 2014;

Yang and Ghose 2010).

Search engine optimization (SEO). As opposed to SEA, search engine optimization (SEO),

entails unpaid or, so-called, organic search listings based on the search engine’s algorithm

(Yang and Ghose 2010). Although, search engine optimization is free of charge, companies

permanently spend time and money to optimize their website positioning on the search engine’s

results page (Dou et al. 2010).

Appendix 209

Social media advertising. Social media advertising comprises a set of advertising platforms

belonging to the field of social media, such as social networks (e.g., Facebook), micromedia

(e.g., Twitter), or other (mobile) sharing platforms (e.g., Instagram) (Kumar et al. 2013). In one

of our data sets, the advertiser uses Facebook as advertising platform, which we define as social.

Hereby, the advertiser designs the ad (including a text and display banner) and selects a target

group to which the ad is shown by their self-declared interests, demographics and location.

Other. In data set 1, other comprises advertisements via Global Mail Exchange (GMX), a

German webmail service provider. As this form of advertising does not clearly fit to one of the

other categories we separated it out, yet continue to implement it into our models as it may

influence the overall advertising effect.

In addition to the above mentioned online channels, online marketing activities may include

further channels, which we do not cover in detail. They include marketing activities on video

or streaming platforms, classifieds, product placements, for instance, on blogs or social media

platforms such as Instagram and Pinterest, in-game advertisements, in-app advertisements and,

furthermore, may also include (or overlap with) different types of devices like mobile devices

and tablets.

7.2 Alternate Model Specifications (Essay 2)

7.2.1 Time-Dependent Covariates (Model 1)

As opposed to time-independent Cox models, the vector of covariates for customer i becomes

time-varying h6!t, Xf$ and the resulting hazard rate is, therefore,

h6!t, Xf$ = h9!t$ × exp !1 β?X6?f$.@

? �

( 14 )

The specification of the predictor covariates remains identical except that the time

component t in the multiplicative expression of the exponential term is added.

Appendix 210

7.2.2 Continuous Covariates (Model 1)

As our primary models apply binary covariates measuring the effect of the occurance of an

incident respective to its non-occurrence, we further compute corresponding models with

continuous covariates to control for the effect of multiple occurences—for instance, of the

micro-journey or channel clicks within one day. Thereby, the models represent the exact

number of micro-journeys, channel exposures, and navigational and informational switches by

the user each day. The model is specified as follows:

exp !1 β?X6?$@

? �= exp !β�AggMicroJourney6 + β#AggAffiliate6 + β3AggDisplay6

+ β2AggOther6 + βMAggNewsletter6 + βTAggPriceComparison6

+ βVAggReferrer6 + βXAggRetargeting6

+ β\AggPaidSearchGeneric6 + β�9AggPaidSearchBrand6

+ β��AggUnpaidSearchGeneric6 + β�#AggUnpaidSearchBrand6

+ β�3AggSocial6 + β�2AggTypeIn6 + β�MAggNaviSwitch6

+ β�TAggInfoSwitch6 + β�VTotalNoChannels6

+ β�XTotalNoClicks6$.

( 15 )

While β17 to β18 remain unchanged, β1 to β16 are modeled as continuous covariates

representing the exact number of exposures on the corrensponding day. We apply these

continuous coviariates accordingly for models 1a and 1b, as well as for all models in the

manuscript. To diminish any biases that might arise from our use of a proportional hazards

model, we apply a setting similar to a logit model. Due to the nature of our data sets and the

logit model, two slight adjustements apply. First, because logit models omit the chronological

order of clicks and the property of time, we add the control covariate of journey duration in

time. Second, we exclude the control covariate, the number of different channels, as it derives

Appendix 211

directly as linear expression from the channel exposure, and therefore becomes redundant.

Except for these two necessary adjustments, the predictors remain identical.

7.2.3 Vector of Covariates (Model 2b)

For Model 2b, the vector of covariates is identical to Model 2a, except for the general micro-

journey covariate. Thus, the coefficients β1 control for the effect of the occurrence of micro-

journeys. The coefficients β2 to β7 include effects of the micro-journey characteristics covering

the number of clicks (β2), the number of different channels (β3), the duration measured in time

(β4), the relative position within the whole journey (β5), and the occurrence of a navigational

switch (β6), respective informational switch within the micro-journey (β7). While β8 to β20

control for channel effects, β21 and β22 control for navigational and informational switches, and

β23 and β24 for the total number of channels and clicks in a similar manner as in Model 1.

Appendix 212

exp !1 β?X6?$@

? �= exp !β�MicroJourney6

+ β#MicroJourney6 × MicroJourneyClicks6

+ β3MicroJourney6 × MicroJourneyChannels6

+ β2MicroJourney6 × MicroJourneyDuration6

+ βMMicroJourney6 × MicroJourneyPosition6

+ βTMicroJourney6 × MicroJourneyNaviSwitch6

+ βVMicroJourney6 × MicroJourneyInfoSwitch6 + βXAffiliate6

+ β\Display6 + β�9Other6 + β��Newsletter6

+ β�#PriceComparison6 + β�3Referrer6 + β�2Retargeting6

+ β�MPaidSearchGeneric6 + β�TPaidSearchBrand6

+ β�VUnpaidSearchGeneric6 + β�XUnpaidSearchBrand6

+ β�\Social6 + β#9TypeIn6 + β#�NaviSwitch6 + β##InfoSwitch6

+ β#3TotalNoChannels6 + β#2TotalNoClicks6$.

( 16 )

7.3 Definition of R2D and R2

PH (Essay 2)

Royston and Sauerbrei’s (2004) R2D is a transformation based on the general index of

determination D proposed by Nagelkerke (1991), and is defined by

gh# = i#/ k#

l# + i#/k#, ( 17 )

where κ2 = 8 / π and σ2 = π2 / 6 concerning proportional hazards models. The formula

for the explained variation statistics in linear regression models with covariate vector x and

parameter vector β may be written as

Appendix 213

g# = mno!/.$l# + mno!/.$, ( 18 )

where D2 / κ2 in R2D plays the identical mathematical role as the var(xβ) in R2. D2 / κ2

may be interpreted as an estimate of the variance of the index xβ. R2PH is an alteration of

O’Quigley, Xu, and Stare’s (2005) modification of Nagelkerke’s (1991) R2 statistic for

proportional hazards models with censored data and defined by

gpq# = rs#/ 6 + r, ( 19 )

where V = ρ2κ / (1 – ρ2

κ) and ρ2κ = 1 – exp(– X2 /e) (O’Quigley, Xu, and Stare 2005;

Royston 2006). For further details refer to Royston and Sauerbrei (2004) and Royston (2006).