Rights / License: Research Collection In Copyright - Non ...48854/eth-48854-02.pdfThree Essays based...
Transcript of Rights / License: Research Collection In Copyright - Non ...48854/eth-48854-02.pdfThree Essays based...
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).