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 more information please consult the Terms of use . ETH Library
242
Embed
Rights / License: Research Collection In Copyright - Non ...48854/eth-48854-02.pdfThree Essays based on Clickstream Data: Analyzing, Understanding and Managing Online Customer Behavior
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Research Collection
Doctoral Thesis
Three Essays based on Clickstream Data: Analyzing,Understanding and Managing Online Customer Behavior
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
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)
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
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
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:
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
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
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:
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
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)
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)
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
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
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
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
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:
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***
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***
… 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).
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
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
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 ***
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***