Which Clicks Lead to Conversions?
Modeling User-journeys Across Multiple Types of Online Advertising
Florian Nottorf
Institut f¨ur Elektronische Gesch¨aftsprozesse, Leuphana University, L¨uneburg, Germany
Keywords:
Online Advertising, User-journey, Consumer Behavior, Purchasing Probabilities, Clickstream Data, Bayesian
Analyis, Mixture of Normals.
Abstract:
With an increase in the potential to allocate financial online advertising spending, managers are facing a so-
phisticated decision and allocation process. We developed a binary logit model with a Bayesian mixture
approach to address consumers’ buying decision processes and to account for the effects of multiple online
advertising channels. By analyzing data from a medium-sized online mail order business, we found inherent
differences in the effects of consumer clicks on purchasing probabilities across multiple advertising chan-
nels. We developed an alternative approach to account for the different attribution of success of advertising
channels—the average success probability (ASP). Compared to standardized metrics, we found paid search
advertising to be overestimated and retargeting display advertising to be underestimated. We further found
that the mixture approach is useful for considering heterogeneity in the individual propensity of consumers to
purchase; for the majority of consumers (more than 90%), repeated clicks on advertisements decrease their
probability of purchasing. In contrast with this segment, we found a smaller segment of consumers (nearly
10%) whose clicks on advertisements increase conversion probabilities. Our approaches will help managers
to better understand consumer online search and buying behavior over time and to allocate financial spending
more efficiently across multiple types of online advertising.
1 INTRODUCTION
In the last decade, the options for online advertising
have become increasingly complex. With the increase
in options to allocate funds for online advertising,
managers have sophisticated decisions to make. Stan-
dardized ratios that evaluate the profitability of adver-
tising campaigns are only partly helpful in evaluating
the short- and long-term effects of specific advertis-
ing channels; such ratios do not address the consumer
process that begins with becoming aware of a product
or brand through specific channels, such as display or
paid-search advertising. This problem may be clearly
illustrated with the following example.
Suppose we are a company that is engaged in on-
line advertising and we follow a user in his daily rou-
tine of Internet surfing; he checks his email, visits
his favorite website, reads his blogs, etc. He sees
dozens of display advertisements, including video,
social media and, retargeted advertisements. Perhaps
he clicks one or more advertisements, and maybe he
even clicks on one of our advertisements; unfortu-
nately, he does not purchase anything. After a few
hours have passed, he sees a display advertisement on
Facebook and clicks again on one of our advertise-
ments but does not purchase. He might remember our
company and our products and search for us using a
search engine such as Google or Yahoo; finally, that
specific user buys something from our company, and
our advertising activities have been effective. Such a
user-journey across multiple types of online advertis-
ing is illustrated in Figure 1.
However, to measure the overall effectiveness of
specific online advertising channels it is important
to evaluate which advertisement was crucial for the
user to become aware of our company and products.
On which type of online advertising should we spend
more to increase our chances of acquiring new con-
sumers in future?
Companies often use standalone metrics to evalu-
ate the profitability of specific online advertising cam-
paigns. However, these metrics do not capture con-
sumers’ decision-making processes over time and do
not account for the interaction effects between multi-
ple advertising activities.
1
Although there have been
1
The profitability and effectiveness of online advertising
141
Nottorf F..
Which Clicks Lead to Conversions? - Modeling User-journeys Across Multiple Types of Online Advertising.
DOI: 10.5220/0004504901410152
In Proceedings of the 4th International Conference on Data Communication Networking, 10th International Conference on e-Business and 4th
International Conference on Optical Communication Systems (ICE-B-2013), pages 141-152
ISBN: 978-989-8565-72-3
Copyright
c
 2013 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Illustration of a User-Journey Across Multiple Types of Advertising.
attempts to attribute online sales success to multi-
ple types of online advertising (uniquedigital, 2012),
the effect of advertising on the individual behavior
of consumers has not yet been comprehensively an-
alyzed.
We build on the specification of Chatterjee et al.
(2003), who model the probability of a consumer to
click on a display advertisement using a binary logit
with a normal prior distribution. However, we extend
their specification in two principal ways. First, we
model the probability of consumers purchasing by in-
cluding multiple advertising channels to address the
complex allocating processes used by companies to-
day to manage online advertising campaigns (such as
display advertising, social media, and paid search ad-
vertising). Second, we employ a Bayesian mixture
of normals approach, which offers more flexibility to
address consumer heterogeneity than standard normal
prior distributions (Nottorf, 2013; Rutz and Bucklin,
2011). The proposed model uses anonymized user-
level data to help managers understand the effects of
specific advertising channels on individual consumer
behavior and online purchasing processes.
This paper is structured as follows. First, we will
briefly review previous research related to this work.
Second, we will examine the general model specifica-
campaigns may be assessed by using different ratios, such
as click-through rate (CTR), which is the ratio of clicks
to impressions for a specific type of advertising; conver-
sion rate (CVR), which is defined as the number of pur-
chases in relation to the number of clicks; and cost per click
(CPC) and cost per order (CPO) as measures of the effi-
ciency of promotional activities (Athey and Ellison, 2009;
Nottorf and Funk, 2013a).
tion and extend the previous work of Chatterjee et al.
(2003) to model consumer probabilities of purchasing
after clicking on multiple advertising channels. This
section is followed by the introduction of our unique
dataset that includes detailed user-journeys and con-
sumer interactions with multiple types of online ad-
vertising. After outlining our findings and discussing
our results, we conclude this work by highlighting its
limitations and providing suggestions for future re-
search.
2 RELATED WORK
This work is related to several streams of research be-
cause we analyze multiple online advertising channels
and their influences on consumer online purchasing
behavior.
An emerging stream of research is analyzing paid
search advertising from the advertiser’s perspective
and providing important initial insights into consumer
clicking and purchasing behavior (see, for instance,
Ghose and Yang (2009) , Rutz et al. (2012), or
Rutz.2011c). For example, Rutz et al. (2011) analyze
the long-term effects of consumers’ online search ac-
tivities on subsequent direct-type-in visitors and de-
velop a hierarchical Bayesian elastic net to address
the “large p, small n” problem. The authors demon-
strate that normal ridge and LASSO regressions push
the limits when analyzing the effects of thousands of
keywords that are normally used in companies’ paid
search advertising campaigns.
Focusing on display advertising, Chatterjee et al.
(2003) and Rutz and Bucklin (2011) find signifi-
ICE-B2013-InternationalConferenceone-Business
142

cant heterogeneity in the propensity of consumers to
click on banner advertisements. In addition, Dana-
her and Mullarkey (2003) introduce user involvement
as an important factor that influences the effective-
ness of banner advertising; characterizing user inten-
tions as either goal-directed or in surfing mode in-
dicates that there will be significant differences in
the effectiveness of banner advertising. Goldfarb and
Tucker (2011) show that the synergies between differ-
ent types of banner advertising are negative; targeted
and obtrusive banner advertising campaigns indepen-
dently increase the purchase intent of consumers,
whereas combining them is less effective.
Limited research has investigated the online be-
havior of users while considering multiple advertis-
ing channels simultaneously (Neslin et al., 2006; Nes-
lin and Shankar, 2009). Previous research on me-
dia synergy effects has primarily focused on within-
media interaction effects of offline advertising efforts,
such as television, radio, and print (Bass et al., 2007;
Naik and Raman, 2003), or model cross media syn-
ergies between offline and online advertising types
(Ansari et al., 2008; Deleersnyder et al., 2002; Ilfeld
and Winer, 2002; Naik and Peters, 2009).
Although studies of paid search advertising have
increased significantly over the last several years,
there have been few attempts to examine an integrated
model of online advertising efforts, for example, to
explain the effects of the interaction between display
and paid search advertising. Although various types
of online advertising arrived in the last decade, sci-
entific works on cross channel advertising effects fo-
cused on modeling data that is aggregated on a spe-
cific time scale (e.g., week or day) on a specific type
of advertising channel (e.g., “display” or “search”).
Although certain studies have focused on the ef-
fects of several types of online advertising, such as
display and paid search advertising (Dinner et al.,
2011; Wiesel et al., 2011), such studies do not inves-
tigate the influence of these types of advertising on
the online behavior of individual consumers as a re-
sult of aggregation biases (Abhishek et al., 2011), and
they do not consider the interaction effects of different
types of online advertising activities on consumers.
The possibility of “following” individual users across
multiple types of online advertising is relatively new;
user-level data is collected through cookie tracking by
companies’ advertising-servers. Nottorf (2013) and
Nottorf and Funk (2013b) employ such data to an-
alyze the effects of repeated paid search advertise-
ments and banner, video and retargeted display ad-
vertising on consumer click probabilities; they find
both differences in the effects of repeated advertising
exposure across multiple types of display advertising
and positive interactive effects between display and
paid search advertising in influencing consumer click
probabilities.
We also use such user-level data to distinguish
multiple types of advertising at the level of individual
consumers. As opposed to Nottorf (2013) and Not-
torf and Funk (2013b), we analyze the effects of mul-
tiple advertising-specific clicks on consumers’ indi-
vidual conversion probabilities. Therefore, we con-
tribute to existing research at several junctures be-
cause we analyze the effects of multiple types of on-
line advertising—such as paid search advertising, dis-
play advertising, and newsletter mailings—on con-
sumer purchasing behavior at the individual user-
level.
3 MODELING INDIVIDUAL
CONVERSIONS ACROSS
MULTIPLE ONLINE
ADVERTISING CHANNELS
3.1 General Specification
We extend the specification of Chatterjee et al. (2003)
and follow Nottorf (2013) to model the probability of
consumer i purchasing at time t in session s, and we
consider multiple advertising channels. Therefore, we
specify the following:
• advertising-specific intercepts, I
ist
,
• variables accounting for short-term advertising-
specific effects, X
ist
, and
• variables accounting for long-term advertising-
specific effects, Y
is
.
I
ist
: Advertising-specific Intercepts
First, we define multiple advertising-specific
intercepts that account for the probability of con-
sumers purchasing after clicking on a respective
advertisement. We include these advertising-specific
intercepts because we expect consumers’ likelihood
of purchasing to vary strongly with the types of
advertising a consumer may click on. For example,
while a click on a paid search advertisement is
followed by a consumer’s active search for specific
terms, a click on a display advertisement may occur
more or less by accident. Chatterjee et al. (2003)
denote this intercept as consumers’ click-proneness.
Because we model consumers purchasing proba-
bilities across multiple advertising channels, we
denote these terms as consumers’ advertising-specific
WhichClicksLeadtoConversions?-ModelingUser-journeysAcrossMultipleTypesofOnlineAdvertising
143

conversion proneness. These intercepts become 1 if
a consumer clicks on a respective advertisement and
remains at 0 otherwise.
X
ist
: Short-term Advertising Effects
While a consumer may not purchase after his first
click on an advertisement within a current session,
that probability may increase with subsequent ones.
Furthermore, that probability of purchasing might
also vary with the type of advertising the consumer
repeatedly clicks on. Therefore, we include the
number of consumer clicks on specific advertising
channels within a current session. These variables
should capture the short-term effects (i.e., session-
length of one hour) of advertisement-specific clicks
on an individual consumer’s conversion probabilities.
If a consumer does not click on a specific type
of advertising in a current session, the respective
variable remains at 0.
Y
is
: Long-term Advertising Effects
The long-term effects of advertising clicks on con-
sumer conversion probabilities must also be modeled.
We accomplish this because we account for the num-
ber of all advertising-specific clicks and not just those
in a current session. We expect the long-term effects
of clicks on conversion probabilities to strongly vary
across consumers (Danaher and Mullarkey, 2003;
Rutz and Bucklin, 2011); while we expect the effect
of repeated clicks on specific advertisements to be
positive for certain consumers (i.e., each click in their
respective user-journey increases the awareness that a
company might be relevant for their buying process),
an increase in clicks on specific advertising channels
may decrease the probability of purchasing for other
consumers (i.e., no matter how often these consumers
click on a specific ad, they cannot be convinced to
purchase something). If a consumer does not click
at all on a specific type of advertising, the respective
variable remains 0.
To model the individual contribution of each ad-
vertising effort and its effect on consumer conversion
probabilities, we specify a binary logit choice model
(Chatterjee et al., 2003). The probability that con-
sumer i converts subsequent to a click on an online
advertisement at time t in session s is modeled as fol-
lows:
Con
ist
=
(
1 if user i converts at time t in session s
0 otherwise,
(1)
with the probability
Pr(Con
ist
= 1) =
exp(I
ist
α
i
+ X
ist
β
i
+Y
is
γ
i
+ ε
ist
)
1+ exp(I
ist
α
i
+ X
ist
β
i
+Y
is
γ
i
+ ε
ist
)
(2)
where I
ist
are advertisement-specific intercepts mod-
eling the consumers’ individual likelihood (prone-
ness) of conversion after clicking on a respective ad-
vertisement. X
ist
are variables varying within (t),
across sessions (s), and across consumers (i), whereas
Y
is
are variables varying across sessions (s) and con-
sumers (i), and α
i
, β
i
, and γ
i
are consumer-specific
parameters to be estimated.
3.2 Variable Specification
The variable X
ist
includes the number of consumer
advertising-specific clicks within a current session.
Furthermore, we follow Chatterjee et al. (2003) and
define the following additional variables incorporated
in X
ist
: x
con
is(t−1)
is the cumulative number of conver-
sions until t − 1 in the current session s for a spe-
cific consumer i, and Con
is(t−1)
is an indicator func-
tion that assumes the value 1 if a consumer has al-
ready purchased in t − 1. We assume both variables
to have a negative influence on subsequent consumer
conversions; if a consumer has already bought some-
thing in t − 1 or within his current session s, it might
be highly unlikely that he will purchase again in the
short-term. Furthermore, we define TLCon
ist
as the
logarithm time since a consumer last made a conver-
sion; if a consumer has never purchased, the variable
remains zero.
Modern tracking software also allows companies
to capture consumers’ on-site clicks and on-site time.
The former refers to the number of clicks a user has
made on the company’s website after he gets redi-
rected by clicking on an advertisement, whereas the
latter denotes the time the user browses through the
company’s website. Therefore, we include the cumu-
lative number of on-site clicks of a user i within ses-
sion s until time t as x
Onsite-clicks
ist
and the cumulative
logarithm on-site time as x
Onsite-time
ist
. We propose that
a user shows a higher involvement in the purchasing
process with increasing on-site time and clicks, which
may increase the probability of a user to convert.
The same might hold true for the number of brand-
related activities. We therefore also include x
Brand
ist
,
which refers to the cumulative number of brand-
related clicks made by user i within session s until
time t, and Brand
ist
, which is an indicator function
assuming 1 if a consumer i performs a brand-related
activity at time t in session s. We define a brand-
related activity as a click on an online advertisement
ICE-B2013-InternationalConferenceone-Business
144

that accompanies prior brand-related knowledge. For
example, if a user clicks on a paid search advertise-
ment after searching for brand-related terms (i.e., the
keyword includes the name of the company in ques-
tion, such as “Staples pen” instead of just “pen”),
we denote that click as a brand-related click. The
same holds true for users’ direct visits to the com-
pany’s website (i.e., the user directly types the name
of the company’s website into the web-browser or
uses bookmarks).
Y
is
accounts for the long-term effects of
advertisement-specific clicks on consumers’ in-
dividual conversion probabilities, i.e., the cumulative
number of clicks on a specific advertisement per
consumer i until session s. Furthermore, y
con
i(s−1)
is
the cumulative number of conversions in previous
sessions. y
Onsite-clicks
is
and y
Onsite-time
is
refer to cumula-
tive on-site clicks and time, respectively. The total
cumulative number of brand-related clicks is modeled
by y
Brand
is
. IST
is
is the logarithm of the intersession
duration between session s and s − 1; it remains at
zero if a consumer is active in only one session.
Session
is
refers to the number of sessions in which a
consumer has already clicked on advertisements.
3.3 Data
We use a dataset from a regular online shop that will
remain anonymous, at its own request. The dataset
consists of information on individual consumers and
the point in time at which they clicked on differ-
ent ads, such as retargeted banner and paid search
advertisements. Following consumers across multi-
ple advertising channels and types is performed with
cookie-tracking software and respective advertising
servers. The data was collected within a one-month
period (between 2012 and 2013). To analyze click-
stream data, we take into account only those user-
journeys that have more than three advertising touch
points. The finale dataset still consists several thou-
sand users.
2
Further, the dataset contains information about the
following types of advertising that a user has clicked
on:
• Search. If a user has searched for a keyword, if
the company in question advertises on search en-
gines such as Google, and if the user clicked on a
respective paid search advertisement of that com-
pany, then we denote this interaction as a “search”
click. Furthermore, the company in question is
listed on the results page of a search engine if
2
The datasets have been sanitized for reasons of confi-
dentiality.
a user searches specific keywords for which the
company has been classified as “relevant” by the
search engine. If a user clicks on links of such
organic results page listings, we also denote this
interaction as a “search” click.
• Price. Companies might pay for becoming linked
on shopping comparison sites such as Nextag. If
a user compares the price of a product and clicks
on a link of the company in question, we denote
this interaction as a “price” click.
• Retargeting. Display advertisements, such as ban-
ners, may be individualized on a user-specific
level. For example, if a user searches for a spe-
cific product and gets redirected to the website of
a company, the company might individualize the
display advertisements in a user’s later browsing
routines to the extent that it re-targets the specific
user and displays the specific (or related) product
in the banner advertisement that the user was orig-
inally looking for at the company’s website. We
denote users’ clicks on such retargeted display ad-
vertisements as “retargeting” clicks.
• Direct. If a user directly visits the website of the
company in question, for example, via bookmarks
or direct-type-ins, we denote this interaction as a
“direct” click.
3
• Other. There are other types of advertising con-
tacts, such as social media or newsletter mailings.
Because of their minimal total advertising con-
tacts, we aggregate them and denote a user’s click
on such types of advertising as an “other” click.
Table 1 lists the number of clicks on each type of on-
line advertising.
Because there is no accessible information avail-
able on the number of consumer sessions and their du-
ration, we manually define a session as a sequence of
advertising exposures with breaks that do not exceed
60 minutes. Demographic information is not avail-
able. We report the descriptive statistics of our final
dataset in Table 2.
3.4 Bayesian Mixture of Normals
Although the standard normal model as it has been ap-
plied by Chatterjee et al. (2003) to model consumers’
clickstream is capable of performing analyses with
many consumers and properly accounts for hetero-
geneity (Allenby and Ginter, 1995; Lenk et al., 1996),
we use a Bayesian mixture approach to account for
3
Although direct visits are not any type of advertising,
we will denote and treat it as an advertising channel for the
sake of convenience.
WhichClicksLeadtoConversions?-ModelingUser-journeysAcrossMultipleTypesofOnlineAdvertising
145

Table 1: General Information on the Dataset.
clicks in total 100.00%
on search 51.42%
on price 16.15%
on retargeting 11.85%
on direct 14.95%
on other 5.63%
conversions 2.14%
Notes: We report the number of advertising-specific clicks,
and the number of conversions within our dataset. For pri-
vacy reasons, the distribution of the advertising channels
have been sanitized.
consumer heterogeneity and to determine the set of
individual parameters. This mixture of normals ap-
proach enables us to find multiple cluster of users
whose conversion probability is modeled significantly
differently from those users of a different cluster (i.e.,
from a different normal distribution). For example,
the conversion probabilities of one cluster of indi-
vidual consumers may increase with each additional
click on a specific type of advertising (wear-in effect),
whereas the conversion probabilities of another clus-
ter may decrease (wear-out effect).
For the sake of convenience, we denote the set of
consumers’ individual parameters to be estimated as:
θ
i
= {α
i
, β
i
, γ
i
} (3)
Because we assume that the tendency of consumers
to purchase will vary significantly (i.e., extensive vs.
impulsive buying decision process), a mixture ap-
proach offers more flexibility for capturing hetero-
geneity than the standard normal approach (Rutz and
Bucklin, 2011). This assumption is consistent with
previous research that classifies the online searches
of users according to navigational, transactional, or
informational intentions (Broder, 2002) and indicates
strong differences in the effectiveness of banner ad-
vertising with respect to consumer involvement lev-
els (Danaher and Mullarkey, 2003; Nottorf and Funk,
2013a). We specify the Bayesian mixture approach,
following Rossi et al. (2005):
θ
i
∼ N(µ
ind
i
, Σ
ind
i
), (4)
ind
i
∼ Multinomial
K
(pvec), (5)
where ind
i
is an indicator latent variable from which
component observation i is derived. ind takes on val-
ues 1,..., K, and pvec is a vector of mixture prob-
abilities of length K. We use uninformative hyper-
priors pvec ∼ Dirichlet(α), µ
k
∼ Gaussian(µ
0
, Σ
0
),
and Σ
k
∼ Wishart(υ, V).
We apply a MCMC algorithm including a hybrid
Gibbs Sampler with a random walk Metropolis step
for the coefficients for each consumer and utilize the
Table 2: Descriptive Statistics of the Covariates.
Variables min max mean sd
Indicator Variables
I
search
ist
0 1 0.26 0.44
I
price
ist
0 1 0.08 0.27
I
retargeting
ist
0 1 0.06 0.24
I
direct
ist
0 1 0.07 0.26
I
other
ist
0 1 0.03 0.17
Intrasession Variables
x
search
ist
0 16 0.50 1.07
x
price
ist
0 286 1.23 13.96
x
retargeting
ist
0 13 0.09 0.44
x
direct
ist
0 46 0.12 1.00
x
other
ist
0 14 0.06 0.42
x
con
is(t−1)
0 3 0.01 0.11
Con
is(t−1)
0 1 0.01 0.09
TLCon
ist
0 7.58 0.15 0.89
x
Onsite-clicks
ist
0 450 2.31 8.95
x
Onsite-time
ist
0 125 1.60 4.47
x
Brand
ist
0 46 0.15 1.03
Brand
ist
0 1 0.09 0.29
Intersession Variables
y
search
is
0 44 0.91 2.61
y
price
is
0 331 1.71 17.93
y
retargeting
is
0 33 0.31 1.66
y
direct
is
0 86 0.64 3.80
y
other
is
0 40 0.19 1.42
y
con
i(s−1)
0 65 0.11 1.92
y
Onsite-clicks
is
0 510 6.28 24.64
y
Onsite-time
is
0 364 4.08 15.24
y
Brand
is
0 86 0.71 3.85
IST
is
0 7.66 1.26 2.02
Session
is
1 82 2.43 6.38
Notes: We report both the minimum/maximum and the
mean/standard deviation
of the covariates used in our study.
R -package bayesm by Rossi et al. (2005). We perform
40,000 iterations and use every twentieth draw of the
last 10,000 iterations to compute the conditional dis-
tributions.
4 RESULTS
4.1 Benchmarking Alternative Models
We benchmark multiple model specifications by mod-
ifying the number of mixture components K. We
compute the log likelihood (LL) and the Bayesian in-
formation criterion (BIC) to analyze fit performances.
The latter criterion penalizes the incorporation of ad-
ditional parameters—such as an additional number of
ICE-B2013-InternationalConferenceone-Business
146

Table 3: Model Comparison: Benchmarking Fit Perfor-
mances.
Model (Num. of Parameters) LL BIC
2 Mixture Components (56) -253.9 374.4
3 Mixture Components (84) -241.2 421.9
4 Mixture Components (112) -241.2 482.2
Standard Normal Distribution (28) -383.3 443.5
Notes: We report the log likelihood (LL) and the Bayesian
information criterion (BIC) for both the proposed and alter-
native model specifications.
mixture components—in the model.
As reported in Table 3, the model with two
mixture components performs best (BIC = 374.4).
There seem to be two cluster of consumers who
react differently to online advertising as the two-
mixture-components model exhibits superior perfor-
mance. The benefit of additional components does
not increase the relative fit performance. As expected,
the model that is reduced to a prior model with a stan-
dard normal distribution performs significantly more
poorly than the models with multiple mixture compo-
nents (BIC = 443.5).
4.2 Key Results
The parameter estimates for the two-mixture-
component model specification are reported in Ta-
ble 4. Our findings vary substantially across the two-
component groups. The larger group, group 1, has
a segment size of 91% and thus clearly represents
the majority of consumers, whereas the second group,
group 2, represents a segment size of 9%, which is a
small number of consumers.
We report the results for the intrasession effects,
intersession effects, and for the advertising-specific
conversion proneness successively for both groups.
Intrasession Effects
Group 1, segment size of 91%
For the majority of consumers, represented by
group 1 (the first component of the mixture model),
we find the cumulative number of clicks on all types
of advertising within a session to decrease purchasing
probability (see x
search
ist
...x
other
ist
),
4
i.e., the likelihood
of consumers purchasing is highest after they have
clicked for the first time on a particular advertisement
and are redirected to the company’s website within a
current session. For this segment of consumers, we
4
The findings are significantly negative for x
direct
ist
and
x
other
ist
.
find further differences in the effects of cumulative
clicks during current sessions across differenttypes of
advertisements. For example, each additional direct
visit (x
direct
ist
= -4.22) and click on other advertisements
such as newsletters or affiliate advertisements (x
other
ist
=
-4.79) significantly decreases consumers conversion
probability, whereas we find that additional search,
price, or retargeting clicks do not decrease the con-
version probability significantly and are not as strong
as the other two types. It seems that this group of
consumersis either very goal-oriented (i.e., these con-
sumers purchase immediately after they have clicked
on an advertisement or directly visited the website of
the companyin question) or is insecure in their buying
process because each additional click decreases their
conversion probability.
As expected, the probability of consumers to
purchase decreases if they have purchased in the last
period (Con
is(t−1)
= -3.70). The cumulative number
of conversions within a current session is not signif-
icant and therefore does not affect further purchases
(see x
con
is(t−1)
). Contrary to expectations, both on-site
time and on-site clicks have no significant influence
on the conversion probability of consumers (see
x
Onsite-clicks
ist
and x
Onsite-time
ist
). Notably, information
about brand-related activities does not influence
consumers’ conversion probabilities significantly
(see x
Brand
ist
and Brand
ist
). This result is surprising
because we expected an increase in brand-related
activities to reflect consumers’ intentions to purchase
from the company.
Group 2, segment size of 9%
We now focus on the intrasession effects on con-
sumer probability to purchase in the smaller segment
represented by the second mixture component(see the
right-hand side of Table 4).
For this small segment of consumers, we do not
find the cumulative number of clicks on advertise-
ments within the current session to significantly de-
crease conversion probabilities; we even find that
each additional search with subsequent clicks signif-
icantly increases such consumers’ probability to pur-
chase (x
search
ist
= 3.76). Therefore, compared to the first
segment of consumers, we uncover important differ-
ences in the effect of advertisements on consumer’s
conversion probabilities.
Furthermore, both variables accounting for past
conversions within consumers’ current session in-
dicate a very strongly negative influence on future
subsequent conversions (see x
con
is(t−1)
and Con
is(t−1)
).
As for the first segment, neither on-site clicks/time
nor brand-related information significantly influences
WhichClicksLeadtoConversions?-ModelingUser-journeysAcrossMultipleTypesofOnlineAdvertising
147

Table 4: Parameter Estimates of the Two Component Mixture Model.
Variables Group 1, segment size: 91% Group 2, segment size: 9%
Mean (95% cov. interval) Mean (95% cov. interval)
Indicator Variables I
ist
I
search
ist
-6.09 (-9.17, -3.05) -19.50 (-32.75, -6.49)
I
price
ist
-3.98 (-6.42, -1.57) -13.30 (-26.01, -0.79)
I
retargeting
ist
-4.79 (-9.00, -0.57) -12.44 (-29.65, 4.67)
I
direct
ist
-3.26 (-7.23, 0.68) -15.00 (-33.92, 3.67)
I
other
ist
-4.06 (-7.40, -0.70) -18.07 (-31.28, -5.02)
Intrasession Variables X
ist
x
search
ist
-1.27 (-4.15, 1.59) 3.76 (0.54, 6.93)
x
price
ist
-2.38 (-6.02, 1.26) 0.44 (-20.43, 21.01)
x
retargeting
ist
-1.41 (-6.52, 3.69) -6.87 (-21.61, 7.47)
x
direct
ist
-4.22 (-6.92, -1.52) 3.96 (-3.21, 11.06)
x
other
ist
-4.79 (-7.57, -2.00) 5.51 (-0.57, 11.57)
x
con
is(t−1)
0.38 (-2.66, 3.41) -31.76 (-49.56, -14.23)
Con
is(t−1)
-3.70 (-6.38, -1.00) -12.58 (-24.62, -1.07)
TLCon
ist
(logarithm hour) -0.85 (-3.73, 2.13) -10.67 (-27.27, 5.24)
x
Onsite-clicks
ist
-1.05 (-2.95, 0.83) 0.46 (-1.61, 2.62)
x
Onsite-time
ist
(logarithm hour) -1.99 (-5.09, 1.02) 0.08 (-2.95, 3.16)
x
Brand
ist
-1.94 (-6.35, 2.40) 0.86 (-3.31, 5.34)
Brand
ist
0.45 (-2.49, 3.40) 1.05 (-5.87, 8.05)
Intersession Variables Y
is
y
search
is
-2.45 (-6.82, 1.91) 3.00 (-0.60, 6.63)
y
price
is
-3.76 (-8.73, 1.20) 4.83 (-2.19, 11.76)
y
retargeting
is
-1.20 (-6.41, 3.99) 5.71 (-2.24, 13.74)
y
direct
is
0.93 (-3.04, 4.93) 6.08 (2.03, 10.31)
y
other
is
1.83 (-1.66, 5.37) -0.25 (-4.10, 3.64)
y
con
i(s−1)
2.02 (-1.29, 5.33) -19.27 (-45.00, 5.99)
y
Onsite-clicks
is
-1.57 (-3.74, 0.62) -0.11 (-2.13, 1.85)
y
Onsite-time
is
(logarithm hour) -0.67 (-3.39, 2.02) 0.42 (-2.42, 3.26)
y
Brand
is
-1.45 (-4.59, 1.67) -1.40 (-8.56, 5.78)
IST
is
(logarithm hour) -0.99 (-3.46, 1.47) 0.19 (-3.55, 3.99)
Session
is
-3.97 (-8.23, 0.23) -1.58 (-6.30, 3.06)
Notes: We report the mean and the 95% coverage interval of the parameter estimates of our proposed model using two-
mixture-component model. The estimates in boldface are significant as they lie in the 95% coverage interval.
consumers’ conversion probabilities.
Intersession Effects
Group 1, segment size of 91%
We now focus on the long-term effects of clicks
on consumers’ conversion probabilities. By contrast
to prior research that analyzed the long-term effects
of repeated display advertisement exposures on
consumers and found them to have a significantly
positive influence on consumer click probabilities
(Chatterjee et al., 2003), we do not find intersession
clicks on any type of advertisement to significantly
influence conversion probabilities for the large
segment of consumers. There do not appear to
be any long-term effects of clicks on conversion
probabilities for this group. On the one hand, this
is extremely helpful information for the company
in question because each consumer nearly has an
identical conversion probability when entering a new
session. On the other hand, if past clicks on online
advertisements do not seem to affect consumer click
probabilities, this information may lead managers
to question what the long-term success rate of past
advertising actually was.
As expected, there is no long-term negative
influence of past conversions on future conversions
(see y
con
i(s−1)
). Furthermore, as for the short-term,
there are no long-term effects of on-site clicks/time
and brand-related activities on consumers’ conver-
ICE-B2013-InternationalConferenceone-Business
148

sion probabilities (see y
Onsite-clicks
is
, y
Onsite-time
is
, and
y
Brand
is
). Although the parameter estimates are barely
not significant, an increase in the number of total
sessions reduces consumers conversion probabilities
(Session
is
= -3.97).
Group 2, segment size of 9%
Compared to the first consumer segment, we again
find inherent differences in the effects of clicks on
conversion probabilities for the second consumer seg-
ment. For example, four out of five click types have
a long-term positive influence influences the proba-
bility that they will purchase. For this small group
of consumers, each search, price, retargeting, and
direct click positively influences the probability that
they will purchase.
5
If we compare the influence of
each click-type, we find that the cumulative number
of search clicks does not have that strong of an ef-
fect on conversion probabilities as retargeting or di-
rect clicks, for example (see y
search
is
, y
retargeting
is
, and
y
direct
is
). We consider these differences in the next sec-
tion in formulating managerial implications.
Similar to the intrasession findings, the influence
of conversions in prior sessions is negative on futures
conversions (y
con
i(s−1)
= -19.27). As with consumers in
the first segment, the variables accounting for on-site
clicks/time, brand-related activities and intersession
time may be neglected (see y
Onsite-clicks
is
, y
Onsite-time
is
,
y
Brand
is
, and IST
is
).
Advertising-specific Conversion Proneness
Group 1, segment size of 91%
Previous research indicates that the click
proneness of consumers in response to display
advertisement exposure is minimal (Chatterjee
et al., 2003)), and our findings also confirm this
for advertising-specific conversion proneness. For
the large segment of consumers, we find that the
initial conversion probability is very small across
all types of advertisement. Consumer proneness to
purchase is smallest after consumers have clicked on
companies’ links through search engines (I
search
ist
=
-6.09). We find the highest probability of purchasing
after consumers directly visited the companies
website (I
direct
ist
= -3.26) or after they clicked on
comparison-shopping sites (I
price
ist
= -3.98). This is
not surprising because these two click types typically
accompany a high level of intention to purchase
5
Please note that only y
direct
is
is significant, whereas the
others are barely not significant.
(e.g., after comparing prices for a specific product or
by directly visiting the company website to purchase).
Group 2, segment size of 9%
By contrast to the first segment that includes con-
sumers with a low initial conversion probability that
decreases with subsequent clicks, we find that the
conversion probability of the second and smaller seg-
ment of consumers increases with subsequent clicks.
By contrast to the first segment, consumers of the sec-
ond segment show an even smaller initial probability
to purchase. All five indicator variables accounting
for the type of advertising show a much stronger ini-
tial and negative effect (see I
search
ist
...I
other
ist
). Our prior
findings on the intrasession and intersession effects
also have shown that this low conversion probability
may be increased with subsequent clicks.
5 IMPLICATIONS
5.1 Real-time Bidding
The knowledge of a consumer’s individual conversion
probability and the respective degree of advertising-
specific influence is vital to the relatively new and
emerging field of real-time bidding (RTB) settings
in which advertising is bought and displayed in real
time on an individual consumer level. RTB provides a
flexible option of matching individual consumers with
suitable advertising content. Within milliseconds, ad-
vertisers place bids for individual advertisement im-
pressions in an auction-based process (Way, 2012).
Assume that the company investigated here is ac-
tive in a RTB setting and dynamically exposes con-
sumers to display advertising. By accounting for con-
sumers’ complete user-journey and applying our pa-
rameter estimates, the company may be able to de-
liver display ads specifically to those consumers who
have a higher probability of purchasing compared
with consumers with a significantly lower probability
of purchasing (i.e., consumers who clicked multiple
times but never purchased).
5.2 Alternative Evaluation
of Advertising Channels
In daily business, there are more- or less-simple
heuristics to evaluate the success of multiple types
of online advertising on an individual user-level
(uniquedigital, 2012). In Table 5, we illustrate an al-
ternative approach of evaluating the profitability of
WhichClicksLeadtoConversions?-ModelingUser-journeysAcrossMultipleTypesofOnlineAdvertising
149

Table 5: Alternative Approach of Evaluating Profitability of Advertising Channels.
Mean Estimates Mean of Variables
Channel α
i
β
i
γ
i
I
ist
X
ist
Y
is
ASP (prop.) CVR (prop.)
search -7.30 -0.81 -1.96 0.26 0.50 0.91 1.65% (5.43%) 1.65% (39.70%)
price -4.82 -2.13 -2.98 0.08 1.23 1.71 0.03% (0.03%) 0.96% (7.25%)
retargeting -5.48 -1.90 -0.58 0.06 0.09 0.31 33.64% (25.46%) 1.09% (6.04%)
direct -4.32 -3.48 1.39 0.07 0.12 0.64 54.23% (51.80%) 4.86% (33.99%)
other -5.32 -3.86 1.64 0.03 0.06 0.19 48.01% (17.28%) 4.94% (13.02%)
Notes: We develop an alternative approach to evaluate the profitability of multiple advertising channels on the basis of
individual user-level data and compare the average success probability (ASP) with the conversion rate (CVR). To calculate
the proportional ASP (see the italicized number in the brackets), we take into account the total number of advertisement-
specific clicks of the complete campaign/investigation period.
multiple advertising channels and contrast that ap-
proach with the conversion rate (CVR) of each chan-
nel type. We will develop that approach simplisti-
cally below because we are not focused on determin-
ing an exact success rate for each advertising chan-
nel. Instead we want to illustrate how it is generally
possible to account for consumers’ individual user-
journeys and to evaluate multiple advertising channels
in a more realistic way than companies currently are
able. By so doing, we propose an alternative evalu-
ation metric, the average success probability (ASP).
The ASP can be interpreted as an advertisement- and
campaign-specific contribution to companies’ success
probabilities because we account for both the short-
and long-term advertising effects of the complete in-
vestigation period.
We begin the evaluation of each advertising chan-
nel by taking into account the mean (proportional for
both mixture segments) of each advertising-specific
parameter. For example, to calculate the mean esti-
mates for the search channel, we take the sum of the
mean parameter estimates for the respective parame-
ter (I
search
ist
, x
search
ist
, and y
search
is
) times the segment size
(91% and 9%, respectively) over both segments:
α
i
= 0.91∗ (−6.09) + 0.09∗ (−19.50) = −7.30
β
i
= 0.91∗ (−1.27) + 0.09∗ 3.76 = −0.81
γ
i
= 0.91∗ (−2.45) + 0.09∗ 3.00 = −1.96
As illustrated in Table 5, a first simplification
is that we only take into account those parameters
that measure the advertising-specific influence di-
rectly. That is, we take the advertising-specific in-
tercepts (I
search
ist
, ..., I
other
ist
), the cumulative number of
advertising-specific clicks within consumers’ current
sessions (x
search
ist
, ..., x
other
ist
), and the cumulative num-
ber of advertising-specific clicks across all sessions
(y
search
is
, ..., y
other
is
).
6
6
Please note that because of simplifications and for con-
venience, we do not distinguish between significant and
non-significant parameter estimates.
We take the sum of the mean parameter estimates
multiplied by the mean of the applicable variables of
the dataset that are reported in Table 1 and Table 5, re-
spectively. The logits of each advertisement-specific
product are reported as the average success probabil-
ity (ASP) in Table 5:
7
ASP
search
=
exp(−4.09)
1+ exp(−4.09)
ASP
search
= 1.65%
The ASPs differ significantly across multiple
channel types. For example, the price channel per-
forms very poorly compared to the retargeting chan-
nel. In analyzing the complete investigation period
and evaluating the proportionalsuccess of each adver-
tising channel, we take into account the total number
of respective clicks (as reported in Table 1) to calcu-
late the proportional ASP (see italicized number in
brackets).
The standard CVRs and their proportional success
rate within the investigation period are reported on the
right-hand side of Table 5. The ASPs calculated here
offer the first advice on how future advertising spend-
ing should be allocated. Compared to the CVRs, we
find that the search channel is overestimated (5.43%
vs. 39.70%) and that the retargeting channel is un-
derestimated (25.46% vs. 6.04%). These differences
may be ascribed to the circumstance that consumers
seem to be attracted by display advertisements—such
as retargeted advertisements—at the early stages of
their search-and-decision process before they conduct
later searches and click on links of the company. In
a test-setting, the company might increase its retar-
geting advertising spending chargeable to the search
channel and then analyze the campaign profitability.
7
Note that we multiply the mean parameter estimates
with the mean of the respective variables: −4.09 = −0.26∗
7.30− 0.50∗ 0.81− 0.91∗ 1.96
ICE-B2013-InternationalConferenceone-Business
150

6 CONCLUSIONS
We develop a binary logit model with a Bayesian
mixture approach to model consumer clickstreams
across multiple types of online advertising and ana-
lyze the individual conversion probabilities of con-
sumers. The mixture approach we utilize outperforms
the standard normal approach and is useful for con-
sidering heterogeneity in the individual propensity of
consumers to purchase; for the majority of consumers
(more than 90%), repeated clicks on advertisements
decrease their probability of purchasing. Thus, for
this segment of consumers, the probability of pur-
chasing is highest after consumers’ first click on an
advertisement. In contrast with this segment, we find
a smaller segment of consumers (nearly 10%) whose
clicks on advertisements increase conversion proba-
bilities.
We successfully demonstrate how to simultane-
ously integrate and evaluate multiple types of online
advertising to gain knowledge that is indispensable to
allocating financial resources. The evaluation of con-
sumers on an individual level along their complete
user-journeyis essential to optimize the auction-based
process, particularly in the emerging new advertising
technology of real-time bidding.
Furthermore, we are able to show inherent differ-
ences in the effects of consumer clicks on purchas-
ing probabilities across multiple advertising channels.
Therefore, on the basis of our parameter estimates, we
develop an alternative approach of accounting for the
success of advertising channels—the average success
probability (ASP)—which may be interpreted as an
advertisement- and campaign-specific contribution to
companies’ success probabilities as we take into ac-
count both short- and long-term advertising effects of
the complete investigation period. Compared to stan-
dardized advertisement-specific conversion rates, we
find the “search” advertising channel to be overesti-
mated and the “retargeting” channel to be underesti-
mated.
In this paper, we analyze a large dataset contain-
ing detailed individual consumer-level information.
Tracking individual consumers across multiple online
advertising types is accomplished by the application
of cookies that are stored on the personal computer of
each consumer. Thus, we do not have combined in-
formation regarding consumer usage of web browsers
across multiple devices (i.e., personal computers at
work versus at home) and are thus unable to model
complete sessions for all consumers. Furthermore,
modern web browsers give consumers the opportu-
nity to deny websites access their personal computers
to store cookies.
Another limitation of this work is that we do not have
any information on consumers’ isolated exposures to
online advertisements that have not been clicked. The
long-term effects of unclicked online advertisements
are thought to be positive on conversions (Yoon and
Lee, 2007) but have not been analyzed yet on an indi-
vidual user-level while analyzing multiple online ad-
vertising channels. We leave this question open for
future research.
Further, there is room for research analyzing the
effects of specific online advertisements on consumer
online behavior. For example, the integration of
consumer-specific information, such as gender or
interests, might uncover further insights into con-
sumers’ individual online click and conversion prob-
abilities. We also see the combination of aggregated
data (that does not suffer from the cookie-deleting
problem) with consumer-level data (as it is analyzed
here) as an important and interesting topic for future
research.
REFERENCES
Abhishek, V., Hosanagar, K., and Fader, P. S. (2011). On
aggregation bias in sponsored search data: Existence
and implications. http://ssrn.com/abstract=1490169.
Allenby, G. M. and Ginter, J. L. (1995). Using extremes
to design products and segment markets. Journal of
Marketing Research, 32(4):392–403.
Ansari, A., Mela, C. F., and Neslin, S. A. (2008). Customer
channel migration. Journal of Marketing Research,
45(1):60–76.
Athey, S. and Ellison, G. (2009). Position auctions with
consumer search. http://ssrn.com/abstract=1454986.
Bass, F. M., Bruce, N., Majumdar, S., and Murthi, B.
P. S. (2007). Wearout effects of different adver-
tising themes: A dynamic bayesian model of the
advertising-sales relationship. Marketing Science,
26(2):179–195.
Broder, A. (2002). A taxonomy of web search. SIGIR Fo-
rum, 36(2):3–10.
Chatterjee, P., Hoffman, D. L., and Novak, T. P. (2003).
Modeling the clickstream: Implications for web-based
advertising efforts. Marketing Science, 22(4):520–
541.
Danaher, P. J. and Mullarkey, G. (2003). Factors affecting
online advertising recall: A study of students. Journal
of Advertising Research, 43(3):252–267.
Deleersnyder, B., Geyskens, I., Gielens, K., and Dekimpe,
M. G. (2002). How cannibalistic is the internet chan-
nel? a study of the newspaper industry in the united
kingdom and the netherlands. International Journal
of Research in Marketing, 19(4):337–348.
Dinner, I. M., van Heerde, H. J., and Neslin, S. A.
(2011). Driving online and offline sales: The cross-
channel effects of digital versus traditional advertis-
ing. http://ssrn.com/ abstract=1955653.
WhichClicksLeadtoConversions?-ModelingUser-journeysAcrossMultipleTypesofOnlineAdvertising
151

Ghose, A. and Yang, S. (2009). An empirical analysis of
search engine advertising: Sponsored search in elec-
tronic markets. Management Science, 55(10):1605–
1622.
Goldfarb, A. and Tucker, C. (2011). Standardization,
standards and online advertising. http://ssrn.com/
abstract=1745645.
Ilfeld, J. S. and Winer, R. S. (2002). Generating website
traffic. Journal of Advertising Research, 42(5):49–61.
Lenk, P. J., DeSarbo, W. S., Green, P. E., and Young, M. R.
(1996). Hierarchical bayes conjoint analysis: Recov-
ery of partworth heterogeneity from reduced experi-
mental designs. Marketing Science, 15(2):173–191.
Naik, P. A. and Peters, K. (2009). A hierarchical mar-
keting communications model of online and offline
media synergies. Journal of Interactive Marketing,
23(4):288–299.
Naik, P. A. and Raman, K. (2003). Understanding the im-
pact of synergy in multimedia communications. Jour-
nal of Marketing Research, 40(4):375–388.
Neslin, S. A., Grewal, D., Leghorn, R., Shankar, V., Teer-
ling, M. L., Thomas, J. S., and Verhoef, P. C. (2006).
Challenges and opportunities in multichannel cus-
tomer management. Journal of Service Research,
9(2):95–112.
Neslin, S. A. and Shankar, V. (2009). Key issues in mul-
tichannel customer management: Current knowledge
and future directions. Journal of Interactive Market-
ing, 23(1):70–81.
Nottorf, F. (2013). Modeling the clickstream across mul-
tiple online advertising channels using a binary logit
with bayesian mixture of normals. Working Paper.
Nottorf, F. and Funk, B. (2013a). A cross-industry analy-
sis of the spillover effect in paid search advertising.
Electronic Markets, Article in Advance.
Nottorf, F. and Funk, B. (2013b). The Economic Value
of Clickstream Data From An Advertiser’s Perspec-
tive. In Proceedings of the 21th European Conference
on Information Systems (ECIS), Utrecht, Netherlands
2013.
Rossi, P. E., Allenby, G. M., and McCulloch, R. E. (2005).
Bayesian statistics and marketing. Wiley, Hoboken,
NJ.
Rutz, O. J. and Bucklin, R. E. (2011). Does banner adver-
tising affect browsing for brands? clickstream choice
model says yes, for some. Quantitative Marketing and
Economics, pages 1–27.
Rutz, O. J., Bucklin, R. E., and Sonnier, G. P. (2012). A la-
tent instrumental variables approach to modeling key-
word conversion in paid search advertising. Journal
of Marketing Research, 49(3):306–319.
Rutz, O. J., Trusov, M., and Bucklin, R. E. (2011). Model-
ing indirect effects of paid search advertising: Which
keywords lead to more future visits? Marketing Sci-
ence, 30(4):646–665.
uniquedigital (2012). Cross-channel management. http://
www.uniquedigital.de/downloads/whitepaper-unique-
cross-channel-management.
Way, H. (2012). Real-time bidding: Greasing the wheels of
digital advertising. http://www.ecommercetimes.com/
story/76000.html.
Wiesel, T., Pauwels, K., and Arts, J. (2011). Marketing’s
profit impact: Quantifying online and off-line funnel
progression. Marketing Science, 30(4):604–611.
Yoon, H.-S. and Lee, D.-H. (2007). The exposure effect of
unclicked banner advertisements. Advances in Inter-
national Marketing, 7(18):211–229.
ICE-B2013-InternationalConferenceone-Business
152