The User-journey in Online Search

An Empirical Study of the Generic-to-Branded Spillover Effect

based on User-level Data

Florian Nottorf, Andreas Mastel and Burkhardt Funk

Institut f¨ur Elektronische Gesch¨aftsprozesse, Leuphana University, L¨uneburg, Germany

Keywords:

Online Search, Online Advertising, Consumer Behavior, Query Log, Spillover.

Abstract:

Traditional metrics in online advertising such as the click-through rate often take into account the users’ search

activities separately and do not consider any interactions between them. In understanding online search be-

havior, this fact may favor a certain group of search type and, therefore, may mislead managers in allocating

their ﬁnancial spending efﬁciently. We analyzed a large query log for the occurrence of user-speciﬁc interac-

tion patterns within and across three different industries (clothing, healthcare, hotel) and were able to show

that users’ online search behavior is indeed a multi-stage process, whereas e.g. a product search for sneakers

typically begins with general, often referred to as generic, keywords which becomes narrowed as it proceeds

by including more speciﬁc, e.g. brand-related (“sneakers adidas”), keywords. Our method to analyze the de-

velopment of users’ search process within query logs helps managers to identify the role of speciﬁc activities

within a respective industry and to allocate their ﬁnancial spending in paid search advertising accordingly.

1 INTRODUCTION

Selling advertising linked to user-generated queries,

the so-called sponsored search has become a criti-

cal component of companies marketing campaigns

(Ghose and Yang, 2008). Although more than half of

all search processes by individual users consist only

of one query (Jansen and Mullen, 2008), consumers

that have a transactional intention often do not reach

their goals by conducting only a single search (Search

Engine Watch, 2006). An aspect also conﬁrmed by

(Rutz and Bucklin, 2011). They demonstrate the so-

called “spillover effect” from generic to brand-related

searches for a company in the hospitality industry:

the generic search (“hotel”) and the corresponding

advertisements by the hotel chain in question signiﬁ-

cantly contributed to the fact that users later turned to

brand-related searches for this hotel chain (e.g. “ho-

tel hilton”) and ﬁnally to corresponding reservations.

Their initial work clariﬁes how traditional metrics

in online advertising such as the click-through rate

(CTR) are alone no adequate tools to control for paid

search advertising campaigns: they only take into ac-

count the users’ search activities singularly and do not

consider any interactions between them.

Our analysis is based on a complete query log

published by AOL in 2006 (Pass et al., 2006) and ex-

plains users’ search activities in more behavioral de-

tail. Unlike (Rutz and Bucklin, 2011), who used

keyword-level data aggregated on a daily basis of a

paid search advertising campaign of a single com-

pany, we analyzed users’ individual queries within

and across entire industries in order to determine

whether the resulting user-journey showed behavior

indicating spillover effects. As evidence of such a

spillover we regarded a user-journey that, for exam-

ple, started with a generic and was followed by a

brand-related search. We investigated and conﬁrmed

the spillover effect and its occurrence in users’ online

search behavior for several industries and, by doing

so, highlighted the role of generic activities as gate-

keepers for companies’ online advertising. To the

best of our knowledge, our work is the ﬁrst to ana-

lyze a complete query log for the occurrence of the

spillover effect and thus makes a contribution to re-

search on consumers’ online behavior. In addition, an

improved understanding of the role of generic activ-

ities and of how consumers actually search for prod-

ucts and brands to satisfy their needs will help ad-

vertisers to allocate their budget on online advertising

more efﬁciently.

The paper is structured as follows: ﬁrst, we will

review existing work on consumer online search be-

havior in general as well as in the speciﬁc context of

145

Nottorf F., Mastel A. and Funk B..

The User-journey in Online Search - An Empirical Study of the Generic-to-Branded Spillover Effect based on User-level Data.

DOI: 10.5220/0004052101450154

In Proceedings of the International Conference on Data Communication Networking, e-Business and Optical Communication Systems (ICE-B-2012),

pages 145-154

ISBN: 978-989-8565-23-5

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

paid search advertising. In the next chapters we will

describe our method of analyzing spillover effects in

query logs and introduce our dataset together with the

ﬁltering procedures applied. Next, we focus on the re-

sults of spillover effects across several industries. The

last sections contain a discussion of our ﬁndings and

will close this paper by mentioning the limitations of

our study and by giving suggestions for further inves-

tigations.

2 RELATED WORK

The detailed records of users’ Internet activities

opened up the possibility of analyzing a variety of

topics, such as consumers’ online search behavior

(see, for example, (Bucklin and Sismeiro, 2009) for a

review and discussion of strengths and limitations of

clickstream data for marketing research). General re-

search classiﬁes consumers’ online searches into nav-

igational, transactional, and informational purposes

(Broder, 2002; Jansen and Spink, 2007). According

to (Moe, 2003), searches heavily depend on the in-

dividual purchase intent such as involvement: while

directed-buying sessions present very narrowly aimed

shopping behavior, consumers with low purchase in-

tention exhibit much broader search patterns for un-

speciﬁc products. (Johnson et al., 2004) conﬁrm that

the depth of consumer search is generally low and

shows no increase with a consumer’s growing expe-

rience. This aspect is conﬁrmed by analyses showing

that the CTR on a search engine’s (sponsored) link

decreases with its position (Agarwal et al., 2011; Ani-

mesh et al., 2011; Ghose and Yang, 2009; Rutz et al.,

2011). Despite the amount of literature focusing on

users’ search behavior no work has yet considered

possible differences of this behavior across industries

- an aspect considered and analyzed in more detail in

the present paper.

The empirical analysis of sponsored search has

only recently begun to become the focus of the sci-

entiﬁc community. (Ghose and Yang, 2009) exam-

ine the general impact of paid search advertising on

measures such as the relationship between the type

and length of keywords and different variables on

consumers’ click and conversion behavior. Although

the authors uncover important differences in the click

and purchasing intensity regarding a speciﬁc group

of keywords, they miss to account for users’ interac-

tions between them. (Chan et al., 2011) build an in-

tegrated model of customer lifetime, transaction rate,

and gross margin accounting for spillovers from spon-

sored search on customer acquisition and behavior in

ofﬂine channels. Their results indicate that customers

who were originally acquired through paid search ad-

vertising on Google have a signiﬁcant higher life-

time value (about 20%) than customers acquired from

other channels. In a recent study, (Rutz et al., 2011)

criticize the common assumption that users respond

homogeneously to keywords. The authors formulate

a consumer-level approach to especially evaluate tex-

tual properties of paid search ads on consumers’ re-

sponses and account in this way for heterogeneity.

Besides several ﬁndings referring to the consumer-

integrated focus, the authors conﬁrm that keyword-

speciﬁc factors, like the distinction between broad

and narrow purposes, are important when linking

searches to a CTR (see also (Ghose and Yang, 2009)).

This stream of research typically uses aggregated-

level data. Although the majority of papers discussed

above claim a more behavioral focus on how users re-

spond and act in the context of search engines, they

only scratch the surface of an analysis of consumers’

actual online search behavior (see (Abhishek et al.,

2011) for discussion of aggregation bias in sponsored

search data).

This paper is most closely related to (Rutz and

Bucklin, 2011), who are able to demonstrate the

spillover effect from generic to brand-related searches

for a company in the hospitality industry. The generic

search (“hotel”) and the corresponding ads of the ad-

vertising hotel chain signiﬁcantly contribute to the

fact that users later turn to brand-related searches for

this hotel chain (e.g. “hotel hilton”) and ﬁnally to

corresponding reservations. We will explore the dif-

ferences between Rutz and Bucklin’s method and our

approach in the following chapter more closely.

3 DATA AND METHODOLOGY

Our analysis is based on a log published by AOL

which consists of over 35 million queries from about

650,000 users over a three month period (March to

May) (Pass et al., 2006).

Although this dataset dates

back to 2006, it is still an unique and comprehensive

query log containing extraordinary information about

users’ search and click behavior, which may be found

in todays search engines like Google or Bing.

In terms of analyzing user-journeys for speciﬁc

behavioral aspects, such as the spillover effect from

generic to brand-related searches, we, ﬁrst, needed

to deﬁne industries and, second, companies within

Since the users who represent the queries are mostly

located in the United States of America, our work is mainly

based on the US region.

We would like to thank an anonymous reviewer for

pointing out this more clearly.

ICE-B2012-InternationalConferenceone-Business

146

these. It was only then that we were able to cat-

egorize user queries into generic and brand-related

types of searches respectively and to make a deﬁnite

statement about both the existence and the extent of

industry-speciﬁc interaction effects between generic

and brand-related search activities.

We will explain this process in more detail in the

following chapters. Please ﬁnd all additional data and

information such as the list of selected companies,

keywords, or our ﬁnal ﬁltered query log on the au-

thor’s website: http://www.nottorf.org.

3.1 The Initial Query Log

The log includes 36,389,567 records structured in ﬁve

columns (see Table 1 for a short excerpt of the data):

• AnonID: A unique identiﬁcation number of an

anonymized user.

• Query: The user’s query.

• QueryTime: The point of time at which the query

was submitted for search.

• ItemRank: In case a user clicked on one of the

result pages, the “ItemRank” shows the site’s po-

sition in the result pages. If no page was selected,

this ﬁeld remained empty.

• ClickURL: Shows the URL of the clicked result

(Pass et al., 2006).

We focused on three industries: the hotel and

hospitality industry to make our ﬁndings compara-

ble with those of (Rutz and Bucklin, 2011); the

clothing industry as a representative of nondurable

goods with the inﬂuence of brand strength being as-

sumed to be strong; and the healthcare industry as

a representative of the insurance sector. We further

restricted our analysis to the top ten companies in

2006, their ranking being based on revenue and brand

strength within a certain industry (Interbrand & Busi-

ness Week, 2006). We did so on the assumption

that these companies represented the major number of

possible brand-related search activities within a spe-

ciﬁc industry. The consequences of this approach will

be discussed later, for example of the fact that we

did not consider the total number of brand-related or

generic search activities recorded in the query log.

On the basis of the industries selected and each of

the ten companies we ﬁltered out generic and brand-

related queries from the initial query log.

3.2 The Filtering Process for Queries

For each of the ten companies within each industry we

deﬁned brand-related keywords to analyze a spe-

ciﬁc query for their occurrence and, as the case may

Figure 1: Structure of the ﬁltering process of brand-related

keywords.

be, marked the query as a branded search. Because

of the fact that many of the selected companies have

subsidiaries, we had to deﬁne a set of keywords that

were related to the parent-company.

We organized

the brand names and keywords in hierarchical order as

shown in Figure 1. Based on this structure we applied

tools (e.g. the online toolset given by Google Ad-

Words) to identify keywords that are often searched

for in the context of the brand names mentioned.

We started with the company names (e.g. “Nike

Inc.”), which were placed in the root section of this

structure. Next to it we put brand and product names

which we derived from the company’s information it-

self (e.g. “NIKEDiD” or “Air Jordan”) and from the

help of the keyword tools mentioned above. On the

basis of these subcategories was deﬁned the ﬁnal set

of keywords (e.g. “nikeid”, “nikid”, or “jordans”) for

our analysis (see the right column of Figure 1). In this

section were also taken into account variant names

and typing errors. However, we deﬁned 120 brand-

related keywords for the healthcare, 228 for the hotel,

and 720 for the clothing industry.

Deﬁning the set of generic keywords required us

to limit which queries could directly be related to the

mentioned industries. Therefore, we may have had

to consider those queries that had no direct relation-

ship to a speciﬁc industry and its companies or prod-

ucts, but still were able to generate clicks on links to

these companies. For example, the analysis of the ho-

tel industry raised the question whether the search for

a country or a ﬂight could have already be seen as a

generic query or not. It is possible that a selection of a

ﬁnal dataset that is too comprehensivemay also cover

users who do not have the intention to purchase some-

thing at all. This is a problem every paid search adver-

tising campaign has to face to some extent, since ad-

vertisers want to become displayed and ranked when

The Harrah’s Entertainment Inc., for example, had

about 21 hotels and hotel chains in 2006.

TheUser-journeyinOnlineSearch-AnEmpiricalStudyoftheGeneric-to-BrandedSpilloverEffectbasedonUser-level

Data

147

Table 1: Extract from the AOL dataset.

AnonID Query QueryTime ItemRank ClickURL

1927 does bcbs cover ci 03.05.2006 00:24

1927 does bcbs ﬂ cover ci 03.05.2006 00:25

7117 www.anthem.com 09.05.2006 09:33 3 http://www.maine.nea.org

7117 www.anthem.com 09.05.2006 09:33 4 http://hr.nd.edu

7117 www.anthem.com 10.04.2006 06:57 1 http://www.myuhc.com

a possible consumer searches for products or services

that the respective company offers. The consequences

of a potentially incorrect or incomplete selection will

be discussed later.

We proceeded from the advertiser’s point of view

and deﬁned generic keywords that reﬂected the in-

tention to gather information about the product or to

purchase it. Hence, we restricted our set of generic

keywords to directly product-related terms, their syn-

onyms and variations. To achieve this, we also made

use of the keyword tools referred to above but mostly

we derived the keywords manually by gathering gen-

eral information from the companies’ websites. Thus,

we deﬁned 196 generic keywords for the healthcare

(e.g. “health care”, “dental insurance”, “medicare”),

335 for the hotel (e.g. “hotel”, “motel”, “suites”), and

197 for the clothing industry (e.g. “shoes”, “shorts”,

“underwear”).

On the basis of our brand-related and generic key-

words we ﬁltered and categorized the query log

records (see Figure 2). First, all log records whose

“Query”-columns contained brand-related keywords

were moved into an industry-speciﬁc table and were

marked as brand-related queries. This ensured that a

record with both brand-related and generic keywords

appearing in the same query was not duplicated. The

second step was to ﬁlter the reduced dataset (with-

out brand-related queries) on the basis of generic

keywords, which were moved to the three industry-

speciﬁc tables with only generic queries. Ideally, the

remaining dataset should no longer have contained

any relevant queries. It is important to mention that

our procedure also ﬁltered out log records that were

not related to the industries selected. This was due

to the fact that some queries occur also within an ir-

relevant context. We handled this problem by manu-

ally scanning the ﬁltered data and by deleting records

giving information about an irrelevant context (e.g.

queries in a pornographic context).

The records not only contain information about

a user’s query itself but also the URL of a clicked

result. We distinguished between URLs on pages

that were related to the ten companies deﬁned

above and between URLs that were not, such as

websites of retailers (e.g. “www.ebay.com” or

“www.amazon.com”). Thus, we were able to sepa-

rate relevant clicks (from the perspective of the ten

companies) from irrelevant ones and were in a posi-

tion to make more precise statements about the effects

of generic and/or brand-related searches. To deter-

mine those (relevant) clicks we analyzed the informa-

tion contained in the “ClickURL”-column of our ﬁ-

nal dataset and manually checked whether the clicked

pages were related to the companies examined. Fol-

lowing this procedure, we identiﬁed 244 websites for

the clothing, 298 for the healthcare, and 838 for the

hotel industry.

The total number of unique users, of impressions,

and of (relevant) clicks in our ﬁnal dataset for each

industry following our ﬁltering procedure are shown

in Table 2. The resulting descriptive statistics con-

ﬁrm strong differences in efﬁciency between generic

and brand-related keywords. Although we ﬁltered out

Table 2: Descriptive statistics on the ﬁnal dataset.

Hotel Users Imp. Clicks CTR

Generic 71,405 327,563 3,801 1.16%

Branded 32,039 97,070 21,154 21.79%

Total 84,408 424,633 24,955 5.88%

Clothing Users Imp. Clicks CTR

Generic 51,105 293,645 350 0.12%

Branded 14,225 46,586 5,695 12.22%

Total 58,166 340,231 6,045 1.78%

Healthcare Users Imp. Clicks CTR

Generic 10,876 33,115 719 2.17%

Branded 9,304 20,516 7,896 38.49%

Total 18,255 53,631 8,615 16.01%

far more generic than brand-related queries from the

initial query log, the number of clicks in response

to branded searches was much higher. This fact re-

sults in quite different CTRs as already shown by

e.g. (Ghose and Yang, 2009; Ghose and Yang, 2010),

(Rutz and Bucklin, 2011), and (Yang and Ghose,

2010). Companies focusing only on these statis-

tics might conclude that the concentration on brand-

related keywords and the neglect of generic terms

would increase proﬁtability since the metrics for key-

ICE-B2012-InternationalConferenceone-Business

148

Figure 2: Schematic illustration of the ﬁltering procedure for a given industry.

words containing brand-speciﬁc information seem to

be more effective than for those keywords describing

generic purposes. The major problem of these types

of keyword-based analyses is, however, that they only

link users’ actions to a speciﬁc keyword one at a time

and do not consider any interactions between several

searches and clicks. We, on the other hand, aimed at

a processual analysis of the data and considered the

development of users’ search activities over time.

3.3 Indicating Spillover Behavior in

Query Logs

The development of the search process from generic

to brand-related searches may be attributed to the fact

that individual brands gain the users’ attention dur-

ing the search process. We deﬁned two levels of the

spillover effect in user-journeys:

Spillover Behavior Level 1. A user-journey shows a

generic-to-branded spillover effect, when a user ﬁrst

searches for generic and next, at any further time, for

brand-related terms.

Spillover Behavior Level 2. A user-journey shows a

generic-to-branded spillover effect, when a user ﬁrst

searches for a generic term and her last search that

leads to a (relevant) click is a brand-related one.

For each of the two deﬁnitions we assigned

all considered user-journeys to the four ﬁelds in

a 2x2 matrix (generic→branded, generic→generic,

branded→branded, branded→generic). For the

level 2 spillover effect we obviously had less user-

journeys than for level 1, since we required those

users to had at least one (relevant) click in their jour-

neys.

Our analysis is different from the one of (Rutz

and Bucklin, 2011) as we utilize user-level instead of

keyword-level data. Rutz and Bucklin use the daily

number of generic searches and clicks as indepen-

dent variables modeling a latent construct of aware-

ness which in turn affects the number of brand-related

searches. In addition, they use data from a paid search

advertising campaign of one company thus having a

clear boundary of the study.

The fact that we built our analysis on a complete

query log further enabled us to consider user-speciﬁc

behavior within a whole industry instead of analyz-

ing keyword-level data aggregated on a daily basis for

just one company. We captured all search activities

for three different industries and determined the num-

ber of users who searched for either generic or brand-

related terms only, alternatively, performed spillovers

from generic to brand-related searches and vice versa.

In addition, we were able to establish the point of time

when a user performed an action and analyzed the ex-

act time span in which possible spillover effects might

occur.

4 RESULTS

4.1 Spillover Results

The results of the spillover analysis for each industry

following the level 1 spillover deﬁnition are shown

in Table 3. It shows the proportion of all unique users

in relation to each group of user-journeys, resulting in

a 2x2 matrix for each industry.

The results indicate that the interaction effects be-

tween generic and brand-related searches differ across

industries. Take, for instance, the healthcare indus-

try. Here, 41.1% of all users conducted only brand-

related and 48.3% only generic searches. Note that

these homogeneous groups also contain user-journeys

with only a single search. The remaining 10.6% of all

users who switched either from generic to branded or

from branded to generic searches during their user-

journeys divide nearly equally into the two hybrid

groups. There is no clear sign of users’ favoring one

speciﬁc interaction direction within the healthcare in-

dustry.

The results for the clothing industry, on the other

hand, indicate at least a small spillover behavior from

generic to brand-related search activities. We found

TheUser-journeyinOnlineSearch-AnEmpiricalStudyoftheGeneric-to-BrandedSpilloverEffectbasedonUser-level

Data

149

Table 3: Proportions of ﬁrst search activities within user-

journeys per industry (spillover behavior level 1).

Hotel to generic to branded

from branded 8.6% (7,258) 20.8% (17,523)

from generic 56.7% (47,849) 13.9% (11,778)

Clothing to generic to branded

from branded 4.7% (2,720) 15.0% (8,784)

from generic 72.6% (42,218) 7.7% (4,444)

Healthcare to generic to branded

from branded 4.9% (897) 41.1% (7,509)

from generic 48.3% (8,821) 5.7% (1,028)

Note: Both the percentage and the total number (in

brackets) of all unique users within one of the

journey-groups are represented for each industry.

a large number of users searching only for generic

terms (72.6%) while a much smaller group searched

only for brand-related ones (15.0%). Although there

were no more than 12.4% of users in total conduct-

ing hybrid searches, there were 3.0% more user-

journeys starting with a generic search that was fol-

lowed by a brand-related one (7.7%) than in the op-

posite group (4.7%). Users looking for articles in

the clothing industry, such as shirts, shoes, or un-

derwear, seemed to switch more likely their type of

search from generic to brand-related terms than from

brand-related to generic ones.

An obvious sign of spillover behavior was shown

by the results for the hotel industry. Here, nearly 14%

of all users switched to brand-related searches after

they initially searched for generic terms. Although the

opposite group, starting with branded and switching

to generic terms, was also relatively large (8.6%), it

signiﬁcantly differed from the actual spillover group.

This can be seen as evidence for the fact that users’

search behavior within the hotel industry initially

started with broad and general search terms (e.g.

“hotel”, “bed and breakfast”, “suite”) which became

more (brand-)speciﬁc as the search proceeded (e.g.

“harrah”, “sheraton”, “hyatt”).

Table 4 shows the results for the level 2 spillover

effect. The focus in this spillover deﬁnition on users’

clicks on (relevant) links led to a shift in favor of

brand-related keywords. This is not surprising since

the brand-related queries received far more clicks

compared to generic ones (see Table 2). But, similar

to the differentiation into an exploratory and a goal-

directed searching mode (e.g. (Janiszewski, 1998)

and (Moe, 2003)), this alternative spillover analysis

has the ability to differentiate between users intending

to purchase something (which results in a click) and

users behaving in a less goal-directed manner (result-

ing in no click on a company’s website). We acknowl-

Table 4: Percentages of ﬁrst search activity and last search

activity before a click within user-journeys per industry

(Spillover behavior level 2).

Hotel to generic to branded

from branded 2.7% (338) 50.9% (6,428)

from generic 12.7% (1,602) 33.7% (4,206)

Clothing to generic to branded

from branded 0.9% (36) 63.9% (2,532)

from generic 6.2% (244) 29.0% (1,150)

Healthcare to generic to branded

from branded 1.0% (58) 80.0% (4,206)

from generic 9.0% (471) 10.0% (523)

Note: Both the percentages and the total number (in

brackets) of all unique users within one of the

journey-groups are represented for each industry.

edge that there might be a purchase intention even if

a user did not click on a link to a company’s website

(e.g., that a user clicks on a third party’s link, such as

Amazon or Ebay).

This analysis shows very strong evidence for

spillover behavior. See, for example, the results

for the hotel industry. Here, 33.7% of all users

that clicked on a company’s link started their user-

journeys with a generic search and switched to a

brand-related one. This means that more than one

third of these users ﬁrst looked for generic terms

before they speciﬁed their searches using brand-

related keywords and ﬁnally clicked and ended their

search-to-click-processes. The magnitude of the ef-

fect for the hotel industry becomes even more pro-

nounced, since the reverse-spillover effect (users that

ﬁrst searched for brand-related keywords before they

conducted a generic search and clicked on a com-

pany’s link) was the smallest of all four groups by

far.

Similar ﬁndings to the hotel industry can be found

for the clothing sector. Here, about 29% of all users

that clicked on a company’s link looked for brand-

related terms after they conducted generic searches.

The reverse-spillover is insigniﬁcantly small, since

less than 1% of the users ﬁrst searched for brand-

related and afterwards for generic terms before they

clicked.

Focusing on the healthcare industry, the ﬁndings

of the level 1 spillover analysis can be partly con-

ﬁrmed. Again, a large number of users seemed to

search only for brand-related or generic terms be-

fore clicking on a respective link. Nearly 90% of

all users who clicked on a company’s link did not

switch their type of search, that is they searched only

for either brand-related or generic terms. This fact

had an immediate inﬂuence on the strength of the

ICE-B2012-InternationalConferenceone-Business

150

Table 5: Time spans for each industry in which the

Spillovers occurred.

Quantiles of time difference in days

Industries 10% 25% 50% 75% 90%

Hotel

Level 1 0.01 0.10 8.95 30.13 55.47

Level 2 0.03 4.85 22.87 51.17 70.15

Clothing

Level 1 0.01 1.73 14.91 37.54 59.07

Level 2 0.04 6.86 23.12 47.86 66.70

Healthcare

Level 1 0.00 0.02 4.95 22.01 49.91

Level 2 0.01 0.72 12.07 37.26 60.90

Note: “Level 1” denotes the time differences in “Spillover

Behavior Level 1”. “Level 2” denotes the time differences

in “Spillover Behavior Level 2” accordingly.

spillover effect from generic to brand-related search:

only a relatively small number of users’ (nearly 10%)

ﬁrst searched for generic terms and next conducted

a brand-related one before they clicked on a corre-

sponding link.

We are not able to fully explain this relatively

small number of interactions between generic and

branded terms, especially when compared to our ﬁnd-

ings for the hotel and clothing industry. Still, there

are some differences in the data of the three industries

(see Table 2) that may at least give a clue to this rela-

tive lack of interactions. For example, the total num-

ber of impressions within the healthcare industry is

considerably smaller than in the two others. In partic-

ular, there are about ten times fewer generic impres-

sions. Also, the CTR of the brand-related keywords is

by far the highest. This can already be seen as an in-

dicator for a relatively small interaction-rate between

generic and brand-related searches, since these two

metrics indicate that searches of a single type alone

seem to be able to lead to a possible solution for the

user.

4.2 Time Differences between Actions

Table 5 shows the time spans and the quantiles for

each industry within which the generic-to-branded

spillover effects occurred. “Level 1” denotes the time

differences of spillover behavior from (ﬁrst) generic

to (ﬁrst) brand-related searches (see the above def-

inition of “Spillover Behavior Level 1”). The time

that elapsed between the users’ ﬁrst generic search

and last brand-related one that resulted in a click on

a company’s link is denoted by “Level 2” (“Spillover

Behavior Level 2”).

Let us ﬁrst focus on the hotel industry. The

results of the analysis of the time differences for

the ﬁrst spillover deﬁnition suggest that 25% of all

users switched from generic to brand-related searches

within only 2.4 hours. These users changed their

search behavior very quickly when compared with the

average time span of about 9 days (8.95) in which the

generic-to-branded spillover occurred. As expected,

the “Level 2”-results indicate a longer time difference

since we focused on users’ last brand-related search

that led to a click on a company’s link. Overall, our

ﬁndings are consistent with the results of (Rutz and

Bucklin, 2011) who ﬁnd that the search process for

lodging seems to be short and to occur mainly in be-

tween a couple of days to two weeks. We conﬁrm an

even shorter time difference on the basis of our results

since we found many users who switched their type of

search in between only a few hours and up to one day.

That rest of the users spread their search process over

the further investigation period, see Figure 3, seems

likely to be a coincidence.

The results for the clothing industry show some

deviations from those for the hotel industry. Although

we expected a search process for garments to change

in a shorter period of time from generic to brand-

related terms, as it does in searches for the right ho-

tel, this was not borne out by the ﬁndings for this in-

dustry. Indeed, the mean time span of the spillover

behavior was about 2 to 3 weeks. This can be ex-

plained by the possibility of users being confronted

with far more possibilities of choice compared to the

hotel industry since competition in the clothing in-

dustry is not restricted to a certain location. In other

words, a user searching for a hotel in a speciﬁc loca-

tion (e.g. “hotel barcelona”) might not have as many

choice alternatives when searching for an appropriate

room as a user looking for a speciﬁc garment (e.g.

“hoodie jacket”). This circumstance could reduce the

time users needed to spend on ﬁnding the right hotel.

Another explanation of the relatively long time differ-

ences compared to the hotel industry is that searching

for the garments might not be that selectively inten-

sive and target-oriented as looking for the right hotel.

In the healthcare industry, the time spans in which

the spillover occurred were the shortest by far. Focus-

ing on the “Level 1”-results, 25% of all users whose

journey showed spillover behavior switched the type

of search within half an hour (0.02). The results of

“Level 2” conﬁrm this short interaction period since

more than one out of four of all spillovers did not take

more than one day eventhough considering the clicks.

Following our previous chain of reasoning, this short

period between switching from one search to the other

appears reasonable: the (health) insurance sector is

easily the least fragmented in comparison with the

other two. This is demonstrated, for instance, by the

TheUser-journeyinOnlineSearch-AnEmpiricalStudyoftheGeneric-to-BrandedSpilloverEffectbasedonUser-level

Data

151

0 20 40 60 80

Hotel

Clothing

Healthcare

0 20 40 60 80

Hotel

Clothing

Healthcare

Figure 3: Time spans within which “Spillover Behavior Level 1” (Left) and “Spillover Behavior Level 2” (right) occurred.

total number of brand-related keywords that we de-

ﬁned for this industry (120) as opposed to the ho-

tel (190) and clothing industry (720) industries. This

limited number of keywords increases the probabil-

ity of users switching to a brand-related search in a

shorter period of time. Either they found the “right”

company more easily or simply were in a more goal-

directed mode compared to users searching within the

clothing industry.

All results are additionally illustrated in Fig-

ure 3, which displays the respective number of user-

journeys for each industry that showed a very short

time span in which the spillover effects occurred (see

the cumulative density at point “0 Days”). Further,

it can be seen that each time span slowly approxi-

mates the total investigation period of 90 days. The

analysis of the time difference provides evidence for

both a substantial percentage of users that switched

from generic to brand-related searches in a very short

period of time (less than one day or several hours)

and the average percentage of users that changed their

type of search over several days or weeks.

5 DISCUSSION

Our results emphasize the role of generic search ac-

tivity in the users’ search and decision process since

it is crucially important for companies and individual

brands to gain the users’ attention during the search

process and thus probably to be considered a poten-

tial solution.

By investigating online search behavior we were

able to show that traditional research vindicates and

can be applied in the online setting. Following this

literature, consumers’ decision processes can be di-

vided into at least two stages: the ﬁrst is a set of

products or brands a consumer is aware of at any

given point of time. This is referred to as the “aware-

ness set”. In the second, the so-called “evoked set”,

a consumer is likely to reduce his or her set of prod-

ucts that the ﬁnal decision is based upon (Howard and

Sheth, 1968). Therefore, the chance of a brand being

considered for purchase does not exist if that brand

is not part of a consumer’s awareness set (Narayana

and Markin, 1975). When a consumer does not know

which brand may satisfy his or her initial need (e.g.,

has no distinctive evoked or awareness set), a search-

process will more likely begin with general searches

and show a narrowing by becoming (brand-)speciﬁc

as it proceeds. We can conﬁrm this assumption within

the setting of online search, since it is indeed crucial

for companies to gain the users’ attention during the

early stage of their search and decision processes.

With respect to understanding users’ search be-

havior in general and with respect to paid search ad-

vertising in speciﬁc, we agree with the work of (Rutz

and Bucklin, 2011), since we found the generic search

activities of primary importance for the users’ search

and decision process: although generic searches re-

ceive far less clicks as brand-related ones, they are

indeed indispensable in managing online advertising

campaigns. Since there are also strong differences

in the extent of the spillover effects (e.g. hotel vs.

healthcare industry), it is recommended for each com-

pany to identify the degree of this interaction effect

between generic and brand-related activities and ad-

just their spending accordingly.

The differences of the extent of the spillover be-

havior across industries might not only result from

different market and competition set-ups, but might

also be due to different factors inﬂuencing users’

search behavior, such as involvement.

Searching for the right hotel room or for the best

health insurance may be more functionally driven

ICE-B2012-InternationalConferenceone-Business

152

than choosing the right sneakers or underwear. A

further differentiation into users with a goal-directed

searching mode and those with a low purchase inten-

tion may reveal some more insights into the extent of

the spillover effect. We focused on that circumstance

by considering clicks on companies’ websites (def-

inition Spillover Behavior Level 2) and have found

a much stronger spillover behavior in the ﬁrst group

compared to users that search only. Further research

should investigate this differentiation in more detail.

The opportunity of investigating users’ search be-

havior upon a very large query log suffers from a vari-

ety of drawbacks. Turning to our ﬁltering and catego-

rization procedures, there may be several sources that

could skew our ﬁndings. Although we have been very

detailed on the keyword-creation-process, we still

might not have captured all of the generic and brand-

related search activities. For instance, we may have

deﬁned too few brand-related keywords, which would

result in too few brand-related queries being ﬁltered

out of the initial query log. If, on the other hand,

we had ﬁltered out more brand-related queries, this

would have, under constant conditions and chargeable

to the generic homogeneous group, increased both

the hybrid and the homogeneous (“from branded -

to branded”) search groups proportionally. We also

could have picked out more (unknown) companies

instead of focusing on the top ten for each industry.

Unknown brands might not receive as many homoge-

neous searches as the top ten industries. Addition-

ally, users would be more likely to become aware

of these unknown brands through generic searches,

which probably would have increased the spillover ef-

fect from generic to branded searches. Restricting our

study to the top companies therefore can be seen as

resulting in conservative ﬁgures since it more likely

leads to an underestimation than an overestimation of

the spillover effect.

6 CONCLUSIONS

To better understand the search behavior of users we

analyzed the query log published by AOL in 2006

(Pass et al., 2006) for the occurrence of the spillover

effect ﬁrst described by (Rutz and Bucklin, 2011).

Our results conﬁrm the occurrence of spillover effects

in online search behavior, but also show that these ef-

fects differ between industries.

Starting from the initial AOL query log we (1)

ﬁltered out every brand-related log record that could

refer to one of the selected companies or their prod-

ucts, and (2) ﬁltered out log records of the remain-

ing queries that contained generic terms for a speciﬁc

industry. Within our analysis we found more user-

journeys starting with generic and switching to brand-

related searches than the other way round. Among the

three industries selected, the effect is most noticeable

in the hotel, closely followed by the clothing industry.

The weakest degree of a spillover was detected in the

healthcare industry.

We could prove the key role of generic activi-

ties and the early anchoring of companies and brands

within user-journeys. Both our understanding of

users’ search activity as a process and our ﬁndings

of the spillover effect within user-journeys allowed

us to transfer the theoretical concepts of the evoked

and (un-)awareness set of (Howard and Sheth, 1968)

and (Narayana and Markin, 1975) to the user’s search

and decision process in the context of search en-

gines: the development of the search process from

generic to brand-related searches may at least par-

tially be attributed to the fact that individual brands

gain the users’ attention during the search process,

who thus “become aware” of these companies or

brands. The focus on generic activities becomes in-

dispensable since they seem to play the role of gate-

keeper for companies in online advertising.

This paper has several limitations. Although we

have been very detailed on the data ﬁltering process

there might be effects that blurred our results, since

the manually selected and revisited generic keywords

for each industry as well as the brand-related ones for

each company might not capture all users and his or

her search intentions. Another limitation is the miss-

ing information on conversions which would have en-

abled us to formulate a more speciﬁc occurrence of

the spillover effect. Our approach, therefore, neglects

the fact that one user might have made several pur-

chases during the investigation period within one in-

dustry. An alternate dataset might overcome these

limitations in a further investigation.

Although we analyzed several industries in this

paper, we neither considered possible interactions be-

tween them nor between companies within these (e.g.

spillovers from company A to company B). Since, for

example, (Ghose and Yang, 2010) found that there

are cross-category purchases within one company for

several keywords in sponsored search, a complete

query log to some extent contains information for

such a further analysis. Our work can be seen as a

ﬁrst step in directing attention more to the behavioral

aspects of users’ online search activities.

TheUser-journeyinOnlineSearch-AnEmpiricalStudyoftheGeneric-to-BrandedSpilloverEffectbasedonUser-level

Data

153

REFERENCES

Abhishek, V., Hosanagar, K., and Fader, P. S. (2011). On

aggregation bias in sponsored search data: Existence

and implications.

Agarwal, A., Hosanagar, K., and Smith, M. D. (2011). Lo-

cation, location, location: An analysis of proﬁtibality

of position in online advertising markets. Journal of

Marketing Research.

Animesh, A., Viswanathan, S., and Agarwal, R. (2011).

Competing “creatively” in sponsored search markets:

The effect of rank, differentiation strategy, and com-

petition on performance. Information Systems Re-

search, 22(1):153–169.

Broder, A. (2002). A taxonomy of web search. SIGIR Fo-

rum, 36(2):3–10.

Bucklin, R. E. and Sismeiro, C. (2009). Click here for

internet insight: Advances in clickstream data anal-

ysis in marketing. Journal of Interactive Marketing,

23(1):35–48.

Chan, T. Y., Xie, Y., and Wu, C. (2011). Measuring the life-

time value of customers acquired from google search

advertising. Marketing Science.

Ghose, A. and Yang, S. (2008). Comparing performance

metrics in organic search with sponsored search ad-

vertising.

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.

Ghose, A. and Yang, S. (2010). Modeling cross-category

purchases in sponsored search advertising.

Howard, J. A. and Sheth, J. N. (1968). A theory of buyer be-

havior. Rivista internazionale di scienze economiche

e commerciali, 15(6):589–618.

Interbrand & Business Week (2006). Interbrand’s best

global brands 2006.

Janiszewski, C. (1998). The inﬂuence of display character-

istics on visual exploratory search behavior. Journal

of Consumer Research, 25(3):290–301.

Jansen, B. J. and Mullen, T. (2008). Sponsored search: an

overview of the concept, history, and technology. In-

ternational Journal of Electronic Business, 6(2):114–

131.

Jansen, B. J. and Spink, A. (2007). The effect on click-

through of combining sponsored and non-sponsored

search engine results in a single listing. Proc. 2007

Workshop on Sponsored Search Auctions, Banff, AB,

Canada.

Johnson, E. J., Moe, W. W., Fader, P. S., Bellman, S.,

and Lohse, G. L. (2004). On the depth and dynam-

ics of online search behavior. Management Science,

50(3):299–308.

Moe, W. W. (2003). Buying, searching, or browsing: Differ-

entiating between online shoppers using in-store nav-

igational clickstream. Journal of Consumer Psychol-

ogy, 13(1-2):29–39.

Narayana, C. L. and Markin, R. J. (1975). Consumer be-

havior and product performance: An alternative con-

ceptualization. The Journal of Marketing, 39(4):1–6.

Pass, G., Chowdhury, A., and Torgeson, C. (2006). A pic-

ture of search. In Proceedings of the 1st international

conference on Scalable information systems, InfoS-

cale ’06, New York, NY, USA. ACM.

Rutz, O. J. and Bucklin, R. E. (2011). From generic to

branded: A model of spillover in paid search adver-

tising. Journal of Marketing Research, 48(1):87–102.

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, 30:646–665.

Search Engine Watch (2006). Delv-

ing deep inside the searcher’s mind.

http://searchenginewatch.com/3406911.

Yang, S. and Ghose, A. (2010). Analyzing the relationship

between organic and sponsored search advertising:

Positive, negative, or zero interdependence? Market-

ing Science, 29(4):602–623.

ICE-B2012-InternationalConferenceone-Business

154