When and whom to retarget?

(1)

When and whom to retarget?

(2)

When and whom to retarget?

Master Thesis

Msc Marketing Management & Intelligence

Faculty of Economics and Business

Department of Marketing

University of Groningen

By:

Tienke van Dijk (S1991558)

Riouwstraat 28a

9715 BW Groningen, The Netherlands

+31 6 25611092

t.van.dijk.6@student.rug.nl

First supervisor: Prof. dr. P.C. Verhoef

Second supervisor: F.T. Beke, Msc

(3)

ABSTRACT

As a solution to the recent surge in online advertising noise, firms increasingly use retargeting

to tailor banner ads to individual consumers in an attempt to increase their advertising

effectiveness (Helft & Vega, 2010, Peterson, 2013; Sengupta, 2013). Nevertheless, this

special form of personalization is

found to be effective only under certain circumstances (Lambrecht & Tucker, 2013) and for certain consumer characteristics (Ho et al., 2008). This research investigates when and for whom personalized retargeting works best by examining the moderating role of timing and consumer characteristics. In particular, the effect of personalization on the click-through and purchase intention of banner ads is assessed while taking into account the effect of the consumer’s position in the purchase decision process and the following consumer characteristics; privacy concern, need for cognition and need for uniqueness. An exploratory online study (N=166) showed personalization strongly increases both click-through and purchase intentions, and banners of relatively higher personalization intensity increase click-through and purchase intention most. The inclusion of a timing factor learnt that banners with medium personalization intensity work best when a consumer is in the information or the post-purchase stage. Taking into account consumer characteristics showed that consumers who care about their privacy might respond less favorably to personalized communication, but it can attract individuals with a high need for uniqueness.

(4)

PREFACE

With my huge passion for online marketing, I am proud to finish this thesis on the effectiveness of personalized online advertising. Unfortunately I failed to find a company in time that was willing to cooperate with me on this thesis, which forced me to make some changes to my initial research idea. However, this experience taught me a lot and in the end I am very pleased with the result. This work is the last project for my Master Marketing Management and Intelligence and thereby the end of my life as a student.

Hereby I would like to thank my supervisors, prof. dr. Verhoef and F. T. Beke, Msc for sharing their knowledge and providing me with useful feedback throughout the whole process of writing the thesis. Their guidance all through the semester helped me in creating the thesis the way that it is now. Also I would like to give special thanks to my (college) friends and family who supported and motivated me while writing this thesis.

I hope you will enjoy reading this thesis as much as I enjoyed writing it.

Tienke van Dijk

(5)

1. INTRODUCTION

With people spending more and more time and money online, the amount of advertising on the internet is ever increasing (eMarketer, 2013). In this online environment banner advertising is soon to become the main driver of total display ad spending ahead of search advertising (eMarketer, 2016). However, due to the growing popularity of banner advertising, it is hard to stand out from the advertising clutter. Because of an aversion to the amount of advertising on the internet, many consumers choose to avoid banner ads (Cho & Cheon, 2004) and response rates have fallen dramatically over time (Hollis, 2005): click-through rates have come down to as low as .08% (Sizmek, 2014). In response, firms are increasingly tailoring their banner ads to individual customers with a method called retargeting (Helft & Vega, 2010). Retargeting banners typically show products that are inferred from the consumer’s last visit to the retailer’s website (Bleier & Eisenbeiss, 2015b). This special form of personalization is an attempt to increase advertising effectiveness (Peterson, 2013; Sengupta, 2013; eMarketer, 2014).

(8)

behavior (Chintagunta, 1999; Homburg & Giering, 2001), previous studies ignored these differences in consumers (e.g. Bleier & Eisenbeiss, 2015a). The moderating role of consumer characteristics has received little attention in the marketing literature regarding retargeting. Therefore this study also investigates the moderating role of consumer characteristics on personalized banner effectiveness.

Retargeting is a largely unexplored instrument of advertisement personalization (Bleier & Eisenbeiss, 2015b). Given the growth of retargeting as an advertising tool (eMarketer, 2016) it is highly important to investigate. Whilst some research is done on the effects of personalization on banner effectiveness (e.g. Lambrecht & Tucker, 2013), little is known about moderators that strengthen or weaken these effects. This research fills this gap by examining the impact of timing on the effectiveness of personalized banner advertisements. In addition, where previous studies usually measure the effectiveness of banners by viewing the audience as a homogenous group (e.g. Bleier & Eissenbeis, 2015a), the current study takes into account the heterogeneous nature of the audience viewing the banner ad. The role of consumer characteristics has not been studied enough in personalized advertising (Ho, Davern, & Tam, 2008). This study addresses the gap in literature by examining the moderating role of consumer characteristics on the effectiveness of personalized banners. More importantly, where previous studies only investigated banner effectiveness by either click-through intentions (e.g Bleier & Eisenbeiss, 2015b) or purchase intentions (e.g. Van Doorn & Hoekstra, 2013; Goldfarb & Tucker, 2011), this study delves deeper by exploring the impact of personalization on both measures of banner effectiveness. This way, a more comprehensive definition of banner effectiveness is explored (Maddox & Snyder, 1997). This study is the first to explore differences in click-through and purchase intention for personalized banners. Furthermore, previous research has only investigated personalization in a single given personalization intensity (e.g. Lambrecht & Tucker, 2013). Banner effectiveness is measured for either personalized or not personalized advertising. In this study various levels of personalization are investigated which will represent reality closer since retailers are also applying various algorithms to personalize their ads in certain instances.

(9)

2. THEORETICAL FRAMEWORK

In this section theoretical concepts relevant for this research are defined and hypotheses are derived. In figure 1 the conceptual model is presented to function as a visualization for this research. Next, the concepts and relationships between them are explained in detail.

Figure 1: Conceptual framework

+

2.1 Degree of content personalization

Personalization is generally seen as a solution to the information overload, which is a consequence of the recent surge in online advertising noise. The roots of personalization lies in relationship marketing and customer management (Crosby, Evans, & Cowles, 1990; Dwyer, Schurr, & Oh, 1987). Just like a helpful sales clerk, the seller (or the online retailer), greets the customer by name, remembers what he or she has purchased or browsed previously, and suggests products he or she might be interested in the future (Bragga, Sunikka, & Kallio, 2012). Firms are using retargeting more and more for tailoring their banner advertisements to individual consumers (Lambrecht & Tucker, 2013). Therefore this study examines personalization in the field of retargeting. Retargeting is a special form of personalization where banners feature product images that match consumer’s interests, inferred from their most recent shopping behavior in the retailer’s online store (Bleier & Eisenbeiss, 2015b). A common measure for personalization in retargeting is the degree of content personalization (DCP), which determines how closely a banner reflects a consumer’s previously viewed items in the retailer’s store (Bleier & Eisenbeiss, 2015a) and thus how closely the ad will relate to the consumer’s inferred interests (Bleier & Eisenbeiss, 2015b). Following Bleier & Eisenbeiss (2015a), this results in the following levels of personalization:

(10)

1. High DCP: A banner shows products sampled from a consumer’s most preferred category and brand combination

2. Medium DCP (brand): A banner shows products sampled from a consumer’s most preferred brand

3. Medium DCP (category): A banner shows products sampled from a consumer’s most preferred category

Note that this conceptualization does not suggest that personalization based on brands rather than categories represents a consumer’s preferences more closely. It only indicates that combining both brand and category choice in a banner (high DCP) represents preferences more closely than either of these choices separately. Banners with no personalization are used as a control condition. This is a banner showing products sampled from random products from arbitrary categories and brands.

2.2 Banner effectiveness

Advertising effectiveness is usually described as being a ‘hierarchy of effects’ (Vakratsas & Ambler, 1999). The effect mentioned is to change the consumer’s mind about a product by changing their attitudes and then by acting it out (Hall, 2002). The banner advertisement would change a consumer’s perception and eventually his or her behavior. Since measuring actual behavior is out of scope, banner effectiveness will be measured by click-through intention and purchase intention.

2.1.1 Click-through intention

Consistent with previous studies (e.g. Bleier & Eisenbeiss, 2015b; Yoo, 2007) banner effectiveness is measured by its click-through intention, which is the extent to which consumers are willing to click to learn more about the product represented in the banner. Click-through intention is seen as an appropriate measure for banner effectiveness (Yoo, 2007). However, this measure alone would not give a full picture of the effectiveness of a banner advertisement. Just like click-through rate is merely a measure of a visit to the retailer’s website (Manchanda, Dubé, Goh, & Chintagunta, 2006), so is click-through intention only a measure of an intention to click on the banner. The banner still has effect for people who do not want to click on the banner; merely viewing a banner has a significant impact on raising brand awareness, brand perception and intent to purchase (Maddox & Snyder, 1997). In order to capture the effects of merely viewing a banner ad as well, purchase intention is chosen as a second measure for banner effectiveness.

2.1.2 Purchase intention

(11)

Purchase intention as a measure for banner effectiveness is in line with previous studies (e.g. Van Doorn & Hoekstra, 2013; Goldfarb and Tucker 2011).

2.3 DCP and banner effectiveness

Bleier & Eisenbeis (2015b) find personalization increases click-through intention for banners of more trusted retailers because consumers perceive these banners as more useful. This finding is supported by eMarketer (2015); they find that more than three in five digital consumers (62%) see retargeting advertisements as useful and believe digital ads are generally well targeted to their interests. According to Bleier and Eisenbeis (2015b) the higher the personalization will be, the closest the products come to that individual’s revealed interests and preferences. Therefore all three levels of DCP are expected to positively influence click-through intention, where high DCP is expected to generate the highest click-through intentions, resulting in the following hypotheses:

H1a: Overall, DCP has a positive effect on click-through intention compared to no DCP.

H1b: High DCP has the most positive effect on click-through intention compared to no DCP.

With regard to purchase intention, Van Doorn & Hoekstra (2013) find higher degrees of personalization negatively affect purchase intention because feelings of intrusiveness are evoked. This negative effect however, is compensated for by an ad that effectively fits a consumer’s current needs. An ad with high fit is tailored to the needs of the consumer and therefore represents relevant information (Tam & Ho, 2006). The higher the DCP will be, the closest the products come to the consumer’s revealed interests and preferences (Bleier & Eisenbeiss, 2015b) and thus would represent more relevant information which results in a high fit. Therefore DCP will have a positive effect on purchase intention, where high DCP will have the most impact, as summarized in the following hypotheses:

H2a: Overall, DCP has a positive effect on purchase intention compared to no DCP.

H2b: High DCP has the most positive effect on purchase intention compared to no DCP.

(12)

2.4 Timing

Lambrecht & Tucker (2013) find personalized banners used in retargeting to be, on average, less effective than banners without personalization. According to their research advertising content that specifically reflects the products consumers viewed during the last shopping trip is, in general not effective. Instead, retargeting consumers with a non-personalized advertisement is found to be effective. Because other literature conflicts with these findings by showing significant positive impact of personalization on banner effectiveness (e.g. Ansari & Mela, 2003; Chen, Pavlov & Canny, 2009) this finding is surprising (Lambrecht & Tucker, 2013). It suggests personalization is dependent on other factors in order to be effective. The central question now, is: when is personalized retargeting effective? An answer can be found in the preference literature. At first a consumer often has only a general idea of what he wants to purchase. At this stage the consumer’s preferences are constructed at a high level and he only focuses on high-level goals. Over time, the consumer knows what he wants to purchase and he develops narrowly construed preferences, which means he has developed a detailed viewpoint of what he wants to purchase. Therefore, the effectiveness of a personalized advertisement depends on whether the concreteness of its message matches how narrowly consumers construe their preferences (Lee, Keller, & Sternthal, 2010; Trope, Liberman, & Wakslak, 2007). Consumers with high-level goals only have a broad idea of what they want so they will respond better to advertising that addresses such goals –advertising without personalization – than to advertising that shows specific products. Only consumers that have narrowly construed preferences respond favorably to personalized advertising (Lee, Keller & Sternthal, 2010; Trope, Liberman, & Wakslak, 2007).

What causes a consumer to change from general ideas to narrowly construed preferences? Here is where the concept of timing comes in. Distant future events are represented in a more abstract, structured, high-level manner than near future events (Trope, Liberman, & Wakslak, 2007). So the further away the purchase, the broader the idea of what to purchase. As a consumer progresses through the purchase decision process, the event of making a purchase comes closer. This will cause consumers to be more aware of what they want to buy. Also they become aware of specific product attributes and determine how much weight they place on each attribute (Hoeffler & Ariely, 1999). The consumer’s position in the purchase decision process would therefore influence the effectiveness of personalized banners. Next, an explanation of the different stages in the purchase decision process is given.

2.4.1 Stage in online purchase decision process

(13)

model that illustrates the main stages of the purchase process. It comprises the need recognition, information search, evaluation of alternatives, purchase and post-purchase stages. It has been the standard model in consumer behavior research and online consumer research.

In this research a four-stage model specialized on the online shopping behavior is used as a base (Li & Chatterjee, 2005). This model is built on the information processing theory of consumer choice and parallels consumer decision-making stages (Howard & Sheth, 1969). The purchase-decision making process contains the following stages. First, a consumer enters the retailer’s site and starts viewing pages with product information, i.e. the information stage. This stage is followed by the consideration stage, when the consumer uses the virtual shopping cart; he either only places products in his shopping cart or he also decides to view the shopping cart page. Last, the post-purchase stage will come if a consumer decides to purchase the products (Bleier & Eisenbeiss, 2015a). Next, the different stages in the purchase decision-making process are explained in detail.

Information stage. A consumer is classified to be in an information search stage when he is

seeking information and is processing activities which one engages in facilitating decision making regarding some goal object (i.e. making a purchase) in the marketplace (Li & Chatterjee, 2005). This is at the beginning of the purchase decision process where the consumer has only viewed product information (i.e. product pages) but not yet conducted in any purchase-related activities.

Consideration stage. A consumer is ought to be in the consideration stage when he used the virtual

shopping cart but not yet made a purchase. The consumer merely placed products in his virtual shopping cart or he viewed the virtual shopping cart as well.

Post-purchase stage. A consumer is defined to be in the post purchase state if he completed a

purchase before exiting the retailer’s online store.

(14)

H1c: The effect of DCP on click-through intention will be most strong if a consumer is in the consideration stage.

H2c: The effect of DCP on purchase intention will be most strong if a consumer is in the consideration stage.

2.5 Consumer characteristics

The mixed results regarding the effectiveness of personalized banner ads suggest personalization depends on other factors. Previous studies usually measure the effectiveness of banners by viewing the audience as a homogenous group (e.g. Bleier & Eissenbeis, 2015a; Lambrecht & Tucker, 2013). Marketing literature has emphasized the importance of consumer’s characteristics in purchase behavior (e.g., Chintagunta, 1999; Homburg and Giering, 2001). This raises the question: whom to target? In order to answer this question, the moderating effects of consumer characteristics are researched.

2.5.1 Privacy concern

The mixed results of the effectiveness of personalization may be caused by the fact that consumers are increasingly becoming worried about their privacy (Langheinrich, Atsuyoshi, Naoki, Tomonari, & Yoshiyuki, 1999). Privacy concern is “an individual’s subjective view of fairness within the context of information privacy” (Malholtra, Kim, & Agarwal, 2004). Consumers’ privacy concerns arise from the feeling that their information is vulnerable and that they are not able to control their personal information (Dinev & Hart, 2004). Those feelings of losing control result in resistance towards sharing individual information (Rubini, 2001) and hence cause resistance towards such information used. Since personalized banners make use of such individual information, individuals high in privacy concern are less inclined to click on those banners. Therefore personalization is expected to be less effective in terms of click-through intention among individuals concerned with privacy, as summarized in the following hypothesis:

H1d: Privacy concern weakens the positive effect of DCP on click-through intention.

(15)

about their privacy are reluctant to purchase products online. Therefore personalization is expected to be less effective in terms of purchase intention among individuals concerned with privacy, as indicated in the following hypothesis:

H2d: Privacy concern weakens the positive effect of DCP on purchase intention.

2.5.2 Need for cognition

Need for cognition is an “individual’s propensity to engage in and enjoy thinking” (Cacioppo & Petty, 1982). It represents the extent to which individuals have the tendency to engage in effortful cognitive activities. Individuals with high need for cognition are intrinsically motivated to engage in cognitive activities (Cacioppo & Petty, The Need for Cognition, 1982). They are more discerning and careful (Petty & Cacioppo, 1986) which makes them more sensitive to the level of preference matching in personalized content (Ho, Davern, & Tam, 2008). Therefore, individuals with a high need for cognition would respond more favorably to personalized banners than individuals low in need for cognition, resulting in the following hypotheses:

H1e: Need for cognition strengthens the positive effect of DCP on click-through intention.

H2e: Need for cognition strengthens the positive effect of DCP on purchase intention.

2.5.3 Need for uniqueness

Need for uniqueness is defined as “the trait of pursuing differences relative to others through the acquisition, utilization, and disposition of consumer goods for the purpose of developing and enhancing one’s self-image and social image” (Tian, Bearden, & Hunter, 2001). The concept consists out of three dimensions: (1) creative choice counterconformity; an individual’s ability to make selections that are likely to be considered as unique but also approved by others, (2) unpopular choice counterconformity; an individual’s desirability to use products that are somewhat deviating from social norm to establish their differentness from others and (3) avoidance of similarity; the avoidance of buying products that become common or the discontinued use of these products (Tian, Bearden, & Hunter, 2001). Individuals high in need for uniqueness have the need to see themselves as being different from others (Snyder & Fromkin, 1977). Personalized banners are tailored to each single consumer's interests and preferences (Ansari & Mela, 2003) and therefore acknowledges the individuality of each recipient by making consumers feel unique (Kalyanaram & Sundar, 2006). An individual with a high need for uniqueness therefore would be more inclined to click on a banner that is personalized, resulting in the following hypothesis:

(16)

For the same reasoning, purchase intention for personalized banners will be higher for individual’s high in need for uniqueness. A personalized banner addresses the need to be unique because it makes an ad more relevant and thus more appealing to a consumer (Chen, Pavlov, & Canny, 2009). Individuals with high need for uniqueness would favor this information more than people low in need for uniqueness, because the individuality and relevance in the ad makes them feel unique (Kalyanaram & Sundar, 2006). Personalization therefore would be more effective in terms of purchase intention among individuals high in need for uniqueness, as stated in the following hypothesis:

H2f: Need for uniqueness strengthens the positive effect of DCP on purchase intention.

2.6 Control variables

(17)

3. RESEARCH DESIGN

3.1 Exploratory online study

To test the hypotheses, a scenario-based online experiment is conducted. For different banners the proposed effects of personalization on banner effectiveness are investigated whilst taking into account the consumer’s position in the purchase decision process and consumer characteristics. The resulting experimental setup was a 4 (high DCP, medium DCP (brand), medium DCP (category), no DCP) x 3 (information, consideration, post purchase stage) between subjects design. Respondents were randomly assigned to one of the 12 conditions.

The experiment consisted of five sections: (1) First, the participant reads a short introduction asking him to imagine he was looking to buy new shoes at the online store of a retailer named Shoestore.com1. Depending on the intended purchase stage scenario, participants then received specific background information that described their purchase decision stage. In particular, the wordings of these scenarios reflected the key characteristics of each purchase decision stage as described by Bleier and Eissenbeis (2015a) and Li and Chatterjee (2005) (for definitions see Appendix A). Scenarios were tested by a pre-test collected from a sample of 20 participants in order to check if scenarios were clear. In the pre-test first the participant read a short introduction defining the three different consumer purchase decision stages based on the definition of Bleier and Eissenbeis (2015a) and Li and Chatterjee (2005). Next, the participant read all three scenarios and were asked to give the corresponding purchase stage. Based on the pre-test some wordings were slightly changed in order to make the different scenarios more clear, but overall all participants labeled each scenario right (for complete wording see Appendix A). (2) Next, in a created Shoestore online web shop participants had to select a pair of shoes that they liked best from the retailer’s assortment. (3) Then, the participant reads a cover story that he either left the online store without purchasing the shoes (in scenario 1 and 2; information search and consideration stage) or bought the shoes (in scenario 3; post-purchase stage). (4) Subsequently, the participant was shown a news site that he supposedly visited the following day (Nu.nl). Here he encounters a banner from Shoestore.com, the retailer’s store he had previously visited. Depending on the treatment group, this banner featured exact the same product as he had selected in the previous question (high personalization), a product from the same brand (medium personalization – brand), a product from the same category (medium personalization – category), a totally other product (control condition)2. As explained in the literature section, this conceptualization is based on Bleier and Eissenbeis (2015a). (5) Last, participants filled out a

1_{A fictional retailer name is used to increase the realism of the experiment and make sure that participants had}

no predetermined opinions about it

2

(18)

questionnaire. Its measures included evaluations of participants’ click-through and purchase intentions, privacy concerns, need for cognition and need for uniqueness. To minimize common method variance and response bias several remedies indicated by Podsakoff, Mackenzie, Lee, and Podsakoff (2003) are implemented in the questionnaire. For instance, in the introduction participants were ensured that their responses would be treated anonymously, only used for this research and that there were no right or wrong answers.

3.2 Sample

The experiment is conducted with Dutch participants who were randomly assigned to one of the 12 conditions. In total 201 participants completed the survey. Before analyzing, the data is checked for outliers, missing values and oddities. 8 cases of participants who did not filled out the entire survey were removed. Also, 11 participants who completed the survey in less than 1 minute were removed (average completion time is 3.29 minutes). Moreover, data from 18 participants who did not pass the attention check is discarded. In this check participants were asked to tick a specific box in order to check if they really paid attention while answering the questions. These are common methods used to detect participants who answer questions without the necessary effort (Huang, Curran, Keeney, Poposki, & DeShon, 2012). After eliminating these unusable data, 166 valid questionnaires remained for further analysis. Of these 166 participants, 70 were male (42%) and 96 female (58%) with an average age of 24.

3.3 Measures

Click-through intention is measured with a single-item based on Yoo (2007): “I would like to click-through the banner ad to acquire further information.” Purchase intention is measured by three items based on Grewal, Monroe, & Krishnan (1998) following Van Doorn and Hoekstra (2013): “The likelihood of purchasing this product is large,” “The probability that I would consider buying the product is large” and “If I am going to buy a product, the probability of buying the product represented in the banner is large.” Privacy concern is measured with four items following Bleier & Eissenbeis (2015b), whose scale is based on Sheng, Nah, and Siau (2008): “It bothers me that firms are able to track information about me,” “I am concerned that firms have too much information about me,” “It bothers me that firms are able to access information about me,” and “I am concerned that my information could be used in ways I cannot foresee”. Following Ho et al. (2008), need for cognition is measured by the five items with the highest factor loadings of the 34-items of the Consumer’s Need for Cognition scale (Cacioppo and Petty, 1982): “I don’t like to have to do a lot of thinking [r]3, “I try to avoid situations that require thinking in depth about something [r], “I prefer to do something that

3

(19)

challenges my thinking abilities rather than something that requires little thought, “I prefer complex to simple problems”, and “Thinking hard and for a long time about something gives me little satisfaction [r]”. Need for uniqueness is measured by the three items (each covering one of the three dimensions of the concept) with the highest factor loadings from the 12-items of the Consumers‘ Need for Uniqueness scale (Ruvio, Shoham, & Makovec Brenčič , 2008): “I actively seek to develop my personal uniqueness by buying special products or brands” (Creative choice), “I enjoy challenging the prevailing taste of people I know by buying something they would not seem to accept” (Unpopular choice), and “I often try to avoid products or brands that I know are bought by the general population” (Avoidance of similarity). All items were rated on a seven-point rating scale ranged from 1 (“Strongly disagree”) to 7 (“Strongly agree”).

As can be seen in table 1, all constructs meet convergent validity. One item (NFC2) of the need for uniqueness scale had to be removed for this construct to meet the required threshold of 0.5 for the average variance extracted (AVE) (Fornell & Larcker, 1981). By removing the item from the scale the AVE increased from 0.447 to 0.518. Also the Cronbach’s alpha increased from 0.674 to 0.677. Furthermore, all constructs exceed the required threshold of 0.6 for the Cronbach’s alpha (Malholtra, 2010). The same applied to the Confirmatory Factor Analysis. For each construct it was appropriate to reduce the items to 1 construct (Eigen value >1, total explained variance >60%). T-values greater than 1.645 for every item indicated that their factor loadings were significantly different from zero (Bagozzi, Yi, & Philips, 1991). For each construct the composite-based reliabilities were above the 0.6 threshold recommended by Bagozzi, Yi, and Philips (1991).

3.4 Model choice

This study investigates the click-through and purchase intention of individual consumer behavior. In order to do so, two predictive models are built; one predicting a consumer’s click-through intention and another predicting a consumer’s purchase intention. For both models is chosen to have a mathematical form that is linear in parameters and variables, also referred to as a linear additive model. The most common method used for estimating the parameters is Ordinary Least Squares (OLS) (Leeflang, Wieringa, Bijmolt, & Pauwels, 2015). Both models have the following structure:

Where:

= value of the dependent variable for individual consumer i ,

= value of independent variable k for individual consumer i, (k=1,…,K), and

(20)

Table 1. Measurement model fit indices SL t-value CA CR AVE Click-through intention 1.000 1.000 1.000 CTI 1.000 999.000 Purchase intention 0.893 0.934 0.825 PI1 0.888 19.461 PI2 0.928 20.166 PI3 0.908 21.124 Privacy concern 0.924 0.946 0.815 PC1 0.916 33.997 PC2 0.928 34.047 PC3 0.922 34.800 PC4 0.843 36.788

Need for cognition 0.677 0.809 0.518

NFC1 0.659 43.011

NFC3 0.554 38.638

NFC4 0.773 41.493

NFC5 0.792 46.779

Need for uniqueness 0.602 0.791 0.558

NFU1 0.734 31.807

NFU2 0.787 26.310

NFU3 0.718 28.271

Note: SL, standardized factor loadings; CA, Cronbach’s alpha; CR, composite reliability; AVE, average variance extracted

3.5 Model specification

The mathematic representation of the first model, which predicts an individual customer’s click-through intention, is as follows:

(21)

= constant for individual consumer i,

= dummy variable indicating whether consumer i has viewed a banner of high DCP or not,

= dummy variable indicating whether consumer i has viewed a banner of medium DCP (brand) or not,

= dummy variable indicating whether consumer i has viewed a banner of medium DCP (category) or not,

= dummy variable indicating whether consumer i is in the information stage or not4, = dummy variable indicating whether consumer i is in the consideration stage or not5, = privacy concern for individual consumer i,

= need for cognition for individual consumer i, = need for uniqueness for individual consumer i, = age for individual consumer i,

= gender for individual consumer i, and

= error term.

The mathematic representation of the second model, which predicts an individual customer’s purchase intention, is as follows: For the descriptions of each variable is referred to the list below the first model for parsimonious reasons.

4_{Reference category is post-purchase stage} 5

(22)

4. RESULTS

4.1 Model estimation

For the model predicting the click-through intention of an individual consumer, a linear regression is performed as represented in equation (1). For the second model, which is predicting individual consumer’s purchase intention, a linear regression is conducted as represented in equation (2). For the model parameters to be estimated with OLS, several assumptions need to hold. In the following, first these assumptions are checked and described. Subsequently, model validation is described, where after the results are interpreted. These steps are first described for model 1 and then for model 2.

4.2 Model 1: CTI

i

4.2.1 Checking assumptions

Multicollinearity In order to check whether multicollinearity is an issue, the variation inflation (VIF) index is used. Variables with a VIF score higher than 5 are problematic. A linear regression performed for the full model as represented in equation (1) shows severe issues of multicollinearity as can be seen from table 2.

Table 2. Problematic VIF values for model 1: CTIi

VIF VIF after mean

centering

centering DCP1 89.212 4.812 DCP3xPC 19.895 1.882 DCP2 99.957 5.597 DCP1xNFC 42.173 1.876 DCP3 87.948 4.978 DCP2xNFC 41.962 1.981 Stage 1 5.115 5.093 DCP3xNFC 38.624 2.226 Stage 2 6.358 5.795 NFU 6.481 6.547 DCP2xStage2 6.175 5.768 DCP1xNFU 23.195 2.747 DCP1xPC 15.093 1.763 DCP2xNFU 28.233 3.242 DCP2xPC 17.206 2.345 DCP3xNFU 21.146 2.847

(23)

correct reference category. Their recommended choice for the reference category is the one who has the highest relative frequency. Initially the post-purchase stage was set as a reference category for the stage. Since the consideration stage has the highest frequency (37%) opposed to the information (34%) and post-purchase stage (28%), this stage is set as reference category. However, this does not decrease the VIF values significantly (Stage 1 VIF= 4.782; Stage 2 VIF= 5.468). For the ease of interpretation the reference category is set back to the post-purchase stage. Because the VIF values are not extremely high (Stage1 VIF=5.093; Stage 2 VIF=5.795) and the parameters does not change when regression is performed twice (once with and once without stage 1 and 2 and its interaction effects) it is assumed that it will not cause any bias. NFU correlates slightly with PC, r(143)=0.141, p<0.10. Since the VIF value of NFC (VIF=6.547) is not that high, it is checked whether the correlation between NFU and PC really makes the parameter estimates unreliable. In order to do so, a regression is performed twice (once with NFU and its interaction effects and once without them). Both regressions showed no differences in parameters and therefore it is assumed that NFU and PC are linearly independent and thus will not cause any bias.

Non-normality. The second assumption that needs to be satisfied is that the disturbances are

normally distributed. To examine whether this is the case, a Kolmogorov-Smirnov and a Shapiro-Wilk test are conducted. The Kolmogorov-Smirnov test is significant, D(143)=0.117, p<0.01, which suggests the residuals are not normally distributed. The Shapiro-Wilk test confirms this result,

W(143)=0.959, p<0.01), indicating the residuals are not normally distributed. The corresponding

histogram shows that the distribution of the residuals is right-skewed. To solve the problem the Box-Cox transformation is applied (Box & Box-Cox, 1964). By taking the log of the dependent variable CTI, the new dependent variable (lnCTI) is normally distributed according to the Kolmogorov-Smirnov test, D(143)=0.056, p>0.10, and the Shapiro-Wilk test, W(143)=0.990, p>0.10, indicating the error terms are normally distributed.

Heteroscedasticity. The third assumption that needs to be satisfied is that the error term needs

to be homoscedastic, which means it has the same variance in all cases cross-sectionally (Leeflang et al., 2015). Before applying the Box-Cox transformation (Box & Cox, 1964), this assumption was violated. The Breusch-Pagan test for heteroscedasticity (Breusch & Pagan, 1979) was significant,

F(25, 117)=1.688, p<0.05), which indicates the error terms were not the same across the conditions.

The White test (1980) supported these results, F(2, 140)=3.657, p<0.05. A remedy is to use an appropriately adjusted formula for the variances and covariances of the parameter estimates (Leeflang, Wieringa, Bijmolt, & Pauwels, 2015). By taking the log of the dependent variable CTIi (explained in

(24)

Omitted variable bias. To see if omitted variable bias is an issue the R2 adjusted and the Ramsey Regression Specification Error test (RESET-test) (Ramsey, 1969) are taken into account. Parameters of the model will be biased if not all explanatory variables are included in the model. The R2 adjusted of the model, R2 adjusted=0.197, does not give any suspicion for omitted variable bias because the explanatory power of the model is normal for predictive models of click-through intention (for comparison see e.g. Bleier & Eisenbeiss, 2015b). For the Ramsey RESET-test two extra variables are included in the model to see if there are omitted variables: the squared predicted values of lnCTIi

(Ŷ2) and the predictor values of lnCTIi to the power of 3 (Ŷ 3

). Both variables are not significant,

p>0.10, which indicates there is no omitted variable bias in the model.

Wrong functional form. The mathematical form of this model is a form that is linear in both

parameters and variables. A misspecification of a model occurs when the wrong functional form is chosen. To examine whether this is the case, again the R2 adjusted and the RESET-test (Ramsey, 1969)are used. The R2 adjusted, R2 adjusted =0.197 does not seem to indicate misspecification since it is fairly normal for a prediction model of click-through intention (for comparison see e.g. Bleier & Eisenbeiss, 2015b). As explained in the previous paragraph, also the Ramsey RESET-test does not give any suspicion that the wrong functional form is used, F(26, 116)=3.386, p>0.10.

Dealing with these assumptions resulted in the following new mathematical representation of the first model: 4.2.2 Model validation

Statistical validity. Looking at the estimated model as represented in equation (3), the overall

(25)

Table 3. Model comparisons for model 1 CTIi

Model 1 Model 2 Model 3

(26)

When looking at the R2 adjusted, model 2 has the highest, R2 adjusted= 0.267, indicating this model has the highest explanatory power. When taking into account the accuracy of estimation the information criteria AIC, AIC3 and BIC6 are used to see which model performs best. As can be seen from table 3, model 2 has the minimum AIC, AIC3 whereas model 3 has the minimum BIC. Because model 3 shows such an extreme decrease in R2 adjusted compared to model 2, namely from 0.267 to 0.241, the variables that are not significant must have some explanatory power in predicting lnCTIi.

Therefore it is concluded that model 2 performs best and will be chosen for further analysis.

Predictive validity. To assess the predictive validity of the model, a (randomly chosen)

validation sample of 10% is used. For the model predicting click-through intention the Mean Absolute Percentage Error (MAPE) is conducted. The MAPE is 12%, meaning that on average the predictions for click-through intention deviate about 12% from corresponding observed values.

4.2.3 Model interpretation

In the following, it is seen to what extent the results support the hypotheses in order to determine to what degree timing and consumer characteristics moderate the effect of DCP on click-through intention. In order to do so, first the main effects are described, followed by a description of the interaction effects.

Main effects. H1a proposes DCP has a positive effect on click-through intention. The results

show full support for H1a. High DCP has a significant positive effect on click-through intention,

β=0.568, p<0.01, indicating high DCP will increase click-through intention by 56.8% compared to no

DCP. Also medium DCP(brand), β=0.314, p<0.05, and medium DCP(category), β=0.304, p<0.05, positively influence through intention, indicating medium DCP(brand) will increase click-through intention by 31.4% and medium DCP(category) will increase click-click-through intention by 30.4%. Hence, it can be concluded that an online banner with high DCP, medium DCP(brand) and medium DCP(category) will increase click-through intention of consumers compared to banners with no DCP. The results also support H1b, which suggests high DCP has the strongest effect on click-through intention. Looking at the strength of effects (so looking at the standardized coefficient), high DCP, β=0.384, has the strongest impact on click-through intention, compared to medium DCP(brand),

β=0.242, and medium DCP(category), β=0.222, meaning high DCP increases click-through intention

the most. Another main effect is found for privacy concern, β=-0.140, p<0.01, indicating consumers concerned about their privacy will have lower click-through intention than consumers not concerned

6_{The following equations are used: AIC = -2 ln(L) + 2 (K+1); AIC3 = -2 ln(L) + 3 (K+1); BIC = -2 ln(L) +}

ln(T) (K+1); Where -2ln(L)= T ln(RSS/T)

BIC -77.209 -152.420 -158.042

(27)

about their privacy. Also a main effect is found for NFC, β=-0.069, p<0.10, indicating consumers high in NFC will have lower click-through intentions compared to consumers low in NFC. The suggested control effects of age and gender does not seem to hold. Click-through intentions do not differ between men or women, p>0.10. Age is also not significant, p>0.10, however , since p=0.161, the variable seems to be almost significant. In addition, when excluding age from the model the R2 adjusted decreases from 0.259 to 0.255. This decrease in R2 indicates that age has some explanatory power in predicting click-through intention. As hypothesized, the effect of age is positive, indicating a one unit increase in age will increase click-through intention by 1.4%. Older consumers therefore have higher click-through intention for online banners.

Interaction effects. For H1c the moderating effect of the consumer purchase decision stage on

click-through intention is investigated. H1c suggests the effect of DCP on click-through intention will be most strong in the consideration stage. The results do not support H1c. There is a moderating effect of a consumer’s position in the purchase decision process, but this effect is not the strongest in the consideration stage. The influence of medium DCP(brand) on click-through intention will be weaker when a consumer is in the information stage compared to the post-purchase stage, β=-0.448, p<0.05, indicating the effect of medium DCP(brand) on click-through intention will be decreased by 44.8% if the consumer is in the information stage compared to the post-purchase stage. Performing multiple regressions with other reference categories reveal the effect of medium DCP(brand) on click-through intention indeed is stronger when a consumer is in the post-purchase stage than the information stage,

β=0.172, p<0.10. It is also found that the effect of medium DCP(category) on click-through intention

(28)

partially supported, meaning that if a consumer has a high need for uniqueness, click-through intention for high DCP and medium DCP(brand) banners will be higher.

When taking into account the strength of effects (so looking at the standardized coefficients), the results show the effect of high DCP compared to no DCP is most strong on click-through intention,

β=0.384, followed by privacy concern, β=-0.380. High DCP and privacy concern thus have the

strongest effect on consumer’s click-through intentions for online banners. Then, medium DCP(brand), β=0.242, and medium DCP(category), β=0.222, follow. Next, the interaction effect of medium DCP(brand) and the information stage, β=-0.194, the interaction effect of high DCP and privacy concern, β=0.141, NFC, β=-0.121, the interaction effects of medium DCP(brand) and NFU,

β=0.120, and the interaction effect of high DCP and NFU, β=0.116. Age has the least effect on

click-through intentions, β=0.104.

4.3 Model 2: PI

_i

4.3.1 Checking assumptions

Multicollinearity. A linear regression performed as represented in equation (2) shows high

issues of multicollinearity as can be seen in table 4. Variables with a VIF score higher than 5 are presumed to be problematic. To lower the VIF values for the interaction terms, the continuous variables PC, NFC and NFU are mean centered. This is also necessary to interpret results in a logical way. However, still some variables show VIF scores above 5: DCP2, Stage1, Stage2, and NFU as can be seen in table 4. For dummy variables the correct reference category has to be chosen in order to solve multicollinearity (Wißmann, Toutenburg, & Shalabh, 2007). Therefore the reference category for stage is set to the consideration stage (the category with the highest frequency). This decreased VIF values minimally (VIF Stage1= 4.927; Stage2=6.132). For the ease of interpretation the reference category is set to purchase stage again. Performing a regression without these variables showed no

Table 4. Problematic VIF values for model 2: PIi

centering

(29)

significantly changing signs and values of the parameters in the model, hence it is assumed that the variables are linearly independent and thus will not cause any biased results. For the DCP dummy variables the reference category is kept at No DCP because it is necessary for interpretation. Also these variables caused no changing parameters while being left out of the regression, assuming no problems of interpretation bias will occur. NFU has a VIF value slightly above 5 because this variable correlates slightly with PC, r(143)=0.141 , p<0.10. However, these variables are assumed to be linearly independent because the parameters showed no significant changes when they were left out of the model.

Non-normality. To see whether the residuals are normally distributed the

Kolmogorov-Smirnov and Shapiro-Wilk test are conducted. The Kolmogorov-Kolmogorov-Smirnov test was significant

D(143)=0.213, p<0.01, indicating the residuals are not normally distributed. The Shapiro-Wilk test

confirms this, W(143)=0.747, p<0.01. The Box-Cox transformation (Box & Cox, 1964) does not seem to solve the non-normality. Another reason for non-normality tests to indicate deviations from the normal distribution is the presence of outliers in the residuals (Leeflang, Wieringa, Bijmolt, & Pauwels, 2015). Outliers indeed are detected in the data. To accommodate for this, the extreme outliers are detected and excluded from the data. Checking for non-normality again, the Kolmogorov-Smirnov and Shapiro-Wilk test, whilst excluding the outliers, still show no normally distributed residuals since the Kolmogorov-Smirnov test is significant, D(139)=0.189, and also the Shapiro Wilk test, W(139)=0.807, p<0.01. Detecting and excluding more outliers from the data will be a possible solution for the non-normality in this sample. However, in doing so, many outliers need to be removed which causes a lot of information loss. To prevent losing valuable information it is checked whether the non-normality actually bias results. By running the regression of the full model twice (once with and once without the outliers) the parameters show no significant changes. Therefore it can be concluded the outliers will not cause any biased results.

Heteroscedasticity. In order to check whether heteroscedasticity is an issue, the

Breusch-Pagan test for heteroscedasticity is conducted (Breusch & Breusch-Pagan, 1979). The test is insignificant, F(25, 117)=1.216, p>0.10, which indicates the residuals are homoscedastic. However, the White test (1980) contradicts these result, F(2, 140)=9.320, p<0.01. Since the Breusch-Pagan test is based on the idea whether the squared residual is related to one or more independent variables (Leeflang, Wieringa, Bijmolt, & Pauwels, 2015), the nature of heteroscedasticity is not caused by a relation between the squared residual and a predictor variable. This is the reason why taking the log of PIi (Box & Cox,

(30)

the full model is performed twice: once with and once without the outliers. No differences are detected in parameters, hence it is assumed that these outliers do not bias results.

Omitted variable bias. To check whether omitted variable bias is present the R2 adjusted and the Ramsey Regression Specification Error test (RESET-test) (Ramsey, 1969) are taken into account. For the Ramsey RESET-test two extra variables are included in the model to see if there are omitted variables: the squared predicted values of PIi (Ŷ

2

) and the predictor values of PIi to the power of 3

(Ŷ3

). Both variables are not significant, F(25, 113)=3.368, p>0.10, which indicates that there is no omitted variable bias in the model. The R2 adjusted of the model, 0.436 does not give any suspicion for omitted variable bias because the explanatory power of the model is high (for comparison see Van Doorn & Hoekstra, 2013; Goldfarb & Tucker, 2011), meaning the level of DCP together with consumer’s position in the purchase decision process and the consumer’s characteristics in this research explain almost half of the variance in purchase intention for the product represented in the banner.

Wrong functional form. The mathematical form of this model is a form that is linear in both

parameters and variables. A misspecification of a model occurs when the wrong functional form is chosen. To examine whether this is the case, again the R2 adjusted and the RESET-test (Ramsey, 1969) are used. The R2 adjusted, 0.436 does not seem to indicate misspecification since it is high for a prediction model of purchase intention (for comparison see Van Doorn & Hoekstra, 2013; Goldfarb & Tucker, 2011). As explained in the previous paragraph, also the Ramsey RESET-test, F(25, 113)=3.368, p>0.10, does not give any suspicion that the wrong functional form is used.

Dealing with these assumptions resulted in no changes in the model and thus the model predicting a consumer’s purchase intention as represented in equation (2) is used for further analysis.

4.3.2 Model validation

Statistical validity. When looking at the model as represented in equation (4), the overall model is

(31)

Table 5. Model comparisons for model 2 PIi

Model 1 Model 2 Model 3

Coefficient

(standard error) p-value

Coefficient

(32)

AIC3 92.835 62.098 57.064

BIC 151.720 97.430 86.506

Note: *p < 0.10; **p<0.05; ***p<0.01.

Predictive validity. To assess the predictive validity of the model the Mean Absolute

Percentage Error is conducted. For the model predicting purchase intention the MAPE is 17%, meaning that on average the predictions for purchase intention deviate about 17% from corresponding observed values.

4.3.3 Model interpretation

In the following, it is seen to what extent the results support the hypotheses in order to determine to what degree timing and consumer characteristics moderate the effect of DCP on purchase intention. In order to do so, first the main effects are described, followed by a description of the interaction effects.

Main effects. H2a suggests DCP has an overall positive effect on purchase intention. The

results confirm this hypothesis. High DCP increases purchase intention, β=2.294, p<0.01, and the same applies to medium DCP (brand), β=1.254, p<0.01, and medium DCP(category), β=0.932,

p<0.01. Indicating an online banner with DCP will increase click-through intention of consumers

compared to banners without DCP. H2b proposes high DCP has the strongest effect on purchase intention. Looking at the strength of effects (standardized coefficients), high DCP has the most impact on purchase intention, β=0.604 for high DCP, β=0.376 for medium DCP(brand) and β=0.266 for medium DCP(category). Therefore H2b is fully supported, which indicates that online banners with high DCP will increase purchase intention most. Also a significant main effect is found for a consumer’s stage in the purchase decision process: purchase intention is higher for consumers that are in the information stage compared to the post-purchase stage, β=0.484, p<0.10. Another main effect is found for NFU, β=0.165, p<0.10, indicating NFU directly influences purchase intention in such a way that consumers high in NFU have higher purchase intention than consumers low in NFU. For the suggested control variables, only age is significant, β=0.093, p<0.01, indicating older consumers have higher purchase intentions for an online banner. The results suggest that a one unit increase in age will result in a 0.093 increase in purchase intention. Purchase intention does not differ between men or women, p>0.10.

Interaction effects. H2c investigates the moderating influence of a consumer’s position in the

(33)

compared to the post-purchase stage. Also the effect of medium DCP(brand) on purchase intention is weaker when a consumer is in the consideration stage compared to the post-purchase stage, β=-0.918,

p<0.05. Performing multiple regressions with other reference categories reveal the effect of medium

DCP(brand) on purchase intention is stronger when a consumer is in the post-purchase stage than the information stage, β=1.234, p<0.01, meaning medium DCP(brand) works best in terms of purchase intention when a consumer is in the post-purchase stage. It is also found that the effect of medium DCP(category) on purchase intention is stronger when a consumer is in the information stage,

β=1.446, p<0.10, and the post-purchase stage, β=0.666, p<0.10, compared to the consideration stage,

indicating medium DCP(category) works best when a consumer is in the information and the post-purchase stage. Privacy concern is hypothesized in H2d to weaken the influence of DCP on post-purchase intention. The results partially support this hypothesis. Privacy concern only moderates the effect of medium DCP(category) on purchase intention, β=-0.551, p<0.01, in such a way that consumers high in privacy concern will have lower purchase intention for banners that have a medium (category) level of DCP than consumers low in privacy concern. H2e proposes that NFC strengthens the effect of DCP on purchase intention. None of the interactions are found to be significant, p>0.10, so H2e is not supported. As hypothesized in H2f, NFU strengthens the effect of DCP on purchase intention. The results support H2f. A significant interaction effect is found for high DCP and NFU, and as expected, this effect is positive, β=0.542, p<0.01. NFU therefore strengthens the influence of high DCP on purchase intention.

When taking into account the strength of effects (so looking at the standardized coefficients), the results show the effect of high DCP is most strong on purchase intention, β=0.604, followed by medium DCP(brand), β=0.376, and the interaction effect of the information stage and medium DCP(brand), β=-0.327. High and medium DCP(brand) thus have the strongest effect on consumer’s purchase intentions, followed by the interaction of the consumer being in the information stage opposed to the post-purchase stage and medium DCP(brand). Next, medium DCP(category), β=0.266, and the interaction effect of privacy concern and medium DCP(category), β=-0.265, follow. Last, age,

β=0.262, the interaction of NFU and high DCP, β=-0.224, the interaction of the consideration stage

(34)

5. DISCUSSION

As a solution to the recent surge in online advertising noise, firms increasingly use retargeting

to tailor banner ads to individual consumers in an attempt to increase their advertising

effectiveness (Helft & Vega, 2010, Peterson, 2013; Sengupta, 2013). Nevertheless, this

special form of personalization is

found to be effective only under certain circumstances (Lambrecht & Tucker, 2013) and for certain consumer characteristics (Ho et al., 2008). This research therefore investigates when and for whom personalized retargeting works best by examining the moderating role of timing and consumer heterogeneity. In particular, an exploratory online study is conducted to assess the effect of personalization on the click-through and purchase intention of a banner ad while taking into account the effect of the consumer’s position in the purchase decision process and the following consumer characteristics; privacy concern, need for cognition and need for uniqueness.

(35)

also found that individuals high in need for uniqueness have higher purchase intention than people low in need for uniqueness. An explanation can be found in the study of Knight & Kim (2007). They found that in the apparel industry individuals high in need for uniqueness desire brands that are of high quality, reputable and prestigious. Also individuals high in need for uniqueness have higher purchase intentions for brands with emotional value and unique identities. The products represented in the banners used in this study match these type of brands; They are all high quality brands (Adidas, Converse, Nike) with a distinct brand image, which explains the higher purchase intention for consumers high in need for uniqueness.

A consumer’s position in the purchase decision process is an important moderator for personalized banner effectiveness. Initially personalization was hypothesized to be most effective in the consideration stage, where consumer’s preferences are most narrowly construed and thus they will respond better to advertising messages that addresses such goals. The results show that medium DCP works best in the information stage and the post-purchase stage. An explanation is found in the preference construction literature. As the event of making a purchase comes closer, consumer’s preferences indeed become more narrowly construed, but they also become more stable (Hoeffler & Ariely, 1999), which makes them less sensitive to a firm’s recommendations (Simonson, 2005). At the beginning of the purchase decision process consumers only have a broad idea of what they want (Lambrecht & Tucker, 2013) and preferences are unstable (Hoeffler & Ariely, 1999), making them highly susceptible to a firm’s recommendations (Simonson, 2005). In the post-purchase stage however, consumers have stable preferences (Hoeffler & Ariely, 1999) and thus should not be susceptive to personalized ads. Results show they are. This can be explained by the assumption that consumers in the post-purchase stage just purchased the product they want, and subsequently this purchase triggers new demand for products, for instance for cross sells or complementary products. Especially medium personalization based on brand works best in this stage, indicating consumers are satisfied with the product they bought and those feelings of satisfaction result in higher purchase intention for products of the same brand.

(36)

6. CONCLUSIONS AND RECOMMENDATIONS

The current study investigated the effect of varying personalization intensities on click-through and purchase intention while taking into account the moderating role of a consumer’s position in the purchase decision process and consumer characteristics. It showed that personalization of banners increases click-through and purchase intention, with relatively high personalization having the most impact. The inclusion of a timing factor learnt that medium personalization works best when a consumer is in the information or the post-purchase stage. Taking into account consumer characteristics showed that consumers who care about their privacy might respond less favorably to personalized communication, but it can attract individuals with a high need for uniqueness. These findings provide clear implications for practice and several limitations should be considered while interpreting the findings, but they can also be regarded as suggestions for future research.

Managerial implications

(37)

consumers. In doing so, they need to assess consumers’ need for uniqueness and target those people who are high in need for uniqueness with highly personalized ads. This strategy would work even better if those ads contain brands that are of high quality, reputable, prestigious and have a distinct brand image (Knight & Kim, 2007). Feelings of uniqueness could be triggered even more by applying content personalization in more ways than just inferring consumers’ future needs from previous online behavior. One example is to use self-referent personalization (Bragga, Sunikka, & Kallio, 2012), that is, using personal information, such as a consumer’s name, to personalize the content of the advertisement. Also both personal information and past online behavior can be used to personalize the advertisement.

Limitations and future research

(38)

7. REFERENCES

Aguirre, E., Mahr, D., Grewal, D., De Ruyter, K., & Wetzels, M. (2015). Unraveling the Personalization Paradox: The Effect of Information Collection and Trust-Building Strategies on Online Advertisement Effectiveness. Journal of Retailing, 91(1), 34-49.

Ansari, A., & Mela, C. F. (2003). E-Customization. Journal of Marketing Research, 60(2), 131–145.

Bagozzi, R. P., Yi, Y., & Philips, L. W. (1991). Assessing Construct Validity in Organizational Research. Administrative Science Quarterly, 36(3), 421-458.

Bleier, A., & Eisenbeiss, M. (2015a). Personalized Online Advertising Effectiveness: The Interplay of What, When, and Where. Marketing Science, 34(5), 669-688.

Bleier, A., & Eisenbeiss, M. (2015b). The Importance of Trust for Personalized Online Advertising.

Journal of Retailing, 91(3), 390-409.

Bosnjak, M., Galesic, M., & Tuten, T. (2007). Personality determinants of online shopping: Explaining online purchase intentions using a hierarchical approach. Jouurnal of Business

Research, 60, 597-605.

Box, G. E., & Cox, D. R. (1964). An analysis of transformations. Journal of the Royal Statistical

Society. Series B (Methodological), 211-252.

Bragga, J., Sunikka, A., & Kallio, H. (2012). An Exploratory Study on Customer Responses to Personalized Banner Messages in the Online Banking Context. Journal of Information

Technology Theory and Apllication, 13(3), 5-20.

Braun, M., & Moe, W. W. (2013). Online Display Advertising: Modeling the Effects of Multiple Creatives and Individual Impression Histories. Marketing Science, 32(5), 753-767.

Breusch, T. S., & Pagan, A. R. (1979). A simple test for heteroskedasticity and random coefficient variation. Econometrica, 47, 987-1007.

Cacioppo, J. T., & Petty, R. E. (1982). The Need for Cognition. Journal of Personality and Social

When and whom to retarget?