1
THE EFFECTIVENESS OF
FIRM-INITIATED TOUCHPOINTS, AND THE
MODERATING ROLE OF AGE AND
COMPARISON WEBSITES
by
MARCO BROUWER
University of Groningen
Faculty of Economics and Business
MSc Marketing
2
THE EFFECTIVENESS OF
FIRM-INITIATED TOUCHPOINTS, AND THE
MODERATING ROLE OF AGE AND
COMPARISON WEBSITES
Master Thesis Marketing Intelligence 15-06-2020 Author Marco Brouwer m.a.e.brouwer@student.rug.nl S3797767 First supervisor
Dr. P.S. (Peter) van Eck p.s.van.eck@rug.nl Second supervisor dr. A. (Abhi) Bhattacharya abhi.bhattacharya@rug.nl
University of Groningen
Faculty of Economics and Business Department of Marketing
PO Box 800
3
SUMMARY
By the rise of the Internet many new, attractive marketing communication channels have been introduced (Danaher & Dagger, 2013). However, because of the rise of new channels, it also becomes more and more complex for firms to control and keep an eye on the customer experience (Verhoef, Kannan & Inman, 2015). While complexity for firms increases, it is proven that understanding of the customer journey could lead to higher conversion rates and improve customer loyalty (Lemon & Verhoef, 2016).
Although various studies investigated the customer journey, there are still some underexplored topics. Based on previous literature, it seems the supply side of firm-initiated touchpoints is not sufficiently considered. To contribute to existing literature and provide deeper insights in advertising effectiveness, in the present study, the effectiveness of affiliate marketing, banner advertising, e-mail marketing, pre-roll advertising, and retargeting was considered. The effectiveness was measured by means of probabilities consumers enter the visit stage and probabilities consumers enter the purchase stage. Furthermore, the effect of age on advertising effectiveness and the effect of comparison websites on retargeting effectiveness were examined. Therefore, this paper sought to answer the following research question:
• What is the effect of different types of firm-initiated touchpoints on the visit stage and the purchase stage, the effect of age on advertising effectiveness, and the effect of comparison websites on retargeting effectiveness?
The effects were assessed by analyzing a large dataset provided by GfK, covering data on 24,985 customer journeys in the travel industry. The event-based, online data entailed information on both the browsing behavior as demographical characteristics of consumers. Generalized Linear Models (GLM) were estimated to analyze these data and to provide answers to the research question.
4 This study suggests, based on its limitations, ideas for future research and recommends marketing managers to take the findings into account when allocating their marketing budgets and tailoring their marketing efforts.
5
PREFACE
Dear reader,
This master thesis is the final report of my Master Marketing Intelligence at the University of Groningen. It finalizes my student time at the University and serves as a start for a career in business. The process of writing a master thesis was a challenging and educational process, in which I learned a lot about how to work with “big data”.
I would like to take this opportunity to express my gratitude towards the people that have supported me during this process. First, I would like to thank my supervisor dr. P. van Eck, for the time he has taken to read my work and provide me with valuable feedback and guidance. Furthermore, I would also like to thank my second supervisor dr. A. Bhattacharya, who took the time to evaluate my master thesis.
Finally, I would like to thank my family and friends, for the support they provided me with during my time as a student at the University of Groningen. My parents deserve a particular note of thanks: your encouraging support helped me a lot.
6
TABLE OF CONTENT
1. INTRODUCTION ... 1
2. LITERATURE REVIEW ... 4
2.1 Firm-initiated touchpoints and firm outcomes ... 4
2.2 Purchase process ... 6
2.3 Age-related differences and advertising ... 8
2.4 Comparison websites and retargeting ... 9
2.5 Conceptual model ... 9 3. METHODOLOGY ... 11 3.1 Data ... 11 3.2 Variables ... 11 3.2.1 Visit stage ... 11 3.2.2 Purchase stage ... 12 3.2.3 Control variables ... 13
3.3 Binomial logistic regression ... 13
3.4 Model specification ... 14
3.5 Analysis plan ... 16
4. RESULTS ... 17
4.1 Preliminary checks ... 17
4.2 Descriptive statistics ... 18
4.3 Assumptions for binomial logistic regression ... 20
4.4 Model selection ... 21 4.5 Testing hypotheses ... 23 4.5.1 Visit stage ... 23 4.5.2 Purchase stage ... 25 4.6 Re-estimation ... 27 4.6.1 Visit stage ... 27 4.6.2 Purchase stage ... 29 4.7 Support of hypotheses ... 29 5. DISCUSSION ... 31
5.1 Summary and discussion of results ... 31
5.2 Implications ... 34
5.3 Limitations and future research ... 36
6. REFERENCES ... 38
1
1. INTRODUCTION
Advertising in the Netherlands has grown from €3.169 billion in 2012 to €3.565 billion in 2017 (Deloitte, 2018). Figures show that especially the market for online advertising grew considerably. By the rise of the Internet many new, attractive marketing communication channels have been introduced (Danaher & Dagger, 2013). According to the annual Online Ad Spend Study, the online advertising market has grown by 9% in 2018. This means that this market outperforms all other traditional medium types combined (Deloitte, 2019). The digital market accounts for €1.832 billion of the total amount spend on advertisements, which comes down to a proportion of 51,4%. Furthermore, it is expected that internet advertising will account for even a greater proportion of the total spending on advertising over the next few years (PwC, 2019). PwC (2019) predicts that internet advertising will reach €2.6 billion in the Netherlands by the time it is 2023, expecting it to grow at an average growth rate of 5.7% (PwC, 2019). Internet becoming the main advertising channel for companies is caused by the shift of offline consumers towards the digital environment (Bughin, 2015). However, as a result of the rise of new channels such as mobile channels and social media, it also becomes more and more complex for firms to control and keep an eye on the customer experience (Verhoef, Kannan & Inman, 2015). A concept that helps to provide insights into the customer experience is the so-called “customer journey”, which refers to the path the customer takes to their decision to purchase an item (Anderl, Schumann & Kunz, 2016). This path consists of multiple “touchpoints”, which are all interactions between the firm and the customer occurring at different moments in time (Homburg, Jozic & Kuehln, 2015). For companies it is essential to gain understanding of the customer experience, as it is proven it could lead to higher conversion rates and improve customer loyalty (Lemon & Verhoef, 2016). Bughin (2015) even found that companies with a high level of digital understanding reach a conversion rate which is 2.5 times greater than companies with a low level of digital understanding. Literature confirms that companies that master the digital channels are gaining a great competitive advantage over companies that do not master it (Bughin, 2015).
2 the effectiveness of different forms of advertising for traffic, conversion, and revenue contribution. Their research did not consider the effect of banner advertising and pre-roll advertising, despite there is a lot of discussion on the true effects of these advertising forms in recent literature (e.g. Muñoz-Leiva, Hernández-Méndez and Gómez-Carmona, 2019; Campbell, Mattison Thompson, Grimm & Robson, 2017; Ghose & Todri, 2015). Based on the above, in this paper the effect of several firm-initiated touchpoints on two stages of the purchase funnel will be investigated.
Additionally, existing studies did not account for heterogeneity across customers. To the best of my knowledge, current research has not investigated the role of demographic characteristics of consumers for the effectiveness of advertising. Yet, previous research reveals demographic factors do affect the responses of individuals on advertising stimuli (Ansu-Mensah & Asuamah, 2013). One of the demographic variables that affect responses to advertisements is age (McKay-Nesbitt, Manchanda, Smith & Huhmann, 2011). Age-related differences play a role in information processing (Williams & Drolet, 2005) and the perceptions on advertising of individuals (Philips & Stanton, 2004). Therefore, this study seeks for greater understanding of the role of age for the effectiveness of different types of firm-initiated touchpoints.
Furthermore, there has not been a study that investigated the effect of comparison websites on advertising effectiveness. Laffey and Gandy (2009) state comparison websites generate visitors further in the purchasing process, whereas Lambrecht and Tucker (2013) argue retargeting is more effective for consumers further in the purchasing process. Despite the related conclusions, there are no studies that examined comparison websites as a moderator on retargeting effectiveness.
This study contributes to the existing literature since it considers advertising forms that have not been considered in a single study yet. It provides insights in the relative effectiveness of the advertising forms. Furthermore, it focuses on multiple stages of the purchase funnel (visit stage and purchase stage), while previous research related to advertising effectiveness mainly focused on only one outcome (e.g. sales) or funnel stage. On top of this, the present study investigates the role of age for different types of advertising forms and the role of comparison websites on retargeting effectiveness. These issues have not been studied in existing literature yet.
3 have limited resources, for marketeers it is important to understand the relative effectiveness of different touchpoints (Baxendale, Macdonald & Wilson, 2015). As in the purchase journey different phases precede a purchase, firms may be interested in the effects of different forms of advertising on different stages of the purchase funnel. Insights on the effect of age and comparison websites on advertising may help advertisers to tailor their marketing efforts more effectively.
This study will provide answers to the following research question:
• What is the effect of different types of firm-initiated touchpoints on the visit stage and the purchase stage, the effect of age on advertising effectiveness, and the effect of comparison websites on retargeting effectiveness?
4
2. LITERATURE REVIEW
2.1 Firm-initiated touchpoints and firm outcomes
A customer journey consists of all interactions between the firm and the customer occurring at different moments in time (Homburg, Jozic & Kuehln, 2015). All these digital interactions between a firm and a customer can be categorized into customer-initiated touchpoints or firm-initiated touchpoints (Bowman & Narayandas, 2001). Customer-firm-initiated contacts refer to moments of contact with a firm, initiated by the customer. Firms do not control these interactions (Wiesel, Pauwels & Arts, 2011). By contrast, firm-initiated touchpoints are controllable by firms. They can be defined as the marketing communications that are initiated by the firm (Wiesel et al., 2011).
Firms could generate higher conversion rates and improve customer loyalty with greater understanding of the customer journey (Lemon & Verhoef, 2016). Li and Kannan (2014) especially emphasize the importance of firm-initiated touchpoints for companies. Firms may benefit from firm-initiated touchpoints, as they allow firms to communicate messages and increase the involvement of the consumer with the brand (Li & Kannan, 2014). Firm-initiated touchpoints can cause consumers to consider the brand, which they might not do without the touchpoints (Li & Kannan, 2014).
However, while firm-initiated touchpoints may increase conversion and brand awareness, according to Li and Kannan (2014), this side of touchpoints is insufficiently considered in existing literature (Li & Kannan, 2014). Furthermore, among managers there are major concerns about the true effects of the touchpoints on firm outcomes (Hoban & Bucklin, 2015). Therefore, the focus of the present study will be on five types of firm-initiated touchpoints, namely: affiliate marketing, banner advertising, e-mail marketing, pre-roll advertising, and retargeting. These firm-initiated touchpoints (or advertising forms) will be discussed in more detail below.
Affiliate marketing
5 website, the partner gets a compensation (Libai, Biyalogorsky & Gerstner, 2003). Libai et al. (2003) describe affiliate marketing is especially useful for the acquisition of customers. Firms can create networks of affiliates in order to drive traffic to their websites (Libai et al., 2003). Moreover, affiliate marketing has become one of the most important methods to increase website traffic and sales (Gregori, Daniele & Altinay, 2014; Fox & Wareham, 2012).
Banner advertising
The most common form of digital advertising is banner advertising (Manchanda, Dubé, Goh, & Chintagunta, 2006; Cho, Lee & Tarp, 2001). Banner advertisements are combinations of visual and textual content that include a link to the website of the advertiser. Urban, Liberali, MacDonald, Bordley and Hauser (2014) mention banner advertisements are used to maximize click-through rates and purchase probabilities. This is partly confirmed by a study by Manchanda et al. (2006). These researchers found a significant increase in purchasing probabilities when using banner advertisements (Manchanda et al., 2006). Furthermore, Dreze and Hussherr (2003) argue banner advertising has positive effects on both brand awareness and advertisement recall. However, managers and more recent literature question whether banners truly affect customer behavior (Hoban & Bucklin, 2015). According to Muñoz-Leiva, Hernández-Méndez and Gómez-Carmona (2019) banners can lead to negative attitudes towards the advertised products, services, websites, and brands. Consumers tend to quickly filter banner ads from the content they intended to see. As a result, consumers do usually not recall the brand since they avoid the banners or do not pay attention to the banners (Muñoz-Leiva et al., 2019). E-mail marketing
6 Melville and Plice (2008) state that the return on investment of e-mail marketing is at least twice as high as compared to the ROI of other advertising forms.
Pre-roll advertising
A relatively novel form of advertising is pre-roll advertising. Pre-roll advertisements can be defined as video advertisements consumers get exposed to before watching the content they intended to watch (Campbell et al., 2017). The effectiveness of pre-rolls is much disputed. Research conducted by YuMe (2017) revealed pre-rolls are effective for increasing various metrics, such as brand awareness and purchasing intentions. Yet, Campbell et al. (2017) question whether pre-rolls have a positive influence on the purchasing intentions of consumers. The authors state pre-rolls cause consumer irritation and annoyance, since consumers are forced to watch a part of the advertisement before they can watch the content they intended to watch (Campbell et al., 2017). The concern about the effectiveness of pre-roll advertising is also raised by Ghose and Todri (2015). Their study showed pre-roll advertising had a strong positive effect on visits, while it did not increase the propensity of consumers to convert (Ghose & Todri, 2015).
Retargeting
An advertising form that uses consumers’ past browsing behavior is retargeting (Lambrecht & Tucker, 2013). Retargeting advertisements are defined as advertisements that are displayed on third-party websites and are related to previous visits of the customer (De Haan et al., 2016). The ads focus on customers that have already visited a product page on a firm’s website. Lambrecht and Tucker (2013) studied the effectiveness of the ads and found that the retargeted advertisements are more effective when the product preferences of the customer are evolved (e.g. by visiting review websites). The authors found that retargeting has a positive impact on firm outcomes, however, on average it is less effective than displaying generic ads (Lambrecht & Tucker, 2013). Furthermore, De Haan et al. (2016) studied the effect of retargeting on traffic, conversion, and revenue contribution. The study pointed out that the effect of this advertising form varies for different product categories (De Haan et al., 2016).
2.2 Purchase process
7 information about different products. They familiarize themselves with the offerings. Subsequently, in the purchase phase the gathered information is used to decide on the purchase. Finally, the consumers get into the post-purchase phase. This phase entails all touchpoints after the purchase has been made (Frambach, Roest & Krishnan, 2007).
Li and Kannan (2014) use a funnel framework to model online purchase decisions. This framework involves three different stages. The first stage is referred to as “the consideration stage”. In this stage the customer recognizes his or her needs and accordingly looks for channels to seek information. Then, the customer will go through “the visit stage”, in which the customer visits the website to gather information and evaluate the alternatives (Li & Kannan, 2014: p.42). Hence, a consumer enters the visit stage as soon as it visits the firm’s website. Therefore, in the present study “a visit” refers to a consumer visiting the website of the focal brand. Finally, the last stage is “the purchase stage”. In this stage, the customer makes the actual purchase (Li & Kannan, 2014). Following this definition, in the present study, “a purchase” occurs when a consumer makes an actual purchase on the website of the firm.
8 the present study expects that all advertising forms positively influence the probability consumers enter the visit stage and the purchase stage. This resulted in the following hypotheses:
• Visit stage
H1A: Affiliate marketing positively influences the probability a consumer enters the visit stage. H1B: Banners positively influence the probability a consumer enters the visit stage.
H1C: E-mail marketing positively influences the probability a consumer enters the visit stage. H1D: Pre-rolls positively influence the probability a consumer enters the visit stage.
H1E: Retargeting positively influences the probability a consumer enters the visit stage.
• Purchase stage
H2A: Affiliate marketing positively influences the probability a consumer enters the purchase stage. H2B: Banners positively influence the probability a consumer enters the purchase stage.
H2C: E-mail marketing positively influences the probability a consumer enters the purchase stage. H2D: Pre-rolls positively influence the probability a consumer enters the purchase stage.
H2E: Retargeting positively influences the probability a consumer enters the purchase stage.
2.3 Age-related differences and advertising
Previous research has shown that demographic factors affect the response of individuals to advertising stimuli (Ansu-Mensah & Asuamah, 2013). One variable that affects responses to advertising is age. Williams and Drolet (2005) state that young and older adults have different modes of information processing of emotions. The researchers found that ads focusing on emotions were more liked and recalled among older consumers. A study by McKay-Nesbitt et al. (2011) revealed younger adults recall emotional ads better than rational ads. Furthermore, Philips and Stanton (2004) found age-related differences affect perceptions of advertising. In their study, younger consumers were less likely to be persuaded by information in the ad, compared to older consumers (Philips & Stanton, 2004). All in all, from prior research one can derive age has an impact on how consumers respond to advertisements.
9 the effect of differences in demographics on advertising might help firms organizing more effective advertising campaigns. Therefore, in the present study the role of age for advertising effectiveness will be investigated. It is unclear from prior research what the expected effect of age on website traffic is. Thus, the following hypotheses are drawn:
H3A: Age moderates the effect affiliate marketing has on the visit stage. H3B: Age moderates the effect banners have on the visit stage.
H3C: Age moderates the effect e-mail marketing has on the visit stage. H3D: Age moderates the effect pre-rolls have on the visit stage. H3E: Age moderates the effect retargeting has on the visit stage.
2.4 Comparison websites and retargeting
A technology that helps consumers to sort all information on the Internet and compare alternatives are comparison websites (Paraskevas & Kontoyiannis, 2005). Firms also benefit from comparison websites; the comparing process drives customers who are more likely to convert towards the websites of the firms (Laffey & Gandy, 2009). Thus, comparison websites generate website visitors that are further in the purchasing process. This might create advertising opportunities for firms, as Lambrecht and Tucker (2013) argue retargeting advertisements are more effective when product preferences of customers are evolved. Lambrecht and Tucker (2013) assumed in their study that customers who had touchpoints of review websites in their journey were further in the purchasing process. While Laffey and Gandy (2009) state comparison websites generate visitors further in the purchasing process, did Lambrecht and Tucker (2013) not consider comparison website related touchpoints. Their study solely focused on the effect of review website touchpoints. Therefore, the current study will examine the moderating effect of comparison websites on the effectiveness of retargeting. Based on the above-mentioned literature, expectations are that comparison websites positively influence the effectiveness of retargeting. This leads to the following hypothesis:
H4A: Comparison websites positively moderate the effect retargeting has on the probability a consumer enters the purchase stage.
2.5 Conceptual model
10 age and the moderating role of comparison websites will be examined. The moderating effect of age will be examined for the visit stage, whereas the moderating effect of comparison websites will be examined for the purchase stage.
11
3. METHODOLOGY
3.1 Data
In order to test the hypotheses, consumer purchase journey data from a Dutch travel agency will be analyzed. The event-based data are provided by GfK, which is a German market research institute. The data were collected in the time period of the 1st of June 2015 until the 31st of September 2016 via GfK Crossmedia Link. The data entails quantitative panel data. A passive measurement system (plug-in) was used to gather data on the browsing behavior of panelists. Complete purchase journeys of panelists were tracked and gathered by the system of GfK. In the dataset each observation represents one touchpoint of the panelist, which can be either customer-initiated or firm-initiated. The sequence of multiple touchpoints demonstrates the purchase journey of the panelist. In the dataset, 29,011 orientations (or purchase journeys) are measured. From all purchase journeys, 3,674 observations ended with a purchase. Moreover, from those 3,674 observations, 192 purchases were done at the focal brand. Next to data on the browsing behavior, there is also data available on demographics. The dataset includes information on demographical variables of 9,678 panelists. It concerns demographics such as the age, gender, income, and education of the panelists.
3.2 Variables
As mentioned before, this study focuses on two stages of the purchase funnel of Li and Kannan (2014). It focuses on both the visit stage and the purchase stage. Therefore, two different models will be specified. The first model focuses on the visit stage and the second model on the purchase stage. The next section will consider the main variables for both stages. Subsequently, the variables that will be controlled in both models will be given.
3.2.1 Visit stage
The first dependent variable is Visit, which represents whether the consumer visited the website of the focal brand. This binary variable will have a value of ‘1’ if a consumer did visit the website of the focal brand in his/her purchase journey and will have a value of ‘0’ if the consumer did not visit it.
firm-12 initiated touchpoints (or advertising forms) are called Affiliates, Banners, Emails, Prerolls, and Retargeting.
In the first model the variable Age is considered as the moderating variable. The moderating effect of this variable will be examined for all five firm-initiated touchpoints. An overview of the operationalization of the main variables of model 1 is shown in Table 1 below.
TABLE 1
Operationalization main variables visit stage
Variables Operationalization
Dependent variable Visit Whether or not the consumer
visited the website of the focal brand
Independent variables Affiliates Banners Emails Prerolls Retargeting
Number of times the firm-initiated touchpoint occurred in the purchase journey
Moderator variable Age Age of the consumer (in years)
3.2.2 Purchase stage
The second dependent variable is Purchase and represents whether the purchase journey of the consumer ended with a purchase at the focal brand. The variable will take a value of ‘1’ if a purchase at the focal brand occurred and ‘0’ if not.
13 TABLE 2
Operationalization main variables purchase stage
Variables Operationalization
Dependent variable Purchase Whether or not the purchase
journey of the consumer ended with a purchase at the focal brand
Independent variables Affiliates Banners Emails Prerolls Retargeting
Number of times the firm-initiated touchpoint occurred in the purchase journey
Moderator variable Comparison Number of times the consumer
visited a comparison website in the purchase journey
3.2.3 Control variables
To rule out alternative explanations (Becker, 2005) and adjust for the effect of external variables on the findings (Leeflang, Wieringa, Bijmolt & Pauwels, 2016), several control variables will be considered. Thus, to increase statistical control (Kutner, Nachtsheim, Neter & Li, 2005) the factors gender, income and, education will be included in the primary analyses. The control variables will be included in both models. In Table 3 an overview of the included control variables and the academic reasoning behind the variables is given.
TABLE 3 Control variables Control variables Academic reasoning Gender
Income Education
Women have much more positive attitudes toward general shopping than men (Alreck & Settle, 2002), whereas men have more favorable perceptions of online shopping than women have (Van Slyke, Comunale & Belanger, 2002).
Online shoppers have a higher income than traditional store shoppers (Zhou, Dai & Zhang, 2007).
Comparing online shoppers to online non-shoppers, online shoppers are wealthier and better educated (Swinyard & Smith, 2003; Li, Kuo & Rusell, 1999).
3.3 Binomial logistic regression
14 consumer. The models that fit marketing problems with binomial response variables best are logit models and probit models (Leeflang et al., 2015). The two types of models are rather similar. Both models are estimated by means of Maximum Likelihood Estimation methods and focus on probabilities an event occurs instead of on observed values. However, due to the mathematical convenience and easier interpretation of the logit model, this model is often preferred (Leeflang et al. 2015). Furthermore, the use of logit models allows for the addition of moderator variables, by including interaction effects. Therefore, in the present study, binomial logistic regression will be applied.
3.4 Model specification
Binary logit models estimate the probability an observation belongs to each group (Malhotra, 2007). Hence, the dependent variables must take a value between 0 and 1 (Hilbe, 2009). The variables need to be non-logistic and thus need to be exponentiated. Equation 3.4.1 shows the structure that ensures the predictor values lie between 0 and 1, provided by Malhotra (2007) and Hilbe (2009).
EQUATION 3.4.1
Binomial regression structure (Malhotra, 2007; Hilbe, 2009) 𝜋𝑖=
exp{𝑥𝑖′𝛽}
1 + exp{𝑥𝑖′𝛽}
Which, according to Allison (2012), can be rewritten as the following equation: EQUATION 3.4.2
Rewritten binomial regression structure (Allison, 2012) 𝜋𝑖=
1
1 + exp(−{𝑥𝑖′𝛽})
15 EQUATION 3.4.3 Model 1 𝑃𝑉𝑖 = 1 1 + exp ( − ( 𝛽0+ 𝛽1𝐴𝐹𝑖+ 𝛽2 𝐵𝐴𝑖+ 𝛽3 𝐸𝑀𝑖+ 𝛽4 𝑃𝑅𝑖+ 𝛽5 𝑅𝐸𝑖 + 𝛽6 𝐴𝐺𝐸𝑖+ 𝛽7 (𝐴𝐹 ∗ 𝐴𝐺𝐸)𝑖+ 𝛽8 (𝐵𝐴 ∗ 𝐴𝐺𝐸)𝑖 + 𝛽9 (𝐸𝑀 ∗ 𝐴𝐺𝐸)𝑖+ 𝛽10 (𝑃𝑅 ∗ 𝐴𝐺𝐸)𝑖+ 𝛽11 (𝑅𝐸 ∗ 𝐴𝐺𝐸)𝑖 + 𝛽12 𝐺𝐸𝑖+ 𝛽13 𝐼𝑁𝐶𝑖+ 𝛽14 𝐸𝐷𝑈𝑖 ) ) Where:
𝑃𝑉𝑖 = Probability consumer visits the website of focal brand in purchase journey i;
𝐴𝐹𝑖 = Count variable for affiliate marketing in purchase journey i;
𝐵𝐴𝑖 = Count variable for banner advertising in purchase journey i;
𝐸𝑀𝑖 = Count variable for e-mail marketing in purchase journey i;
𝑃𝑅𝑖 = Count variable for pre-roll advertising in purchase journey i;
𝑅𝐸𝑖 = Count variable for retargeting in purchase journey i;
𝐴𝐺𝐸𝑖 = Age of the consumer in purchase journey i;
(𝐴𝐹 ∗ 𝐴𝐺𝐸)𝑖 = Interaction variable for affiliate marketing and age of user in purchase journey i;
(𝐵𝐴 ∗ 𝐴𝐺𝐸)𝑖 = Interaction variable for banner advertising and age of user in purchase journey i; (𝐸𝑀 ∗ 𝐴𝐺𝐸)𝑖 = Interaction variable for e-mail marketing and age of user in purchase journey i;
(𝑃𝑅 ∗ 𝐴𝐺𝐸)𝑖 = Interaction variable for pre-roll advertising and age of user in purchase journey i;
(𝑅𝐸 ∗ 𝐴𝐺𝐸)𝑖 = Interaction variable for retargeting and age of user in purchase journey i; 𝐺𝐸𝑖 = Gender of user in purchase journey i;
𝐼𝑁𝐶𝑖 = Income of user in purchase journey i;
𝐸𝐷𝑈𝑖 = Education of user in purchase journey i. EQUATION 3.4.4 Model 2 𝑃𝑃𝑖 = 1 1 + exp (− ( 𝛽0+ 𝛽1 𝐴𝐹𝑖+ 𝛽2 𝐵𝐴𝑖 + 𝛽3 𝐸𝑀𝑖+ 𝛽4 𝑃𝑅𝑖 + 𝛽5 𝑅𝐸𝑖 + 𝛽6 𝐶𝑊𝑖+ 𝛽7 (𝑅𝐸 ∗ 𝐶𝑊)𝑖+ 𝛽8 𝐺𝐸𝑖+ 𝛽9 𝐼𝑁𝐶𝑖+ 𝛽10 𝐸𝐷𝑈𝑖)) Where:
𝑃𝑃𝑖 = Probability consumer purchases at the website of focal brand in purchase journey i;
𝐴𝐹𝑖 = Count variable for affiliate marketing in purchase journey i;
𝐵𝐴𝑖 = Count variable for banner advertising in purchase journey i;
𝐸𝑀𝑖 = Count variable for e-mail marketing in purchase journey i;
𝑃𝑅𝑖 = Count variable for pre-roll advertising in purchase journey i;
𝑅𝐸𝑖 = Count variable for retargeting in purchase journey i;
𝐶𝑊𝑖 = Count variable for comparison website in purchase journey i;
(𝑅𝐸 ∗ 𝐶𝑊)𝑖 = Interaction variable for retargeting and comparison website in purchase journey i;
𝐺𝐸𝑖 = Gender of user in purchase journey i;
𝐼𝑁𝐶𝑖 = Income of user in purchase journey i;
16 3.5 Analysis plan
17
4. RESULTS
4.1 Preliminary checks
First, the data were checked for missing responses. Missing responses are values of a variable that are unknown, due to ambiguous answers of respondents or no proper recordings of the answers (Malhotra, 2007). In the dataset related to the demographics of the panelists, several missing values were found. For the variables Age, Gender, and Income 1,603 missing values were found, whereas for Education 2,031 missing values were found. A possible explanation for these missings could be that the panelists did not consistently participate. Moreover, 1,338 panelists refused or were not able to answer the question related to Income. This resulted in a total of 2,941 missing values for Income. Since for 1,603 panelists there is no data on demographics at all, these missing values are removed from the dataset. After this listwise deletion of the missing values, data of 8,075 panelists were left. The remaining missings for Education (428) and Income (1,338) were imputed applying Multiple Imputation (MICE-package in R). The widely used method that has been used to impute the missings is predictive mean matching (PMM), which imputes missings by means of nearest-neighbor algorithms. It matches the outcome variable with the outcome of the missing variables (Vink, Frank, Pannekoek & Van Buuren, 2014).
18
FIGURE 2 FIGURE 3
Number of touchpoints (before removal) Number of touchpoints (after removal)
Lastly, multicollinearity between predictor variables was checked. If multicollinearity occurs, two or more dependent vectors are highly linearly correlated. The presence of multicollinearity leads to unreliable parameter estimates (Leeflang et al., 2015). To check whether multicollinearity was present, a linear model including the relevant main variables were estimated. Subsequently, VIF scores were assessed. In general, it is assumed a VIF score greater than 5 indicates collinearity is an issue (Leeflang et al., 2015). In Table 4 the VIF scores of all predictor variables are shown. None of the VIF scores exceeded the threshold of 5, thus, one can assume multicollinearity was not an issue for the analyses.
TABLE 4
VIF scores predictor variables
Variable VIF score
Affiliate Banners Emails Prerolls Retargeting Age Comparison Gender Income Education 1.002 1.017 1.046 1.018 1.074 1.131 1.044 1.075 1.199 1.260 4.2 Descriptive statistics
19 is 17 years old and the oldest is 94 years old. The income class that occurred most frequently is the class of €40.000 between €67.000, which comes down to an income that is between 1 and 2 times the average yearly income (see Figure 6). In Figure 7 the distribution of the education of the respondents is shown. Education class 4 occurred in 29% of the purchase journeys, which stands for an education level of at least MBO.
FIGURE 4 FIGURE 5
Distribution of gender Distribution of age
FIGURE 6 FIGURE 7
Distribution of income level Distribution of education level
20 touchpoints occurred on average in less than 10% of the purchase journeys. Furthermore, comparison websites were visited 5.8 times on average. The maximum value of comparison websites is 1,159, meaning in one purchase journey 1,159 comparison websites touchpoints occurred.
TABLE 5 Descriptive statistics
Variable Min Max Mean Standard deviation
Visit Purchase Affiliates Banners Emails Prerolls Retargeting Age Comparison Gender Income Education 0 0 0 0 0 0 0 17 0 1 1 1 1 1 79 69 135 124 1219 94 1159 2 7 8 0.105 0.006 0.045 0.063 0.087 0.063 0.903 51.930 5.828 1.597 3.804 4.722 0.306 0.078 0.941 1.110 1.838 1.032 17.941 15.778 29.363 0.491 1.595 1.867
4.3 Assumptions for binomial logistic regression
As opposed to regression models that are based on Ordinary Least Squares (OLS), logistic regression does not require many of the principle assumptions such as non-normality and homoscedasticity. However, according to Park (2016) and Bewick, Cheek and Ball (2005) some assumptions still apply.
First, for logistic regression the dependent variables must be dichotomous. Since logistic regression focuses on probabilities an event occurs, the variable should only take two different values. The desired event should be coded as a ‘1’ (Park, 2016). In the present study, both a visit and a purchase are coded as ‘1’.
Second, only meaningful variables should be included in the model (Park, 2016). In the present study, both the predictor variables and the control variables are included in the model based on literature. Because of the theoretical underpinning it is assumed only meaningful predictor and control variables are included in the models.
21 Fourth, large sample sizes are typically required for logistic regression (Park, 2016). Leeflang et al. (2015) use between 5 and 10 observations for each variable as a rule of thumb. The dataset that will be used for testing the hypotheses entails 24,985 observations. Thus, the assumption is satisfied.
4.4 Model selection
To determine which models will eventually be used to test the hypotheses, for both dependent variables multiple models were estimated. Ultimately, one model will be selected for each dependent variable. For estimating the models, a stepwise approach was followed, which means the control variables were taken out one by one. In each step the control variable with the highest p-values was taken out. Furthermore, for each dependent variable, a model without the interaction terms was estimated. The variables that are included in each model are shown in Table 6.
TABLE 6 Compared models
Model DV IVs Control variables
1.1 1.2 1.3 1.4 1.5 2.1 2.2 2.3 2.4 2.5 Visit Visit Visit Visit Visit Purchase Purchase Purchase Purchase Purchase
FICs + Age + Interactions FICs + Age + Interactions FICs + Age + Interactions FICs + Age + Interactions FICs + Age
FICs + Comparison + Interaction FICs + Comparison + Interaction FICs + Comparison + Interaction FICs + Comparison + Interaction FICs + Comparison
Gender + Income + Education Gender + Income
Income None
Gender + Income + Education Gender + Income + Education Gender + Income
Income None
Gender + Income + Education
The estimated models are compared by means of four different performance criteria. The models that perform best on these metrics will eventually be used to test the hypotheses. The first metric that has been used is the Akaike Information Criterion (AIC). It is a metric to estimate the relative quality of the models and penalizes models that include variables that do not explain a lot (Leeflang et al., 2015). Therefore, model selection based on the AIC leads to a preference for parsimonious models. When comparing models by means of the AIC, the model with the lowest AIC is preferred (Leeflang et al., 2015).
22 the model predicted the dependent variable was ‘1’, and the other group consisted of observations for which the model predicted the dependent variable was ‘0’. The probability of the occurrence of a ‘0’ is greater than the probability of the occurrence of a ‘1’ since the dependent variables rarely occur in the dataset (see Table 5). That is why a cut-off value of .25 was used for calculating the hit rates, instead of the default cut-off value of .5. Using the value of .25 resulted in a greater number of predicted ‘1s’ and is found to be a more reasonable cut-off value after experimenting with different values. Besides, due to unequal distribution does the overall hit rate not give a complete view on the prediction accuracy. Therefore, the hit rates for only the ‘1s’ were also calculated for all models.
The models were also compared by means of the Pseudo R². The Pseudo R² value indicates how well the model fits the data. The higher the Pseudo R², the better the model fit. In the present study the Nagelkerke R² index was used, which is a corrected version of the Cox and Snell R² index (Walker & Smith, 2016).
Moreover, the Top Decile Lift (TDL) was used to compare the estimated models. The metric represents the ability of the model to identify customers that have a high probability of entering the visit stage or purchase stage (Leeflang et al., 2015). The higher the TDL value, the better the model identifies the purchase journeys that comprise a visit or purchase. In Table 7 the comparison of the models is shown.
TABLE 7
Model fit and performance comparison Model AIC Hit rate
(overall) Hit rate (1s) Pseudo R² TDL 1.1 1.2 1.3 1.4 1.5 2.1 2.2 2.3 2.4 2.5 15325 15340 15344 15405 15318 1763.0 1765.2 1763.4 1815.4 1773.5 90.77% 90.75% 90.74% 90.76% 90.75% 99.35% 99.35% 99.35% 99.36% 99.39% 87.98% 87.62% 87.56% 88.61% 87.26% 15.38% 16.67% 16.67% 18.18% 25.00% 0.119 0.117 0.117 0.111 0.119 0.086 0.077 0.077 0.042 0.079 2.729 2.580 2.503 2.542 2.740 4.416 4.287 4.027 4.611 4.157
23 indicating these models have the best model fit. The TDL is the highest for model 1.5. The model is performing 2.729 times better than a random model would perform. It is performing slightly better than model 1.1, while in model 1.1 five additional parameters were estimated. A visualization of the performance of model 1.5 is shown in Appendix B. Additionally, a likelihood ratio test revealed the model is significantly (p = < .001) better than a null model. Based on the above, model 1.5 will be used to test the hypotheses in section 4.5.
For the purchase stage model 2.1 is considered as the best performing model. Model 2.1 has the best AIC (1763.0) and Pseudo R² (0.086). Hence, compared to the other models, model 2.1 fits the data best. Model 2.4 has a better TDL and slightly better hit rates, but it underperforms on the other two metrics. In Appendix B the performance of model 2.1 is visualized. To test the performance of the model relative to a null model, a likelihood ratio was performed. The test revealed model 2.1 is significantly (p = < .001) better than a null model. Hence, model 2.1 will be used to test the hypotheses related to the purchase stage.
4.5 Testing hypotheses
To interpret the outcomes of the logistic regressions, three means of interpretation were generated: β-coefficients, odd ratios, and marginal effects. The three means all have different ways of interpreting. Firstly, for the β-coefficients, a positive (negative) and significant estimate indicates an increase in the predictor variable leads to an increase (a decrease) in the probability of observing an ‘1’ (a consumer entering the visit or purchase stage). Secondly, odd ratios were generated. Odd ratios are the exponentiated values of the coefficients. A negative β-coefficient will lead to an odd ratio below 1, whereas a positive β-β-coefficient will lead to an odd ratio larger than 1. Lastly, marginal effects were generated. Marginal effects show the effect of each predictor variable when all other variables take their average value. An important note is that only the statistically significant outcomes can be interpreted (i.e. p-values = < .05) 4.5.1 Visit stage
24 not displayed in the table. The full results of the estimation (including the results for Age and the control variables) are shown in Appendix C.
TABLE 8 Excerpt results model 1
Variable β-coefficients Odds ratio Marginal effects Sig. (Intercept) Affiliates Banners Emails Prerolls Retargeting -2.501 0.022 0.025 0.069 0.064 0.484 0.082 1.022 1.025 1.072 1.066 1.623 0.003 0.003 0.008 0.007 0.056 <0.001*** 0.279 0.062 <0.001*** 0.008** <0.001*** Significance codes: ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Based on the output, the following conclusions can be drawn:
- Affiliates (p = .279) and Banners (p = .062) are not statistically significant. Based on the p-values, one must assume the variables do not affect the probability consumers enter the visit stage. Hence, no support has been found for H1A and H1B.
- Emails is highly significant (β = .069; p = < .001). The odds ratio shows that for each e-mail touchpoint, the probability of a consumer entering the visit stage, increases by a factor of 1.072. Furthermore, the marginal effects show each e-mail touchpoint increases the probability of the average consumer entering the stage by 0.8%. Hence, support has been found for H1C. A graph of the curve is shown in Appendix D.
- Prerolls is significant (β = .064; p = .008). An increase in Prerolls of one, increases the probability of consumers entering the visit stage by a factor of 1.066. Furthermore, per pre-roll touchpoint the probability of the average consumer entering the stage increases by 0.7%. Hence, support has been found for H1D. A logistic curve is shown in Appendix D.
25 4.5.2 Purchase stage
The hypotheses concerned with the purchase stage are the direct effects of the firm-initiated touchpoints (H2A, H2B, H2C, H2D, H2E) and the interaction effects of comparison websites and retargeting (H4A). The model comparison revealed the model including the interaction effect is the best performing model. Thus, in this model, both the direct effects as the interaction effect are examined. Table 9 displays the estimation results of the purchase stage model. The full results, including estimates for the control variables, are shown in Appendix E.
TABLE 9
Excerpt estimation results model 2
Variable β-coefficients Odds ratio Marginal effects Sig. (Intercept) Affiliates Banners Emails Prerolls Retargeting Comparison (Retargeting * Comparison) -7.356 0.007 -0.360 0.048 -0.739 0.009 0.004 <-0.001 <0.001 1.007 0.698 1.049 0.477 1.009 1.004 1.000 <0.001 -0.001 <0.001 -0.003 <0.001 <0.001 <-0.001 <0.001*** 0.923 0.554 <0.001*** 0.363 <0.001*** <0.001*** 0.001* Significance codes: ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Based on the output in Table 9 the following conclusions can be drawn:
- Affiliates (p = .923), Banners (p = .698), and Prerolls (p = .363) are not significant. The touchpoints do not affect the probability a consumer enters the purchase stage. Hence, no support was found for H2A, H2B, and H2D.
- Emails is highly significant (β = .048; p = < .001). The odds ratio shows that for each e-mail touchpoint the probability of a consumer entering the purchase stage, increases by 1.049. Furthermore, the marginal effects show positive yet small effects (< 0.1%) on the probability of the average consumer entering the stage. Hence, support has been found for H2C. The logistic curve is shown in Appendix F.
26 output, one can conclude support has been found for H2E. In Appendix F a visualization of the logistic curve is shown.
- The interaction of Comparison and Retargeting is significant (β = < -.001; p = .001). Hence, comparison website touchpoints do affect the effectiveness of retargeting. The output shows comparison website touchpoints weaken the relationship between retargeting touchpoints and purchases, whereas the opposite was hypothesized. However, a plot (see Figure 8 below) of the interaction effect reveals the effect varies for different levels of comparison website touchpoints and retargeting touchpoints. The figure shows that comparison website touchpoints positively affect the effectiveness of retargeting, until 625 retargeting touchpoints are reached. Thus, exposing consumers to more than 625 retargeting touchpoints is more effective for consumers with no comparison website touchpoints in their journey. This indicates that advertising via retargeting should be limited up to 625 times for consumers with comparison website touchpoints in their journey. Based on the above, one can conclude partial support has been found for H4A. Comparison website touchpoints can increase retargeting effectiveness, yet, it only increases the effectiveness for journeys with less than 625 retargeting touchpoints.
FIGURE 8
The interaction effect of comparison websites on retargeting effectiveness
27 4.6 Re-estimation
To explore the hypotheses that are not supported, the models were re-estimated with some adjustments. The adjustments for both stages are discussed below.
4.6.1 Visit stage
In the model concerning the visit stage, Affiliates and Banners turned out to be insignificant. The interaction terms were not estimated since the model without interaction terms was considered as the better performing model.
For the first adjustment, all firm-initiated were transformed from count variables into binary variables. Hence, the variables could only take two values and now indicated whether the firm-initiated occurred at least once in the purchase journey or not at all. In Table 10 the results of the re-estimation are shown. The full results are shown in Appendix G.
TABLE 10
Excerpt results re-estimation with binary variables Variable β-coefficients Odds ratio Marginal
effects Sig. (Intercept) Affiliates1 Banners1 Emails1 Prerolls1 Retargeting1 -2.647 0.584 0.537 1.451 0.005 4.179 0.071 1.794 1.710 4.267 1.005 65.310 0.062 0.056 0.210 <0.001 0.774 <0.001*** <0.001*** 0.003** <0.001*** 0.975 <0.001*** Significance codes: ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Looking at the re-estimation results in Table 10, the following conclusions can be drawn: - Affiliates, Banners, Emails, and Retargeting turned out to have significant p-values and
positive estimates. Hence, compared to not using the advertising forms, does using the advertising forms positively affect the probability a consumer enters the visit stage. The results of this model supported the previously not-supported hypotheses: H1A and H1B. While the count variables for Affiliates and Banners were not significant, the binary variables do show significant p-values. Thus, the variables positively affect visit probabilities, but do not lead to a constant change in the response variable. Hence why the GLM was not able to fit significant logistic curves of the relationships.
28 a significant effect. Besides, including the interaction effects, took away the significance of the main effects, which might be due to the correlation between the main variables and their interaction variables. Afterward, the model was re-estimated, but now the binary advertising form variables were included. The results are shown in Appendix I. Neither in this model significant interaction terms were found.
For the third adjustment, a new variable was created to distinguish four different age groups. The variable Age was replaced with the new variable Agegroup. The replacement was done first for the model containing the count variables. An excerpt of the results is shown in Table 11. The full output of the model is given in Appendix J.
TABLE 11
Excerpt results re-estimation with age groups Variable β-coefficients Odds ratio Marginal
effects Sig. (Affiliates*Agegroup2) (Affiliates*Agegroup4) (Affiliates*Agegroup4) (Banners*Agegroup2) (Banners*Agegroup3) (Banners*Agegroup4) (Emails*Agegroup2) (Emails*Agegroup3) (Emails*Agegroup4) (Prerolls*Agegroup2) (Prerolls*Agegroup3) (Prerolls*Agegroup4) (Retargeting*Agegroup2) (Retargeting*Agegroup3) (Retargeting*Agegroup4) -0.009 0.038 -0.010 0.140 0.134 0.154 -0.355 -0.348 -0.331 0.068 0.033 0.010 0.222 0.148 0.108 0.991 1.038 0.990 1.151 1.144 1.167 0.701 0.706 0.718 1.070 1.033 1.010 1.249 1.159 1.114 -0.001 0.004 -0.001 0.016 0.016 0.018 -0.042 -0.041 -0.039 0.008 0.004 0.001 0.026 0.017 0.013 0.979 0.909 0.976 0.468 0.489 0.426 0.174 0.175 0.201 0.685 0.843 0.951 0.035* 0.137 0.346 Significance codes: ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05
Based on the estimation results the following can be concluded:
29 - None of the other interaction terms have significant p-values, thus, for these variables, age does not play a role in retargeting effectiveness. Additionally, when changing the reference level still no significant interaction terms were observed. Hence, no support has been found for H3A, H3B, H3C, and H3D.
For the last adjustment, the binary variables for the firm-initiated touchpoints and the variable Agegroup were used. The results of the estimation are given in Appendix K. For this model no significant interaction terms were found.
4.6.2 Purchase stage
In the second model Affiliates, Banners, and Prerolls turned out to be insignificant. Emails and Retargeting, and the interaction of Comparison did show significance.
The model was re-estimated using the binary variables instead of the count variables for the firm-initiated touchpoints. The full results of the estimation are shown in Appendix L. Emails and Retargeting were significant again, while Affiliates, Banners, and Prerolls remained insignificant. Affiliates, Banners, and Prerolls not showing any effect on the purchase stage, could be due to the rare occurrence of purchases (154) in the dataset.
4.7 Support of hypotheses
30 TABLE 12
Overview hypotheses and conclusions
Hypotheses Supported?
H1A: Affiliate marketing positively influences the probability a consumer enters the visit stage. H1B: Banners positively influence the probability a consumer enters the visit stage.
H1C: E-mail marketing positively influences the probability a consumer enters the visit stage. H1D: Pre-rolls positively influence the probability a consumer enters the visit stage.
H1E: Retargeting positively influences the probability a consumer enters the visit stage.
Yes Yes Yes Yes Yes
H2A: Affiliate marketing positively influences the probability a consumer enters the purchase stage. H2B: Banners positively influence the probability a consumer enters the purchase stage.
H2C: E-mail marketing positively influences the probability a consumer enters the purchase stage. H2D: Pre-rolls positively influence the probability a consumer enters the purchase stage.
H2E: Retargeting positively influences the probability a consumer enters the purchase stage.
No No Yes No Yes
H3A: Age moderates the effect affiliate marketing has on the visit stage. H3B: Age moderates the effect banners have on the visit stage.
H3C: Age moderates the effect e-mail marketing has on the visit stage. H3D: Age moderates the effect pre-rolls have on the visit stage. H3E: Age moderates the effect retargeting has on the visit stage.
No No No No Partially
H4A: Comparison websites positively moderate the effect retargeting has on the probability a consumer enters the purchase stage.
31
5. DISCUSSION
In section 5.1 the results presented in the previous section will be discussed and compared to the literature mentioned in the introduction. Subsequently, in section 5.2, the theoretical and practical implications of the results will be presented. Lastly, in section 5.3, limitations and suggestions for future research will be discussed.
5.1 Summary and discussion of results
The present study aimed to provide answers to the research question, consisting of three components. The first component was concerned with the effect of affiliate marketing, banners, e-mail marketing, pre-rolls, and retargeting on the visit stage and the purchase stage. The second component aimed at the moderating effect of age on the effectiveness of the five advertising forms, while the third component focused on the moderating role of comparison websites for retargeting effectiveness. The research question that resulted in 16 different hypotheses, will be discussed below.
The effect of different advertising forms on the visit stage and purchase stage (H1 and H2) Affiliate marketing did positively affect the probability consumers enter the visit stage; however, it does not affect the probability consumers enter the purchase stage. In previous studies, affiliate marketing is considered as one of the most important methods to acquire customers, increase sales, and drive traffic towards the website (Libai et al., 2003; Gregori et al., 2014; Fox & Wareham, 2012). The findings in the present study confirm affiliate marketing generates more visitors, yet, it does not increase the number of visitors with purchase intentions. A closer look to the literature reveals the type of affiliate the advertiser cooperates with, might play a role. According to Edelman and Brandi (2015), advertisers do often have difficulties selecting appropriate affiliates. The affiliates typically have no well-known brand names or established reputations. Thus, there might not be a good fit between the advertiser and affiliate, resulting in users immediately bouncing from the advertisers’ website.
32 found an increase in purchase probabilities when using banner advertisements. A possible explanation for not finding a significant effect on purchases in the present study might be provided by Muñoz-Leiva et al. (2019). According to the authors, consumers do usually not pay attention to banner advertisements. They state banners could even lead to negative attitudes towards firms (Muñoz-Leiva et al., 2019). A striking result from the present study is that banner advertising did significantly affect visit probabilities, whereas it did not affect purchase probabilities. According to Tolomei, Lalmas, Farahat, and Haines (2019), users sometimes accidentally click on an advertisement, visit the advertisers’ website, and immediately bounce back without spending time on the landing page. In this way, consumers without any purchase intentions will be driven to the website (Tolomei et al., 2019). This might explain the positive significance of visit probabilities, and not of purchase probabilities.
33 Campbell et al. (2017), since, they state pre-rolls lead to consumer annoyance and irritation. Consumers are forced to watch an advertisement, while they intend to watch other content (Campbell et al., 2017). Besides, as with the findings of banner advertising, positive effects on visit probabilities were found. Accidental ad clicks might play a role in pre-roll advertising as well.
It is found retargeting has a positive effect on both the visit stage and the purchase stage. These findings are in line with the findings of De Haan et al. (2016) and Lambrecht and Tucker (2013). De Haan et al. (2016) found a positive effect of retargeting on website traffic, while Lambrecht and Tucker (2013) mentioned the advertising form has a positive impact on firm outcomes. Additionally, in the present study is found retargeting has among the five advertising forms, the strongest effect on visit probabilities. This might relate to by what is emphasized by De Haan et al. (2016); retargeting focuses on consumers that have already visited the website. It uses the initial intention of the consumer to visit the website and evokes the consumer to re-visit the website. The consumers that are interested in the products, and were exposed to retargeting advertisements, already know the way to visit the website (Blake, Nosko & Tadelis, 2015). Thus, the fact that consumers exposed to retargeting advertisement have already visited the website in the past, might explain the great impact of retargeting on visit probabilities.
The moderating role of age for the effectiveness of firm-initiated touchpoints (H3)
34 The moderating role of comparison websites for retargeting effectiveness (H4)
In the present study, comparison website touchpoints are found to have a positive effect on retargeting effectiveness. This corresponds with the study of Lambrecht and Tucker (2013) about the moderating role of review websites, in which is argued dynamic retargeting becomes more effective for consumers that visited review websites. They mentioned that consumers who visit review websites have more developed product preferences, which is according to Laffey and Gandy (2009) also the case for comparison websites. However, in the present study it is also found the effect of comparison website touchpoints becomes negative after the threshold of 625 retargeting touchpoints is reached. It seems that the use of a lot of retargeting for one consumer is less effective for evoking hesitant consumers (or consumers that compare alternatives on comparison websites) to visit the website. Thus, the moderating effect of comparison websites depends on the number times the comparison touchpoints and retargeting touchpoints occur. Lambrecht and Tucker (2013) only considered whether review website touchpoints occurred, and not whether the effect changes for different numbers of touchpoints. Besides, an important note is that Lambrecht and Tucker (2013) made a distinction between dynamic and generic ads. They found a positive influence of review touchpoints on the effectiveness of dynamic ads, but no effect on generic ad effectiveness. In the present study, no distinction was made between different types of retargeting advertisements. Likewise, the effect of comparison website touchpoints might also differ for various types of retargeting ads (e.g. dynamic versus generic).
5.2 Implications
35 research concerning banner and pre-roll advertising, did this study consider the effects on multiple stages of the purchase funnel. Using existing literature, it sought for possible explanations for the opposing effects on the purchase funnel stages. This contributed to the discussion about the true effects of banner and pre-roll advertising.
Furthermore, the present study investigated the effect of age on advertising effectiveness. This extended the research of De Haan et al. (2016) since they raised the call for distinguishing advertising form effectiveness for different types of customers. The findings of the present study pointed out that age might play a role for different types of advertisement content rather than for different advertising forms, which can be considered in future research.
Moreover, this research elaborated on the study of Lambrecht and Tucker (2013) about the effectiveness of retargeting. It investigated the role of comparison websites in retargeting effectiveness, while Lambrecht and Tucker (2013) did only study the effect of review websites. It revealed the effect of comparison website touchpoints not only depends on whether comparison website touchpoints occur in a journey, but also on the number of times they occur. Lambrecht and Tucker (2013) did not consider this when investigating the effect of review websites on retargeting effectiveness.
Not only does this study contribute to academic thinking, it also has some practical relevance. According to Lemon and Verhoef (2016), it is critical for firms to gain understanding of the customer experience, as it could lead to higher conversion rates and improved customer loyalty. Marketing managers may benefit from the findings of this paper, since it provides insights in the effectiveness of different firm-initiated touchpoints. While all investigated advertising forms increase the probability of consumers visiting the website of the advertising firm, only e-mail marketing and retargeting increase the probability of consumer entering the purchase stage. Managers may take this into account when implementing marketing efforts. Moreover, this study reveals the extent to which the touchpoints increase the probabilities, which can come in handy when allocating budgets to different advertising forms.
Furthermore, the results show firms could better use retargeting for people between 30 and 49 years old than for people younger than 30 years old. This finding might also be useful for firms to tailor their marketing efforts more effectively and target customer segments.
36 for consumers with at least one comparison website touchpoint. Hence, marketing managers should limit the extent to which they use retargeting for customers who have comparison website touchpoints in their journey.
5.3 Limitations and future research
This study is not without limitations. Yet, the limitations in this study may serve as opportunities for future research. First, the dataset that was used for this study entailed, after data cleaning, only 154 focal purchases. This might have led to the insignificant parameters for the hypotheses related to the purchase stage. Thus, future research concerning advertising effectiveness could use datasets comprising more purchases.
Second, the data that were used for this research were collected in the Netherlands, which makes the generalizability of the results questionable. According to Chau, Cole, Massey, Montoya-Weiss, and O’Keefe (2002), cultural differences do affect the online behavior consumers exhibit. Moreover, the data were collected specifically in the travel sector. The results might thus not be generalizable to other industries. For future research it might be interesting to investigate to what extent advertising effectiveness differs across various industries and across various countries.
Third, for the model concerning the purchase stage, there was not accounted for the sequence in which comparison website touchpoints and retargeting touchpoints occurred in a purchase journey. Therefore, comparison website touchpoints that have occurred after retargeting touchpoints, might have been used to assess the effect on retargeting effectiveness.
Fourth, this study investigated the impact of age for advertising effectiveness, however, it did not account for the content of the advertisements. Researchers willing to investigate the influence of age on visit or purchase probabilities should make a distinction between the content of the advertisements (e.g. rational versus emotional).
Fifth, in this study the number of focal visits might be affected by clicks happened by mistake. Therefore, future researchers might want to account for the effect of accidental clicks. Tolomei et al. (2019) detected accidental clicks by tracking the time a user spends on a landing page (so-called dwell-time). In the future, researchers might consider taking this into account when investigating the effect of advertising forms on visit probabilities.
38
6. REFERENCES
Allison, P. D. (2012). Logistic regression using SAS: Theory and application. SAS Institute. Alreck, P., & Settle, R. B. (2002). Gender effects on Internet, catalogue and store shopping. Journal of Database Marketing & Customer Strategy Management, 9(2), 150-162. Anderl, E., Schumann, J. H., & Kunz, W. (2016). Helping firms reduce complexity in multichannel online data: A new taxonomy-based approach for customer journeys. Journal of Retailing, 92(2), 185-203.
Ansu-Mensah, P., Asuamah, S. Y., & Amankwah, J. (2013). Consumers’ attitude towards advertisement elements: A survey of marketing students in Sunyani Polytechnic, Ghana, West Africa. International Journal of innovative research in management, 2(4), 13-24.
Baxendale, S., Macdonald, E. K., & Wilson, H. N. (2015). The impact of different touchpoints on brand consideration. Journal of Retailing, 91(2), 235-253.
Becker, T. E. (2005). Potential problems in the statistical control of variables in organizational research: A qualitative analysis with recommendations. Organizational Research Methods, 8(3), 274-289.
Bewick, V., Cheek, L., & Ball, J. (2005). Statistics review 14: Logistic regression. Critical care, 9(1), 112.
Blake, T., Nosko, C., & Tadelis, S. (2015). Consumer heterogeneity and paid search effectiveness: A large‐scale field experiment. Econometrica, 83(1), 155-174.
Bowman, D., & Narayandas, D. (2001). Managing customer-initiated contacts with manufacturers: The impact on share of category requirements and word-of-mouth behavior. Journal of marketing Research, 38(3), 281-297.
Bughin, J. (2015). Brand success in an era of Digital Darwinism. Retrieved March 1, 2020 from https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/brand-success-in-an-era-of-digital-darwinism
