Guiding consumers towards a purchase
Investigating the effect of timing of firm-initiated touchpoints within
the customer journey
by
Alex Buis
Guiding consumers towards a purchase
Investigating the effect of timing of firm-initiated touchpoints within
the customer journey
By:Alex Buis
University of Groningen Faculty of Economics and Business
Master thesis MSC Marketing Intelligence & Management (Final version)
Completion date: January 12, 2020
Vierde Drift Noorderhaven 20 9712 AH Groningen
(06)55468080 a.j.buis@student.rug.nl Student number: s3495256
Management summary
Over the past few years, the interest in customer experience has increased in both the practical and theoretical field. Where customer experience is conceptualized as a customer journey across multiple touchpoints with a company. Several studies have focused and contributed in gaining a deeper understanding of the customer journey. However, no study yet has examined the influence of marketing instruments and their timing within a customer journey. The purpose of this study is to examine the role of timing of marketing instruments and their role on conversion probability within a customer journey. In other words; how can a firm guide consumers towards a purchase. Therefore the main research question is:
“What is the influence of the timing of firm-initiated touch-points on the conversion probability within a customer journey?”
In order to investigate the research question a dataset of a Dutch travel agency was made available by GFK. The dataset contains path data, in which each touchpoint is registered within a customer journey. Between the registered touchpoints a distinction is made between touchpoints which are initiated by the consumer or which are initiated by the focal firm. Where touchpoints which are initiated by the focal firm, represents their marketing mix. The marketing mix (firm-initiated touchpoints) of the focal firm consists out of: affiliate
marketing, email marketing, banner advertising, pre-roll advertisements and retargeting. The customer journeys within the dataset are about booking a trip faraway towards a sunny destination.
The term customer journey is not always used in previous literature. Path to purchase and purchase funnel are both also commonly used. Despite the different terminology, they all describe the process in which a consumer goes through certain stages prior to making a purchase. Within previous literature, the stages within a customer journey also differ. But most seem to be based on the behavioral sequence of: need recognition, information search, evaluation, purchase and post-purchase. Within this study, the customer journey consist out of three subsequent stages: the consideration stage, the visit stage and the purchase stage.
In order to analyze the research question and to test the hypotheses, three different methods are used. Although the customer journey exist out of three stages, only the
consideration stage and the visit stage are analyzed. Since the purchase stage solely consists out making the purchase itself and completing the transaction. The three methods used are; a logistic regression, a multinomial logistic regression and a Markov model. Via a logistic regression the consideration stage was first analyzed in isolation to determine the effect of firm-initiated touchpoints during that stage. Second, via the multinomial logistic regression the visit stage was analyzed in isolation to determine the effect of firm-initiated touchpoints on conversion during that stage. Lastly a Markov model was used in order to analyze the entire journey to determine the overall effectiveness of firm-initiated touchpoints.
Overall the results show that the consumer-initiated touchpoints have a greater effect on conversion than the firm-initiated touchpoints. The Markov model only attributes 18% of the conversion towards the firm-initiated touchpoints. Which is similar as the results of analyzing the consideration stage and visit stage in isolation. Within the consideration stage only a few significant different and positive effects were found on conversion compared to a generic search for the firm-initiated touchpoints. Again in the visit stage there were not many significantly different positive effect for the firm-initiated touchpoints on conversion, this time compared to a visit to a competitors website and towards moving back to the
consideration stage. While most of the firm-initiated touchpoints seem to increase the odds for ending the journey without a purchase.
In conclusion consumer-initiated touchpoints play a big role in conversion. Within this study a visit to the competitors website seems to be the most important touchpoint in relation to conversion. A possible explanation is that consumers want to validate their choice (e.g. see whether they get a good deal). For the firm-initiated touchpoints retargeting and email
Preface
This master’s thesis is the final project of the master Marketing Intelligence and Management at the University of Groningen for me. Writing this master’s thesis allowed me, to put the skills I have developed during the master to the test as well as it allowed me to learn and explore new things. Studying the customer journey fitted perfectly with my interests. Since, my previous education of commercial economics at Windesheim in Zwolle, I have always been interested with the phenomenon customer journey. Especially mapping the customer journey, analyzing the points of pain/pleasure and coming up with ideas for improvement fascinates me. Via studying a phenomena that has interested me for a long time seems a fitting end for my educational period.
I am very thankful to my supervisor dr. Peter van Eck for guiding me throughout the process of writing my master thesis. The feedback and sessions were very helpful in
overcoming obstacles along the road. Also I am thankful for the opportunity to work with a real life dataset. I am also grateful towards my fellow group members, for providing me with feedback during the group meetings and the pleasant atmosphere in general during these meetings.
Writing this thesis and studying the influence of firm-initiated touchpoints within a customer journey was something I have really enjoyed. Therefore, I hope that you will enjoy reading this thesis as well.
Alex Buis, Groningen, January 2020
Table of contents
1. Introduction ... 1 2. Theoretical background ... 2 2.1 Customer journey ... 2 2.2 Firm-initiated touchpoints ... 4 2.2.1 Affiliate marketing ... 5 2.2.2 Banner advertising ... 5 2.2.3 Retargeting ... 6 2.2.4 Pre-roll advertising ... 7 2.2.5 Email advertising ... 8 2.3 Consumer-initiated touchpoints ... 8 2.4 Conceptual model ... 9 3. Methodology ...10 3.1 Data ...10 3.2 Plan of analysis ...123.2.1 Logit models and the multinomial model...12
3.2.2 Higher-order Markov Chain ...14
4. Data preparation and cleaning ...16
5. Results ...18
5.1 Results for the consideration stage ...18
5.1.1 Validating the conversion probability model ...18
5.1.2 Face validity of the logistic regression ...19
5.1.3 Hypotheses testing via the logistic regression ...20
5.1.4 Comparing the results with the transition probability model ...22
5.2 Results for the Visit stage ...22
5.2.1 Validating the multinomial logistic regression ...22
5.2.2 Face validity of the multinomial regression ...24
5.3 Results of analyzing the entire journey ...27
5.3.1 Validating the model ...27
5.3.2 Results of the Markov Chain ...28
5.4 Meaningful results for the firm-initiated touchpoints ...30
5.4.1 Affiliate marketing ...30
5.4.2 Banner advertisements ...31
5.4.3 Email ...32
5.4.4 Pre-roll ...32
5.4.5 Retargeting ...33
6. Conclusion & managerial implications ...33
1
1. Introduction
When nearly 400 decision makers were asked about their top priorities for the upcoming year, improving the customer experience received the most number one rankings (21%), followed by growth in revenues (17%) and improving the differentiation (16%) (Accenture, 2015). A related and important aspect is the understanding of the customer journey, since customer experience is conceptualized as a customer’s journey with a company during a purchase cycle across multiple touchpoints (Lemon & Verhoef, 2016).
The influence of the use and migration between touchpoints in customer journeys has mainly been studied in multichannel, online and service marketing literature (Lemon & Verhoef, 2016). For example, Li and Kannan (2014) developed a measurement model to analyze consumers consideration of online channels and subsequent visits through these channels. They were the first to examine these effects in an online environment while considering different stages of a consumers’ path to purchase. Another study examined the effectiveness of various online forms of advertising throughout the purchase funnel. In which the authors try to determine; which form of advertising is the most effective, when the effects of advertising occur and where in the purchase/website funnel these effects occur (de Haan, Wiesel, & Pauwels, 2016). Other research has shown that the most effective type of
advertisement on visiting and conversion behavior depends on the individual’s advertisement impression history. In which the authors also account for a wear-out effect of each type of advertisement (Braun & Moe, 2013).
Although there are already several studies that focus on gaining a deeper
2 The investigation of the research question will contribute to the theoretical field by addressing a relevant research gap; the role of timing of firm-initiated touchpoints within a customer journey (Lemon & Verhoef, 2016; Marketing Science Institute, 2018). The results of this study will contribute to the practical field by gaining insights in how touchpoints (both firm-initiated and customer-initiated) and their timing within a customer journey influence the conversion probability. A deeper understanding, will enhance the ability of managers to use marketing instruments (firm-initiated touchpoints) to guide their customers towards a purchase. Although the attribution of touchpoints (or channels) is not something new, the influence of their timing has not been examined. Since most channels or touch-points are paid for either in time or money (or both), it is vital to have a good understanding of the value of each touch-point.
In the next section, the theoretical framework will be discussed with all the relevant concepts that relate to the research question and will be summarized in the conceptual model. Then, in the subsequent section, the research design will be discussed. Containing the plan of analyses; which includes a description of the methods and the data that will be used.
Subsequently, the results will be discussed in terms of findings and in terms of reliability and validity. Lastly; the conclusions, recommendations and limitations will be discussed.
2. Theoretical background
2.1 Customer journey
Within the literature there are several terms that relate to the same phenomena; the path to purchase, customer journey and purchase funnel. All similar in the fact that they are used to describe the process where a consumer goes through certain stages prior to making a
purchase. Throughout this paper, the term customer journey will be used. Which and how many stages there are however differ, table 1 gives an overview of the different stages that are used within a customer journey.
3 customer journeys (regardless of the stages used) seems to be based on the behavioral
sequence of: need recognition, information search, evaluation, purchase and post-purchase.
Table 1:
Overview of definitions of customer journeys
Only the customer journey of Hoban & Bucklin (2015), out of table 1, does not really
resemble that sequence. Their focus is more on a consumer’s interaction with the focal firm’s website and the extent of information availability on a consumer. Since the focus of this study Authors Subsequent stages used within the customer journey
Neslin, Grewal, Leghorn, Shankar, Teerling, Thomas & Verhoef (2006)
1. Need recognition: Customer recognizes his/her needs. 2. Information search: Via various channels.
3. Purchase: Customer decides on which channel to use to make the purchase. 4. After Sales: Customer receives sales support (e.g. for insurance: advice on
increased coverage). Lemon & Verhoef
(2016)
1. Pre-purchase: Encompasses all customer’s interaction with a company/brand before the transaction. Like: need recognition, search and consideration. 2. Purchase: Encompasses all customer’s interaction during the transaction
itself. Like: choice for product/service, ordering and payment.
3. Post-purchase: Encompasses all customer’s interactions after the purchase. Like: consumption, engagement and service requests.
Hoban & Bucklin (2015)
1. Non-visitor: Consumers who have never interacted with the firm’s website. 2. Visitor: Consumers who have visited the website but have not provided
personal information necessary to sign up for an account.
3. Authenticated user: Consumers who have created an account, but no transaction has taken place.
4. Converted customer: Consumers who have completed a transaction. Li & Kannan (2014) 1. Consideration stage: Consumers recognize their needs and considers
different channels for information search.
2. Visit stage: Consumers visit the website via specific channels to search information and evaluate alternatives.
4 goes beyond interactions with only the firm’s website, the approach of Hoban & Bucklin (2015) is less suitable. Moreover, stages after a purchase are beyond the scope of this research since the focus is on whether the purchase takes place or not. To determine when the firm-initiated touchpoints have the most effect within a customer journey on the purchase probability, more and various stages prior to a purchase could provide more detailed information. Neslin et al. (2006) and Li and Kannan (2014) use both two stages prior to the purchase. Both use need recognition and information search, however Li and Kannan (2014) also include the evaluation of alternatives. So in some extent they use a more detailed approach in the stages prior to a purchase. Therefore the approach of Li and Kannan (2014) will be used in this study.
Each of the stages in the customer journey consists of touchpoints which are either consumer-initiated or firm-initiated. To define them more clearly; customer-initiated
touchpoints are touchpoints where the customers seeks out information on their own. While with firm-initiated touchpoints, firm’s initiate marketing communications (Li & Kannan, 2014). It is important to make a distinction between the two since the objective of this study is to examine how firms can guide a consumer towards a purchase.
2.2 Firm-initiated touchpoints
For analyzing the influence of firm-initiated touchpoints a dataset is obtained from a Dutch Travel agency, containing online path data. The dataset will be discussed more in chapter 3; methodology. The firm-initiated touchpoints that will be examined in this study are; affiliate marketing, banner advertisements, retargeting, pre-roll advertisements and email marketing. Within the Netherlands the amount of money spend on digital advertising has increased by 9% in 2017, which was mainly driven by search, mobile, social and online video advertising. For example; revenue from display ads on mobile devices increased by 19%, while
5
2.2.1 Affiliate marketing
Affiliate marketing is a commission-based form in which an affiliate partner (e.g. a review site) is rewarded for referring a user towards the website of a firm (Anderl, Becker, von Wangenheim, & Schumann, 2016). Based on the definition affiliate marketing can only occur close or during the visit stage of the journey. In their study across 4 different datasets (Anderl et al., 2016) found that the results for the effect of affiliate marketing were inconclusive. For two of the three datasets, which contained affiliate marketing as a channel, they found a positive influence on conversion probability. In fact, for those datasets affiliate marketing was the strongest predictor for conversion. However, for the other dataset, they found a significant negative effect on the conversion probability. Other studies focused on how affiliate
marketing interacts with search-engine advertising (Olbrich, Bormann, & Hundt, 2019) or how firms can manage their affiliate programs in terms of risk, incentives, and information (Edelman & Brandi, 2015). To the best of my knowledge, the timing of affiliate marketing within a customer journey has not been examined. Therefore, it is difficult to derive a
hypothesis about where in the customer journey affiliate marketing is expected to be effective in relation to the conversion probability. However, since (Anderl et al., 2016) found affiliate marketing as the strongest predictor for two of their used datasets and that, based on the definition of affiliate marketing, consumers have to visit a partner website (e.g. a review site) first. The expectation is that affiliate marketing will have a significant positive influence on conversion probability in the visit stage.
H1: Affiliate marketing will have a significant positive effect on the conversion probability in the visit stage.
2.2.2 Banner advertising
Banner advertising or display adverting is the graphical object with the adverting message into a website and the effect of generic banners are stronger when a consumers has no narrowly construed preferences (Anderl et al., 2016). The results of Li and Kannan (2014) indicate that display ads are effective in the short run. While (Hoban & Bucklin, 2015) found that online display advertising has a small and positive effect in three of the four purchase stages (non-visitor, visitor, authenticated visitor and converted visitor) they examined. Moreover, they found that in earlier stages the marginal returns to more impressions
6 significant effect of display ads on their behavior (Hoban & Bucklin, 2015). Another study found that display advertisements have a low direct effect on purchase conversions, however, they are likely to stimulate subsequent visits through other online formats (Xu, Duan, & Whinston, 2014). Since Li and Kannan (2014) and Hoban and Bucklin (2015) found a significant and positive effect of display advertising the expectation is that within this study display advertising will have a significant and positive influence on the conversion
probability in the consideration and visit stages. However, since generic advertisements are more effective when consumers do not have clear preferences the expectation is that banners will be more effective in the consideration stage then in the visit stage.
H2a: Banner advertisements will have a significant positive effect in the consideration stage on the conversion probability.
H2b: Banner advertisements will have a significant positive effect in the visit stage on the conversion probability.
H2c: Banner advertisements will be more effective in the consideration stage than in the visit stage.
2.2.3 Retargeting
Retargeting focuses, by definition, on customers who have previously visited the advertiser’s website and require therefore anteceding channels (Anderl et al., 2016). A study examined the effect of two types of retargeting advertisements; a generic ad (a brand ad) and a dynamic ad (product someone previously saw online). They found that a generic advertisement works better when a consumer has no narrowly construed preferences and that a dynamic ad works better when the preferences of a consumer are more narrow and clear. The authors proposed that in an online environment a visit to a site that provides detailed information about specific products (e.g. a product review site) signals that a customer is thinking about specific product attributes and is, therefore, more likely to be developing (or already have developed) narrowly construed preferences (Lambrecht & Tucker, 2013). Another study found somewhat
7 a consumer moves backwards within their customer journey or the consumer has had a
previous journey. Moreover, on one hand the expectation is that retargeting will have a positive effect when a consumer thinks about more specific product attributes. However that effect will decrease when the consumer the consumer moves closer towards a purchase decision.
H3a: Retargeting will have a significant and positive influence on the conversion probability in the visit stage.
H3b: The effect of retargeting on conversion probability will diminish when the consumer moves closer towards a purchase decision.
2.2.4 Pre-roll advertising
Pre-roll advertising is a relatively new form of advertising. Pre-roll advertisements are video’s that are played in exactly the place where a consumer is expecting to view their intended content. Pre-roll advertisements enable consumers with an option to skip after viewing a brief forced segment (Campbell, Mattison Thompson, Grimm, & Robson, 2017). Consumers are in a heightened state of attention because there are anticipating to fulfill their search goal. A pre-roll ad intervenes that process and forces consumers to watch at least a few seconds
(Campbell et al., 2017). No study yet has examined the effect of pre-roll advertising on conversion probability, other studies have focused on the underlying mechanisms of skipping the advertisement (Campbell et al., 2017) or on how to reduce the perceived intrusiveness (Goodrich, Schiller, & Galletta, 2015). However, Goodrich et al. (2015) found that a high level of informativeness reduces the perceived intrusiveness and Campbell et al. (2017) found that a lower perceived intrusiveness leads to a decrease in the likelihood to skip the
advertisement. Since a consumer is searching for information in the consideration stage and in the beginning of the visit stage. The expectation is that then a pre-roll add will have a positive influence on the conversion probability.
H4a: Pre-roll advertising will have a significant and positive effect on conversion probability during the consideration stage.
H4b: Pre-roll advertising will have a significant and positive effect on conversion probability in the visit stage.
8
2.2.5 Email advertising
For consumers who already have had an email intervention within their path to purchase, another email intervention can lower the conversion probability (Li & Kannan, 2014). Although, the results of that study showed overall that email marketing is effective in the short-run. An important lesson is that sending multiple emails is not effective for all visitors, even within the same segment there can occur differences in effectiveness(Li & Kannan, 2014). These findings are somewhat similar to other studies. Others found that an increase in the number of emails, to obtain more clicks, did not lead to a substantial change in website visits or revenues (de Haan et al., 2016). This is in line with their overall conclusion that firm-initiated touchpoints are less effective then consumer-firm-initiated touchpoints at each stage of the used website funnel (de Haan et al., 2016). As a possible explanation for why email
marketing might be less effective, the authors reason that most costumers only respond to an email when they have clear preferences of what and where to buy, based on other sources and are reminded about their purchase intention via email. Another study also found that the effect of an email was stronger if it was preceded by other consumer-initiated touchpoints (Anderl et al., 2016). The authors also found that emails have a strong long-term effect but a weak short-term effect. Moreover, the authors state that emails may serve more as a facilitator for
subsequent visits through other channels rather than lead directly to a conversion (Anderl et al., 2016). To relate the above back to where email is expected to influence the conversion probability within a customer journey. The expectation is that email will have a positive effect on the conversion probability close to when the consumer makes their purchase decision. As a facilitating role to remind the consumer of their purchase intention. Moreover, when the frequency of emails increases (for a consumer who already had at least one email intervention) the conversion probability will decrease.
H5a: Email will have a positive effect on conversion probability when a consumer is in the visit stage.
H5b: Email will have an increasing effect on conversion probability when a consumer is in the visit stage.
2.3 Consumer-initiated touchpoints
9 example previous display impressions have significant spill-over effects towards consumer-initiated touchpoints like direct type in and online search behavior (Anderl et al., 2016). Results of a study, across four different companies (Anderl et al., 2016), showed that all customer-initiated touchpoints account for the majority of the conversion contribution for each company. The attribution varies between 68% and 92%. However, the contribution of each specific consumer-initiated touchpoints varied per company, for one company search engine optimization was far more effective than for the other. The finding that consumer-initiated touchpoints are contributing more to conversions than firm-consumer-initiated touchpoints is underlined in other studies (de Haan et al., 2016). In other words, although firm-initiated touchpoints are the main focus of this study, consumer-initiated touchpoints are not something to ignore. Since previous research, that examined the influence of consumer-initiated touchpoints within a customer journey, showed that the influence of each specific consumer-initiated touchpoint varies across companies. And because the main focus is on the exact influence of firm-initiated touchpoints within this study. The hypothesis for consumer-initiated touchpoints will be more generic. Based on the literature, the expectation is that consumer-initiated touchpoints will have a greater influence on the conversion probability for each stage of the customer journey.
H6a: Consumer-initiated touchpoints will have a greater significant influence on the conversion probability than firm-initiated touchpoints in the consideration stage. H6b: Consumer-initiated touchpoints will have a greater significant influence on the conversion probability than firm-initiated touchpoints in the visit stage.
2.4 Conceptual model
The hypothesis derived from the reviewed literature are usually clearly represented in a simple manner in the conceptual model. However, due to the multiple firm-initiated
10 Figure 1:
The conceptual model
3. Methodology
3.1 Data
To determine when a firm should intervene with marketing interventions within a customer journey. A dataset containing path data will be used provided by GFK, the focal company within the dataset is a Dutch travel agency. The paths in the dataset concern Dutch individuals which customer journeys are targeted on purchasing a trip faraway towards a sunny
destination. The data is collected in a panel form meaning that over a time period individuals have been monitored, with their consent, which resulted in individual observations. The observation period for this data set is from 01-06-2015 until 31-09-2016. Each observations is registered as a touchpoint, an overview of all the touchpoints in the dataset are displayed in table 2. Consider the following example; a consumer searches via google for flight tickets and subsequently clicks on a flight tickets website and sees flight tickets in which they are
interested. Then that consumer visits a comparison site to search for a suitable
accommodation. Unfortunately (s)he couldn’t find anything (s)he likes and ends the journey without a purchase. This example of a customer journey exist out of three touchpoints
11 Table 2:
Overview of the touchpoints within the dataset
Type of touchpoint Description Initiated by Stage
Affiliates Consumer is referred to the focal firm via a partner Firm Consideration/visit
Banner Consumer sees a banner advertisement of the focal firm Firm Consideration/visit
Email Consumer receives an email from the focal firm Firm Consideration/visit
Pre-rolls Consumer watches a pre-roll advertisements of the focal firm Firm Consideration/visit Retargeting Consumer sees a retargeted banner advertisement of the focal firm Firm Consideration/visit Accommodation search Consumer searches via a search engine for accommodations Consumer Consideration Information / comparison search Consumer searches via a search engine for a comparison site Consumer Consideration Competitor search Consumer searches via a search engine for a competitors Consumer Consideration Focal brand search Consumer searches via a search engine for the focal firm Consumer Consideration Flight tickets search Consumer searches via a search engine for flight tickets Consumer Consideration Generic search Consumer searches via a search engine for generic information Consumer Consideration
Accommodation website Consumer visits an accommodation website Consumer Visit
Accommodation app Consumer uses an accommodation application Consumer Visit
Information / comparison website Consumer visits an information / comparison website Consumer Visit Information / comparison app Consumer uses an information / comparison application Consumer Visit
Competitor website Consumer visits an website of an competitor Consumer Visit
Competitor app Consumer uses an application of an competitor Consumer Visit
Focal brand website Consumer visits the focal firm website Consumer Visit
Flight tickets website Consumer visits a flight tickets website Consumer Visit
12 3.2 Plan of analysis
The objective is to investigate where, within a customer journey, a firm-initiated touchpoint influences the conversion probability. In order to analyze this phenomenon and to test the hypotheses, three main methods will be used for estimation; Markov chains, logistic regression and multinomial regression. Two logit models will be created to determine the predictive ability of the firm-initiated touchpoints within the consideration stage and one multinomial model for the visit stage. While a Markov chain model will be used to determine the overall contribution of a firm-initiated touchpoint on the conversion probability. Table 3 is an overview of the models, which are going to be discussed in the next sections.
Table 3:
Overview of the models used
Model Dependent variable Stage Answer hypothesis
(1) Logit model Conversion probability Consideration stage H2a & c, H4a, H6a (2) Logit model Transition probability Consideration stage
(3) Multinomial model
Conversion probability Visit stage H1, H2b & c, H3a & b, H4b & c, H5a, b, & c. H6b
(4) Markov chain Transition probabilities Both stages H6a & b
3.2.1 Logit models and the multinomial model
To determine, whether the firm-initiated touchpoints have a significant and positive effect at each stage, both a logistic regression model and a multinomial regression will be used. For the consideration stage two logistic regression models will be estimated. The consideration stage has two possible outcomes, either the consumer moves towards the visit stage or the journey ends. Therefore a logistic regression will suffice (Leeflang, Wieringa, Bijmolt, & Pauwels, 2015). The first model will be estimated for the conversion probability. The second model will be estimated for the transition probability of going to the subsequent visit stage. Via estimating both models, it is possible to infer the effect of firm-initiated touchpoints on moving further through the customer journey and on the effect on conversion probability. It could be that for example a banner in the consideration stage is highly significant in guiding the consumer towards the next stage, while it decreases the overall conversion probability.
For the visit stage a multinomial regression model will be estimated. Which has an extra possible outcome. If the journey moves towards the purchase stage it ends with a
13 moves back to the consideration stage. Therefore a logistic regression does not suffice for this stage, since there are three possible outcomes instead of two. A multinomial regression is suitable for estimation, when an individual chooses between a certain number alternatives (Leeflang et al., 2015).
The first logistic regression will be used to estimate the conversion probability at the consideration stage. Anderl et al. (2016) test two types of logistic regression in analyzing the customer journey. One model is estimated based on the amount of clicks on each channel and their second model uses an order effect which reflects the sequence of channels within a customer journey. The second model outperforms the first one in predictive ability (Anderl et al., 2016). Since, the focus of logit models is on the predictive capabilities (Leeflang et al., 2015), the second model will be used within this study. The first logit model, which estimates the conversion probability is:
!"#$%(') = * + - .(/012) 3
045
605+ .(/017)607+ .(/015)602+ ./060/ (1)
Y = the customer journey ends with a purchase or without a purchase
* = the error term
.4i = the coefficient of the last touchpoint xi within the sequence
.4i-1 = the coefficient of the second-last touchpoint xi within the sequence
.4i-2 = the coefficient of the third-last touchpoint xi within the sequence
.4i-3 = the coefficient of the fourth-last touchpoint xi within the sequence
xi1,2,3,4 = Touchpoint xi out of the n options that occur in the consideration stage (see table 2).
With the first model, the last four touchpoints of the consideration stage are taken into account to make the number of variables within the logit model tractable (Anderl et al., 2016). The dummy variable xit is coded 1 if a channel is present at position t and if not it is set to 0. The dependent variable (Y) is either a 1 meaning that the consumer has converted or a 0 which means that the journey did ended without a purchase. The second model, estimating the transition probability to the visit stage is almost identical to the first formula. Only the
14 !"#$%(') = * + - .(/012)
3 045
605+ .(/017)607+ .(/015)602+ ./060/ (2)
Y = the consumer moves towards the visit stage or the journey ends
The third model is going to be estimated for the visit stage and since that stage as one extra possible outcome, the third model will be a multinomial logit. Not only the dependent variable is different. This model will be estimated with different data/touchpoints. The formula for the third model is:
:('$ = ;) = * + - .(/012) 3
045
605+ .(/017)607+ .(/015)602+ ./060/ (3)
P(Yi = k) = the probability that the observed outcome (Yi) is equal to one of the possible outcomes k (ending the journey with purchase | ending the journey without purchase | moving back to the consideration stage)
xi1,2,3,4 = Touchpoint xi out of the n options that occur in the visit stage (see table 2).
The estimated coefficients of the logistic regression and the multinomial logistic regression can’t be directly interpreted. They can only be used to derive the direction of the effect (positive or negative), whether the effect is significant and to compare the magnitude of the significant effects with each other (Leeflang et al., 2015). For a more meaning full
interpretation the odds ratio is going to be used. Which measures the probability of Yi=1
divided by the probability of Yi = 0. If an odds ratio is between 0 < 1, it means a negative
relation of the variable. An odds ratio of 1 means no relation and an odds ratio > 1 means a positive relation (Leeflang et al., 2015). Consider the following example, where purchase is coded to 1 and no purchase to 0. An odds ratio of three for a variable would mean that the probability for conversion is 300% larger than for no conversion. A odds ratio of 0.25 would mean that the odds for conversion have decreased with 75%.
3.2.2 Higher-order Markov Chain
15 into account. A method for analyzing time series data, especially if one is dealing with a categorical data sequence, is a higher-order Markov Chain model (Ching & Ng, 2006). A higher-order Markov Chain assumes that each data point takes (X(n)) a value in the set; M = {1,2,……, m}. Where m is finite, in the case of the data set at hand m is a touchpoint out of the finite options displayed in table 2. With a Markov Chain model in general, transition probabilities are estimated between states (touchpoints). In other words; what is the
probability of going to touchpoint Z given that a consumer is at touchpoint Y. With a higher-order Markov chain it is possible to estimate the transition probability on a sequence of states. So for example; what is the probability of going to touchpoint Z given that a consumer went through touchpoints W, X, Y. The formula to calculate transition probabilities is (Ching & Ng, 2006): :=>(3) = ? @ A >(315) = ?5, … , >(31D) = ?D) = - E0 D 045 FGHGI (4) Where: - E0 D 045 = 1 (5)
And Q = [qij] is a transition matrix where each column sums equal to one, such that:
0 ≤ - E0
D 045
FGHGI ≤ 1, ?@, ?0 N O (6)
Formula (4) represents the probability of a data point (X(n)) being the current touchpoint (j 0) based on the previous sequence of touchpoints (X(n-1) = j1, … , X(n-k) = jk). Where k represents the number of touchpoints one looks back into the journey. So if the current touchpoint is being estimated based on the previous four touchpoints, k equals four. A condition of formula (4) is formula (5). A Markov Chain is estimated with a Poisson distribution, so formula (5) should equal one. Since all the probabilities of all the possible outcomes should equal 1 (Ching & Ng, 2006). Furthermore, formula (6) represents that the probability that the current touchpoint is (j0) given the previous sequence of touchpoints (ji), should be between 0 and 1. Given that all the touchpoints that are included are present in the defined set of touchpoints
16 To interpret the results of the higher order Markov chain, the removal effect is going to be used. The removal effect is the change in probability for reaching the conversion state from the start when a specific touchpoint is removed. In other words, it reflects the change in conversion rate if that specific touchpoint was not present (Anderl et al., 2016). Suppose email has a removal effect of 0.50. It would mean that if email was left out of all journeys, 50% of the conversions would be lost.
4. Data preparation and cleaning
Before diving into the results, a few adjustments to the dataset are discussed that were made prior to the estimation. First, the original dataset contained 31760 unique journeys across 9678 unique users. Of those 31760 journeys only 2080 contained at least one firm-initiated touchpoint. Since the focus is on determining the effect of firm-initiated touchpoints within this study, those 2080 journeys contain valuable information. However, journeys that do not contain a firm-initiated touchpoint can also be very valuable for determining the effects. On the other hand, too include all 29680 journeys that do not contain a firm-initiated touchpoint might be unnecessary and inefficient during the estimation. Therefore 2080 journeys will be randomly selected out of those 29680 journeys that do not contain a firm-initiated touchpoint to create a 50-50 split in the dataset.
17 the touchpoint before and the time difference was smaller than 118 seconds, it got removed. In total 540.858 subsequent touchpoints were removed in order to debias results.
Lastly, the dataset was divided into two subsets. One containing all the consideration touchpoints and another containing all the visit touchpoints. To illustrate, consider the following customer journey which displays a hypothetical example of a journey which ends without a purchase:
C > C > F > F > V > V > F > C > V > V
Where C is a consumer-initiated touchpoint that is classified in the consideration stage, where F is a firm-initiated touchpoint and where V is a consumer-initiated touchpoint that is classified in the visit stage. See table 2, for an overview of the classifications and which touchpoints C, V or F could represent. All the firm-initiated touchpoints are classified according to the previous touchpoint. So the first firm-initiated touchpoint in the example gets classified into the consideration stage and therefore the second firm-initiated touchpoint as well. And via the same logic, the third firm-initiated touchpoint gets classified into the visit stage. When a journey starts with a firm-initiated touchpoint it gets assigned to the
consideration stage. For the example it would mean that this single customer journey would exist out of four separate sequences, namely; sequence 1 ( C > C > F >F), sequence 2 ( V > V > F), sequence 3 (C) and sequence 4 ( V > V).
The data preparation and cleaning resulted into three different datasets, see table 4. The first data set contains all touchpoints which will be used to estimate the Markov model in order to analyze the entire journey. The second dataset is a subset of the first one. It contains all the consideration touchpoints and is used to analyze the consideration stage. The third dataset is also a subset of the first one but contains all the visit stage touchpoints and is used to analyze visit stage.
Table 4:
Overview of the datasets
Dataset Used for estimating stage Users Journeys Paths Dataset 1 Visit and consideration stage 3134 4160 4160
Dataset 2 Consideration stage 1387 1756 10273
18
5. Results
The results are analyzed per section per stage. First the consideration and visit stage are analyzed in isolation and afterwards the entire journey is analyzed based on the Markov model. Each section begins with validating and comparing different models, next the face validity of the estimated model is briefly discussed and afterwards the hypotheses are tested. This chapter ends with obtaining results in a meaningful way per firm-initiated touchpoint.
5.1 Results for the consideration stage
5.1.1 Validating the conversion probability model
In order to validate the estimated logistic regression model, the base model (see formula 1 at the chapter methodology), other models were estimated to compare it’s performance. Similar too Anderl et al. (2016) the base model is going to be compared to a count model. In which the independent variables represent the number of occurrences of each touchpoint. The base model is also going to be compared to a simplified version in which some consumer-initiated touchpoints are grouped. The fourth and last comparison model is similar as the base model, the only difference is an extra variable which accounts for the length of the sequence. See table 5 for an overview of the models used for comparison.
Table 5:
Models to compare the Logistic regression with Model: Independent variables:
Model 1A The base model, as described in formula 1 (see methodology) Model 1B Count variables which represents how many times each touchpoints
occurs in the sequence.
Model 1C Similar as model 1, only now grouped variables for the consumer-initiated touchpoints to decrease the used degrees of freedom. Model 1D Similar as model 1, only added one extra variable “length” which
represents the amount of touchpoints in the sequence.
19 model 1A and 1D are preferred over model 1B and 1C. While the hit rate is very similar across all four models. A likelihood ratio test could determine which model fits the data the best. A prerequisite of the likelihood ratio test is that models are nested within each other (Leeflang et al., 2015). Only model 1A and 1D are nested within each other. However it is still possible to perform a likelihood ratio test between non-nested models that are estimated via maximum likelihood (Vuong, 1989). The likelihood ratio test for non-nested models indicates that model 1A fits the data better than model 1B (z = 10.94, p < 0.01) and better than model 1C (z = 6.91, p < 0.01). In order to determine whether there is a significant difference between model 1A and model 1D the normal likelihood ratio test is used. The likelihood ratio test between 1A and model 1D indicates that model 1D is significantly better than model 1A (X2 = 2.82, P < 0.05). Therefore, model 1D will be used for analyzing the
results. Model 1D is similar as formula 1 (model 1A) described at the methodology, it only includes a single extra parameter which accounts for the length of the sequence.
Table 6:
Comparison of the logistic regression models
Sequence model (1A) Count model (1B) Sequence grouped (1C) Sequence + length (1D) AIC 1351 2069 1593 1328 BIC 1636 2152 1794 1620 Hit rate 98% 97% 98% 98% McFadden R2* 0.40 0.03 0.27 0.41 Nagel R2* 0.43 0.04 0.30 0.44 Cox Snell R2* 0.10 0.01 0.07 0.10
*R2 are pseudo R2 measures instead of the regular measure of variance
5.1.2 Face validity of the logistic regression
20 competitor search touchpoint in the conversion probability model and the transition
probability model. As well as for email in the transition probability model.
5.1.3 Hypotheses testing via the logistic regression
The results of model 1D, the best model out of the four models compared to each other, are displayed in table 7. Since the touchpoints are dummy coded for each point in time within the sequence, the estimated effects for the touchpoints are based on a reference level. The
reference level used for estimation of the results is a “generic search”. So for each significant effect, it means that for that point in time the variable has a significantly different effect on conversion probability than a generic search. Where time represents the point in the sequence and only the last four touchpoints are taken into account. So b4iXi4 is the last touchpoint in the
sequence. b(4i-1)Xi4 the second last and so on. Only the significant effects are displayed, in
order to keep the results more clearly due to the amount of estimators. Important to note, as discussed at the methodology: the coefficients in table 7 can only be interpreted as a direction of the effect, positive or negative. It does not represent a change in conversion probability. For a more meaningful interpretation the odds ratio is going to be used later on.
Within the results, displayed in table 7, no evidence is found to accept H2a which states that banner advertisements will have a significant effect on conversion probability within the consideration stage. At every point within the sequence the effect of a banner was not significantly different than a generic search (p > .10). Subsequently the results of model 1D do not provide any evidence to accept H2c which states that banner advertisements will be more effective in the consideration stage than in the visit stage. Since no significant different effect were found. Furthermore, the results do not provide evidence to accept H4a which states that pre-roll advertisements will have a significant and positive on conversion probability.
21 Table 7:
Results of the logistic regression model (1D) on conversion probability
Time / point in sequence
Variables b b(4i-3)Xi1 b(4i-2)Xi2 b(4i-1)Xi3 b4iXi4
Intercept -3.80***
Length of path -0.005**
Accommodation search 0.955***
Information/comparison search 2.54*** 1.67*** -1.98***
Competitor search 3.057*** -2.46***
Focus Brand search 3.03*** 1.41***
Flight tickets search -1.94***
Affiliate 1.77***
Email 1.96** 1.23*
Retargeting 3.23*** -3.14** 1.73***
* P < 0.1 ** P < 0.05 ***P < 0.01
Table 8:
Results of the logistic regression model (1D) on transition probability
Time / point in sequence
Variables b b(4i-3)Xi1 b(4i-2)Xi2 b(4i-1)Xi3 b4iXi4
Intercept 1.45***
Length -0.005***
Accommodation search 0.63*** 1.60***
Information/comparison search 1.45*** 0.87***
Competitor search 0.45** -0.46** 1.93***
Focus brand search 2.44*** 1.43*** -2.12***
Flight tickets search 0.25* 0.27**
Banner 1.85** -1.97**
Email 1.67* -2.38** 3.11**
Pre-Roll advertising -1.32***
Retargeting -0.69**
22
5.1.4 Comparing the results with the transition probability model
To compare whether effects of touchpoints differ between moving further into a customer journey and the conversion probability. A model to estimate the probability of moving towards the visit stage is estimated (see formula 2) in order to compare results with the conversion probability model. The transition model is the same as the model of which the results were discussed in section 5.1.3, the only difference is the dependent variable. Which is the transition probability towards the visit stage, instead of conversion probability. Table 8 displays the results of the transition probability model.
As mentioned in the previous section, no significant effects were found for banner advertisements on conversion probability. However the results in table 8, provides evidence for significant different effects for banner advertisements in comparison to a generic search on the probability of moving towards the visit stage. At the begin of the sequence a banner has a positive effect on the transition probability (b = 1.85, p < 0.05). Moreover, further in the sequence a negative effect on transition probability has been found (b = -1.97, p < 0.05). 5.2 Results for the Visit stage
5.2.1 Validating the multinomial logistic regression
In order to validate the multinomial logistic regression model, the base mode (see formula 3 at chapter 3 methodology), other models were estimated in order to compare it’s performance. Similar as for the consideration stage, the sequence model is going to be compared to a count model which counts the amount of occurrences of each touchpoint in a unique sequence. Both the sequence model and the count model are going to be compared with a simplified model in which some independent variables are grouped. Since for example, the competitor website and their app are registered as two separate touchpoints. Within the grouped models, the website and app touchpoints are combined into one variable in order to simplify the model.
Table 9:
Overview of the models for comparison Model: Independent variables:
23 Table 10 displays measures to compare the models with each other. Based on the AIC criterium. Model 2A and 2B seem to have a better trade-off between model fit and complexity than model 2C and 2D. The hit rate however seems rather similar across all models as well as the pseudo R2 measure of McFadden. Since, among all the measures used to compare the
models, difference seems rather small the likelihood ratio test for non-nested models could be decisive. The likelihood ratio test for non-nested models indicates that model 2A significantly is a better fit than model 2B (z = 4.30, p < 0.05), while the AIC value of 2A is higher than for 2B. Moreover, it also indicates that model 2A outperforms 2C (z = 2.38, p < 0.05) and model 2D (z = 2.64, p < 0.05). Hence model 2A will be used for estimation and analyzing the data.
Table 10:
Measures to compare the multinomial logit models
Sequence model (2A) Grouped Sequence (2B) Count model (2C) Grouped count model (2D) AIC 17.797 17.781 17.916 17.924 Hit rate 70,5% 70,5% 71,4% 71,3% McFadden R2* 0.076 0.073 0.062 0.060 Nagel R2* 0.135 0.131 0.111 0.109 Cox Snell R2* 0.103 0.099 0.084 0.083
*R2 are pseudo R2 measures instead of the regular measure of variance
24
5.2.2 Face validity of the multinomial regression
For the visit stage (see table 11), the results for the firm-initiated touchpoints show no
inconsistencies across their point in the sequence on a given outcome. All significant different effects are either positive or negative. However the results show some unusual aspects, since all effects are set to the reference level competitor website. It seems very strange that pre-roll advertising only has significantly different negative effects on conversion probability relative to a visit to the competitor website. This would mean that a visit to the website of a
competitor is better for the conversion probability at the focal firm, than seeing a video advertisement from that same focal firm. Moreover, all firm-initiated touchpoints have significantly different positive effects on ending the journey without a purchase compared to visiting the website of a competitor.
5.2.3 Hypotheses testing via the multinomial regression
The results for the estimated model (model 2A) are displayed in table 11. Since it is a multinomial logistic regression, the estimates are relative measures of effects to one of the possible outcomes. The reference level for the outcome variable is set to going back to the consideration stage. So for example a positive significant effect for a banner advertisement on the purchase outcome means that the banner advertisement has a significantly different effect on purchase than on moving backwards to the consideration stage. Furthermore, the
independent variables are dummy coded. Therefore all estimated effects are also relative to a reference level. All estimates are relative to the touchpoint “competitor website”. Lastly the coefficients in table 11 can only be used to determine the direction of the effect, either positive or negative. Which is the same interpretation as for the coefficients of the normal logistic regression.
25 Table 11:
Results of the multinomial logistic regression
*P < 0.10 **P<0.05 ***P<0.01
Ending journey with a purchase at the focal firm Ending journey without a purchase at the focal firm
Time / point in sequence Time / point in sequence
Independent variables b b(4i-3)Xi1 b(4i-2)Xi2 b(4i-1)Xi3 b4iXi4 b b(4i-3)Xi1 b(4i-2)Xi2 b(4i-1)Xi3 b4iXi4
Intercept -3.610*** -1.241*** Accommodation website 0.517*** 0.398*** 0.325*** -0.100* Accommodation app -12.13*** -12.18*** -16.12*** -19.80*** Information/comparison website 0.452* 0.474* -0.52** 0.211** -3.518*** Information/comparison app -13.81*** -13.71*** Competitor App 1.545*
Focal brand website 1.074*** 0.9958*** 0.6934*** 1.112*** 0.319** -0.746***
Flight tickets website 0.739*** -0.847** 0.579*** 0.3917*** 0.212* -0.236***
Flight tickets app 0.826*
Affiliate 1.517* 1.29*** 1.07**
Banner 0.757** 0.761** 0.6914**
Email 1.253* 1.145*** 0.842***
Pre-roll -9.39*** -9.812*** -11.92*** 0.8617** 0.684* 1.318*** 1.339***
26 The results in table 11 provide no evidence for H2b, which states that banner
advertisements will have a significant positive effect in the visit stage on conversion probability. At no point within the sequence, a banner advertisement has a significantly different effect on purchase compared to moving back to the consideration stage (p > 0.10). While banner advertisement has a significant influence and positive effect on ending the journey, compared to going back to the consideration stage at the first (b = 0.757, p < 0.05), second (b = 0.761, p < 0.05) and last point in the sequence (b = 0.691, p < 0.05). Which provides some support for H2c, which states that banner advertisements will be more effective in the consideration stage than in the visit stage. Since in both stages no significant effect was found on conversion probability. And in the visit stage a banner increases the probability of ending the journey compared to going backwards.
Moreover, the results in table 11 provide evidence to support hypothesis H3a which states that retargeting will have a significant and positive influence on conversion probability during the visit stage. Since at the first point of the sequence a significant and positive effect is found (b = 0.81, p < 0.05) and at the last point of the sequence (b = 0.78, p < 0.05).
However retargeting also increases the probability for ending the journey without a purchase during the visit stage at the first (b = 0.60, p < 0.01) and the last (b = 0.95, p < 0.01) point of the sequence. Therefore hypothesis H3a is partially accepted. The results of table 11 do not provide sufficient evidence to support H3b which states that the effect of retargeting will diminish when the consumer moves closer towards a purchase decision. Although the effect at the last point is smaller than at the first point, since no effect is found in between there does not really exist a diminishing effect.
Furthermore, the results in table 11 provide no evidence to support H4b, which states that pre-roll advertising will have a significant and positive effect on conversion probability. The results out of table 11 suggest the opposite. For the second point (b = -9.39, p < 0.01), the third point (b = -9.812, p < 0.01) and for the last point (b = -11.92, p < 0.01), it indicates a negative effect on conversion probability. On the other hand, it increases the probability of ending the journey without a purchase at the first (b = 0.861, p < 0.05), second (b = 0.684, p < 0.10), third (b = 1.31, p < 0.01) and last point of the sequence (b = 1.33, p < 0.01). For
27 For H5a which states that email will have a positive effect on conversion probability in the visit stage, some evidence to support it is presented in table 11. In the second point of the sequence, email has a positive significant effect on conversion probability relative to moving back and to the website of a competitor (b = 1.253, p < 0.1). Since only in the second point a significant and positive effect was found on conversion probability, table 11 does not provide evidence to support H5b. Which states that email will have an increasingly positive effect on conversion probability when a consumer is in the visit stage.
Lastly, table 11 provides partial evidence to support hypothesis H6b, which states that consumer-initiated touchpoints will have a greater significant influence on conversion
probability than firm-initiated touchpoints during the visit stage. Since, table 11 shows that for consumer-initiated touchpoints 15 significant different effects are found on conversion. While for firm-initiated touchpoints only 7 significant different effects are found on
conversion.
5.3 Results of analyzing the entire journey
5.3.1 Validating the model
To determine which order for the Markov model is most suitable, various different orders are compared with each other via the area under the ROC curve; the AUC measure. Which is one of the best ways to evaluate a classifier model on performance (Bradley, 1997). The model with the greatest area under the curve, has the best performance in classifying the outcome correctly. Via maximizing the AUC, a third order Markov chain has the best performance (with a maximum order of 10). Figure 2 displays the ROC curve for the first order, the second order and the third order Markov models. The AUC value for the third order Markov model is 0.95 which is distinctively higher than the second order (AUC = 0.86) and the first order (AUC = 0.78). An AUC value of 0.95 means that the third order Markov model classifies 95% of the data correctly. So a third ordered Markov model seems the best fit.
28 they predict almost every journey to end without a purchase. With a hit-rate threshold of 0.30. And only 92 out of the 4160 journeys end with a purchase.
Figure 2: ROC Curve
5.3.2 Results of the Markov Chain
Table 12 displays how many conversions each method contributes to each channel.
Interesting to see is that with the Markov model only 18% of the conversions get contributed to a firm-initiated touchpoint. Furthermore the last click attribution seems to over evaluate the accommodation website, the focal brand website and retargeting. While it underestimates generic searches and the flight tickets website. The first click attribution seems to over evaluate the focal brand website and email. While it underestimates, the
information/comparison website and generic searches.
29 effective. The removal effects provide more evidence to support H6b, which states that
consumer-initiated touchpoints will have a significantly greater effect on conversion probability than firm-initiated touchpoints during the visit stage. The touchpoints
accommodation website (0.89), information/comparison website (0.76), competitor website (0.92), focal brand website (0.71) are all more effective than any of the firm-initiated touchpoints.
Table 12
Conversion attribution and removal effect
Conversion attribution Removal effect
Channel Third order Markov model Last click attribution First click attribution Third order Markov model Accommodation website 13% 18% 12% 0.89 Accommodation app 2% 0% 0% 0.12 Accommodation search 3% 0% 1% 0.19 Information/comparison website 11% 8% 4% 0.76 Information/comparison app 2% 0% 1% 0.11 Information/comparison search 1% 0% 0% 0.08 Competitor website 13% 13% 13% 0.92 Competitor app 1% 0% 0% 0.03 Competitor search 2% 0% 0% 0.14
Focal brand website 11% 39% 46% 0.78
Focal brand search 1% 0% 0% 0.05
Flight tickets website 10% 4% 3% 0.71
Flight tickets app 1% 0% 0% 0.06
Flight tickets search 3% 0% 0% 0.18
30 5.4 Meaningful results for the firm-initiated touchpoints
5.4.1 Affiliate marketing
For affiliate marketing the odds are 590% higher for conversion than a generic search in the last point of the consideration stage (odds = 5.90). For all the other points in the sequence, no significantly different effect on conversion is found compared to a generic search. Moreover, affiliate marketing does not significantly increases or decreases the odds for transitioning to the visit stage compared to a generic search.
Table 13
Odds ratio for the firm-initiated touchpoints
Time / point in sequence b(4i-3)Xi1 b(4i-2)Xi2 b(4i-1)Xi3 b4iXi4
Consideration stage: Conversion probability Affiliate 5.90 Banner Email 7.15 3.43 Pre-roll Retargeting 25.49 0.04 5.64 Consideration stage: Transition probability Affiliate Banner 6.40 0.13 Email 5.34 0.09 22.50 Pre-roll 0.26 Retargeting 0.49 Visit stage:
Purchase at focal firm
Affiliate 4.55 Banner Email 3.50 Pre-roll 0.00 0.00 0.00 Retargeting 2.26 2.21 Visit stage:
31 For the visit stage affiliate marketing the odds of conversion are 455% higher (odds = 4.55) than moving backwards to the consideration stage compared to a visit to a competitors website. Whilst, the odds of ending the journey are also greater than moving backwards into the journey. Since at the second last point the odds of ending the journey are 310% (odds = 3.1) higher and at the last point 292% higher (odds = 2.92) than moving backwards to the consideration stage. Compared to a visit to the competitors website. So on one hand, affiliate increases the odds for a purchase. On the other, it also increases the odds of ending the journey without a purchase, compared to moving back.
The results of the third order Markov model are similar. It only attributes 2% of the conversions towards affiliate marketing and the removal effect (10%) belongs to the lower categories. Which is not surprising since the odds ratios for the consideration and the visit stage show positive effects on conversion but also on ending the journey without a purchase.
5.4.2 Banner advertisements
For the banner advertisements no significantly different effect was found compared to a generic search on conversion, in the consideration stage. While, the odds for transitioning to the visit stage are 640% higher (odds = 6.40) than for a generic search at the first point within the sequence. However for the third point in the sequence the odds for transitioning are 87% lower than for a generic search (odds = 0.13).
Almost identical to the consideration stage, no significantly different effect was found for a banner advertisement on conversion probability in the visit stage. Only now compared to a visit to a competitor website instead of a generic search. On the other hand the odds for ending the journey without a purchase are greater for three out of the four points within the sequence compared to a visit to a competitors website. The odds are approximately 200% higher at the first point (odds = 2.13), the second point (odds =2.14) and the fourth and last point in the sequence (odds = 1.99).
Since there are no positive significantly different effects found for banner
32
5.4.3 Email
For email positive effects on conversion probability are found compared to a generic search within the consideration stage. The odds for conversion are 715% higher (odds = 7.15) at the first point of the sequence and are 343% higher (odds = 3.43) at the last point in the sequence. The odds for transitioning to the visit stage are somewhat reversed. The effect is stronger at the end of the sequence than at the beginning in the visit stage. The odds are 534% higher (odds = 5.34) at the first point in the sequence and even 2250% higher at the fourth and last point in the sequence (odds = 22.50). However at the second point in the sequence the odds for conversion are 91% lower (odds = 0.09).
For the visit stage, the odds of conversion are 350% higher (odds = 3.50) than moving backwards to the consideration stage, compared to a visit to a competitors website. While the odds are greater for conversion, they are also greater to end the journey without a purchase compared to moving back. At the third point within the sequence the odds of ending the journey without a purchase is 314% higher (odds = 3.14) than moving backwards. And at the fourth point in the sequence the odds are 232% higher (odds = 2.32) to end the journey without a purchase.
The Markov model confirms the results of the consideration and the visit stage to some extent. It only attributes 4% of the conversions towards email, which seems rather low since the odd ratio for conversion are high in both the consider and the visit stage. However the removal effect is 30%. Which means that without a single email across all journeys, 30% of the conversions would be lost.
5.4.4 Pre-roll
Pre-roll advertisements seem to have no significant different effects on conversion in the consideration stage compared to a generic search. However, on the second point in the
sequence the odds are distinctively lower (odds = 0.26). Meaning that a pre-roll advertisement decreases the odds for transitioning to the next stage with 74% compared to a generic search.
33
5.4.5 Retargeting
Retargeting increases the odds for conversion in the consideration stage by 2500% compared to a generic search (odds = 25.49) on the first point of the sequence. At the third point in the sequence it increases the odds for conversion by 560% (odds = 5.64). While at the second point it decreases the odds for conversion by almost 96% (odds = 0.043). It is strange that at the first and third point retargeting increases the odds of conversion whilst it decreases the odds at the second point. Multicollinearity could lead to such a ordinary result. However among all the independent variables there exist no multicollinearity. For the transition probability there is only a significant decreasing effect in odds. At the second point in the sequence retargeting decreases the odds for transitioning by 51% (odds = 0.49). The negative effect on transitioning probability is consistent with the negative effect on conversion
probability at the same point within the sequence.
In the visit stage, retargeting increases the odds for conversion by 226% at the first point of the sequence (odds = 2.26) and by 219% at the last point in the sequence (odds = 2.19), compared to moving back towards to the consideration stage. However retargeting also increases the odds for ending the journey without a purchase with 183% at the first point (odds = 1.83) and with 259% (odds = 2.59) at the last point of the sequence. So while
retargeting increases the odds for conversion it also increases the odds for ending the journey without a purchase, compared to moving back towards the consideration stage.
The Markov model attributes the most conversions towards retargeting (8%), among all the firm-initiated touchpoints. The removal effect for retargeting is also the highest for retargeting (58%) compared to the other firm-initiated touchpoints. Which suggests that retargeting is the most effective firm-initiated touchpoint in converting consumers.
6. Conclusion & managerial implications
The goal of this study, as mentioned in the introduction, is to generate insights in how the timing of firm-initiated touchpoints influence the conversion probability within a customer journey. For each firm-initiated touchpoint the effectiveness is going to be discussed in a separate section and the consumer-initiated touchpoint are going to be discussed in a more general manner in a single section. Each section ends with relevant managerial implications.
6.1 Affiliate marketing
