THE IMPACT OF ONLINE MARKETING ON THE ONLINE CUSTOMER JOURNEY: THE INFLUENCE OF FIRM-INITIATED CONTENT ON THE CUSTOMER STAGES
by
John J. Malkoun
University of Groningen Faculty of Economics and Business
ABSTRACT
We draw the picture of the online customer journey as a series of customer-initiated contents symbolizing the journey steps on which firms exert pressure using firm-initiated contents to convert consumers into visitors, then buyers. We then look at the effectiveness of those tools on the probabilities to transition from one step in the journey to another. We base our approach on literature, design a Markov Chain attribution model to describe the researched phenomena, illustrate the journey using transition matrices, compare the transition probabilities using t-tests and a journey without any firm-initiated content as a control group, and finally, inspect the specific stage differences. We showcase and summarize the wide range of results, and evaluate the model used using evaluation criteria and a cross-validated ROC curve analysis. Managerial implications, limitations and recommendations for future research are eventually listed.
Keywords: Online; journey; customer; channel; attribution
Research Theme: Conversion attribution and the consumer purchase journey Seminar Supervisor: Dr. Peter van Eck
INTRODUCTION
When a customer interacts with a form of advertisement (e.g. banner ad) or embarks on a path toward a firm’s website (e.g. search engine), these become channels leading to that customer’s conversion (Martin, 2009; Mulpuru et al. 2011). Understanding how customers interact with these channels during their journey is important for many firms because online advertisement spend is an essential factor in their marketing mix and budgeting decisions (Raman et al. 2012). In fact, the Marketing Science Institute has listed it as top research priority1.
Evaluating the effectiveness of and attributing credit to different channels are challenges that firms must face (Nelsin & Shankar, 2009). Even when this is achieved, an increasing number of channels, leads to increasing complex budgeting strategies (Raman et al. 2012). A step toward addressing these issues is understanding the relationship between the different types of channels and their respective effectiveness.
Studies have looked into the different roles of various online channels in a customer’s purchasing action (Neslin & Shankar, 2009; Verhoef, 2012) without considering the effect of marketing (advertising) on the purchase funnel or considering a multichannel credit attribution approach. Some studies solely focused on a particular stage of the online journey or a particular channel such as search engine advertisement (e.g. Agarwal, Hosanagar & Smith, 2011; Skiera & Nabout, 2013) or brand keyword paid search (Blake, Nosko & Tadelis, 2015; Jansen & Schuster, 2011)). Although insightful, this is not fully realistic since multiple stages usually precede the purchase decision (Frambach, Roest & Krishnan, 2007; Gensler, Verhoef & Böhm, 2012) and these touch points have different effects on conversion likelihood (Braun & Moe, 2013; Li & Kannan, 2014). Other research aimed at developing a framework to better define and predict the effectiveness of different channels in leading to conversion (Abhishek, Fader & Hosanagar, 2010; Anderl, Becker, von Wangenheim & Schumann, 2016). Despite groundbreaking models with strong practical relevance, these frameworks do not develop our understanding on the relationship between channels and how some types of channels influence the customer funnel stages represented by other types of channels (e.g. how retargeting affects moving to the consideration stage represented by price comparison websites). They inform us on how much each touchpoint
1
contributes to conversion and how important these touchpoints are to the model but not on how different types of touchpoints influence each other (e.g. retargeting and price comparison website visits).
As channels serve different purposes in a customer journey, it makes sense to classify them. Li & Kannan (2014), and de Haan, Wiesel & Pauwels (2016) classify channels into firm-initiated content (FIC) and customer-initiated content (CIC). The first type pushes a message to the customers (Shankar & Malthouse, 2007) and include, among others, display advertisement, affiliates and email. While the second requires a customer’s action (Li & Kannan, 2014) and include, among others, search keywords, comparison website and direct website URL entry.
In this paper, we use this classification to answer our research question: do the probabilities of moving from one CIC to another differ when a FIC is involved in the same journey? And, do the probabilities of moving from one CIC to conversion differ when a FIC is involved in the same journey?
The results will enrich our knowledge about the customer journey and the firm tools that influence the steps in that journey. We hope to further our understanding of the different types of channels and their roles in a customer journey using a real and rich dataset from a travel agency and a Markov chains statistical model. To the extent of our knowledge, research has yet to touch upon this topic. The closest are Li & Kannan (2014) who looked at the carryover and spillover effects of the different types of channels, and de Haan, Wiesel & Pauwels (2016) who compared the respective effectiveness of the two types in driving conversion.
Practitioners benefit from knowing how the different channels are related, how they perform and how influencing the customer journey through firm-initiated content affects the steps of that journey. This allows them to optimize their promotional strategies and budgeting decisions as they use marketing tools such as display ads or email ads when and where these will be effective. It also allows them to fairly distribute credits between channels.
LITERATURE
Customer Journey
Researchers have long tried to map the process through which a customer goes and which includes a purchase stage. Howard & Sheth (1969) suggested that consumers go through three phases: extensive problem solving, limited problem solving and routinized response behavior. The first step is characterized by consumers with no information about a market and its brands and who seek information. In the second stage, consumers have little or imperfect information about a market and seek comparative information about brands. In the last stage, consumers have enough information to make a purchase decision. Similarly, Lavidge & Steiner (1961) coined the AIDA model: attention-interest–desire–action which is a representation of the steps and mental states a consumer goes through. In multichannel marketing, Neslin et al. (2006) took inspiration on Howard & Sheth’s (1969) model to define the customer journey as problem recognition-search-purchase-after sales.
Standing on the shoulders of giants, we simplify this process for the purpose of our analyses. We break down the customer funnel into three steps (Figure 1): awareness (acquiring basic information about a market/brand), consideration (collecting enough information about a market/brand to be able to select a preference) and decision stage (choosing and purchasing a product). The latter includes conversion. In our case, conversion represents the purchase through the travel agency, the firm’s goal (Montgomery & Srinivasan, 2014). In this paper, we use the terms funnel and journey interchangeably.
Figure 1: Simplified Customer Journey
Different Types of Content
Firm-initiated content are, as their noun indicate, content such as display advertisement that is created and initiated by the firm in order to communicate a product or service to consumers (Shankar & Malthouse, 2007). Customer-initiated content is the result of a customer’s action (Li & Kannan, 2014) such as entering a comparison website or opening the firm’s application.
FICs and CICs have different purposes and results. Since they require a customer’s action, CICs when triggered indicate concern or interest from the customer (Bowman & Narayandas, 2001; Shankar & Malthouse, 2007). For instance, a customer who searches for flight tickets on a search engine either plans on traveling or is curious about this prospect. Both indicate a level of interest. Looking at the previously introduced customer journey, CICs concern customers in the consideration stage (Alba et al. 1997). Alternatively, since FICs push a message to consumers (Shankar & Malthouse, 2007), they can be used to raise awareness about a brand (first stage of the funnel). A banner ad about flight tickets for a summer trip to Argentina fulfills such a purpose. However, we can imagine situations where FICs influence customers in a later stage in the funnel: a customer considering a trip to Argentina sees the previously mentioned banner and decides to read the offer by clicking it. As such, CICs can be seen as a representation of the customer journey and FICs as an external influence on different stages of that journey (Figure 2).
Little research, that we know of, has inspected the influence of online FICs on the probabilities of moving from one CIC to another and on the contribution of CICs in conversion in an online multi-channel journey. We will, however, discuss the papers that studied the effect of a specific firm’s online marketing tool on its customer’s purchase funnel. The tools researched include paid search, display advertising, email, retargeting and affiliate marketing.
Paid Search
are consistent with the purpose of search engines which is the collection of information about service brands by consumers (Brown et al. 2007). Animesh, Viswanathan & Agarwal (2011), using empirical data from a firm in the mortgage industry, suggested that click-through rate is driven by a seller’s rank in a sponsored search listing, a consequence of paid search. Click-through is the result of a customer clicking on the link leading to a firm’s website. Thus, this action indicates interest in the form of information collection (Brown et al., 2007). With regard to performance, Rusmevichientong & Williamson (2006) have developed a model of keyword selection for profitable keywords proving that bidding on the appropriate keywords can positively affect conversion. Similar studies also developed frameworks to optimize advertising budgets through the allocation of spending on keywords (e.g. Kitts & Leblanc, 2004; Fruchter & Dou, 2005; Muthukrishnan, Pal, & Svitkina, 2007; Gopal, Li, & Sankaranarayanan, 2011). It is important to note that the effectiveness of keywords is heterogeneous as retailer-specific and brand-specific keywords indicate different purchase behaviors (Ghose & Yang, 2009). They still both indicate a position in the consideration and decision stages of the funnel. As such research showed that paid search can impact the customer journey and conversion. Although paid search is not part of the data we used, it is part of the scarce literature on the impact of a marketing tool on the customer journey, which may be relevant.
Display Advertising
the effectiveness of pre-roll ads and their impact on the customer journey has not been investigated by previous research, yet. Since such video platforms contain an endless catalog of videos, we can argue that they may be visited by consumers in any information-seeking stage (awareness and consideration). Of course, we also note that they may be visited by non-consumers. We posit that since they are a type of display advertisement, they have an effect similar to other display ads. In a nutshell, display ads influence the customer journey and conversion.
Retargeting
Retargeting is a personalized recommendation aimed at customers who previously visited a firm’s website (Lambrecht & Tucker, 2013). It has been proven more effective than generic ads on customers visiting review websites, and therefore in the consideration stage (Lambrecht & Tucker, 2013). An A/B test on individual-level data was also used in that case. Because retargeting targets customers who previously visited a brand’s website without purchasing, they may increase conversion using personalization (Dias et al. 2008; Linden, Smith & York, 2003). However, the types of studies leave room for extraneous effects such as consumers not being online and therefore not being able to convert (Lewis, Rao & Reiley 2011). Lambrecht & Tucker (2013) reiterate this concern. We admit that assumptions need to be made in experiments like these while considering the limitations. In conclusion, retargeting is an effective way to influence the customer journey and a likely driver of conversion.
Email Marketing
“Email marketing is to directly market a commercial message to people using email” (Wu, Li & Liu, 2015). Emails can be used to bring traffic to a firm’s website (Ansari & Mela, 2003), signaling a customer’s position in the consideration phase. The results of Anderl, Becker, von Wangenheim & Schumann’s (2016) algorithm suggested that email marketing is a facilitator between channels rather than a direct driver of conversion. Their Markovian attribution model indicated that the effect of email advertising is stronger when preceded by a CIC which resonates with our research topic which views the journey as a sequence of steps that can be impacted by an external marketing tool. In fact, email initially serves to capture attention and then interest in a brand (Ellis-Chadwick & Doherty, 2012). As such, we expect emails to be a facilitator in the customer journey increasing the likelihood to trigger other CICs and an indirect driver of conversion.
“Affiliate marketing is a prominent, contemporary type of performance-based internet marketing whereby a company compensates affiliates for each customer referred through the affiliate’s marketing efforts” (Gregori, Daniele & Altinay, 2014). The aim of affiliate marketing is to increase traffic to a company’s website and eventually increase sales (Malaga, 2007). Fox & Wareham (2010) also mentioned its potential to increase sales. Previous research however mainly relied on surveys and not real empirical data. In sum, we can assume that affiliate marketing influences the three steps of the purchase funnel. It can serve as to promote awareness: a fashion blogger introducing her readers to a beauty product and providing a link to that page or special promotional voucher to be used. A same blogger writing a guide to choosing a face cream (offering a coupon code) which will likely be read by consumers considering to buy one. Thus, we expect affiliate marketing to affect the customer journey and conversion.
Building on the conclusion of previous research on the customer funnel and on the role different types of channels play in that journey, we will use several advertising types, namely affiliates, email, pre-roll, banner and retargeting ads, to test our hypotheses on. We expect that these firm-initiated contents, when in a funnel, have a significant impact on the probabilities of moving from one customer-initiated content to another.
H1: Probabilities of moving from one CIC to another are significantly different when the FIC is used in the same journey than when it is not.
The hypothesis is tested for a. affiliates, b. email, c. pre-roll, d. banner and e. retargeting ads. For example, a customer is exposed to a pre-roll video after searching for flight tickets on an accommodation website. This customer later visits the travel agency’s website. As CICs can be of a great variety, we posit that the probabilities will be different but not necessarily higher.
In this research, we mainly focus on the impact of these specific FICs on a variety of CICs rather than some specific CICs: first, to keep the results manageable. Second, because it fits in our view of CICs as the customer journey and the latter can take many forms. Third, because it would require splitting the data in many smaller subsets, increasing information loss.
literature review to propose that CICs in journeys that include such a combination are more likely to lead to conversion.
H2: Probability of moving from a CIC to conversion is higher in journeys that include the FIC.
The hypothesis is tested for a. affiliates, b. email, c. pre-roll, d. banner and e. retargeting ads. Again, specific CICs are not differentiated based on literature, but rather compared.
Figure 2: Conceptual Graph
METHODOLOGY
Data
Description
between the 01-06-2015 and 31-09-2016. The time series and cross-sectional nature of panel data also allows for generalizability (Leeflang et al. 2015). The variables we use for analysis are the user identification (variable name: UserID), the purchase journey identification (PurchaseID), the Date (Date), the touchpoint type (type_touch) and the purchase on the focal brand’s website (purchase_own). Additional implied variables were created and are discussed later in this section. The touchpoints include customer-initiated contents (Table A) which as we discussed earlier, are synonym of consideration and/or decision steps in a customer journey and give a representation of that journey (Bucklin & Sismeiro, 2009). In this case, the journey ends with either a purchase on the focal travel agency’s website, a purchase on a competitor’s website or no purchase at all. Since purchase is difficult to measure passively, a survey was used for this variable. CIC touchpoints include among others accommodation website, comparison app, competitor tour operator/travel agency website, generic search and focus brand website. The other type of touchpoint available are advertisement contents (FIC): banner, affiliate, email, pre-roll video, and retargeting.
There are 9,678 distinct customers with 29,012 unique purchase journeys. 192 of these journeys end with a focus brand purchase and 2.080 of these contain at least one FIC. The FICs are distributed as follows: 1.580 affiliates, 1.809 banners, 2.835 emails, 1.929 pre-rolls. and 34.916 retargeting.
During the data processing (done before handing us the data), some restrictions were applied to keep data relevant for analysis. We list the most important: touchpoints were initially kept if they occurred during the 6 months leading to the end of the journey; journeys end when there is no activity for 4 weeks; advertisement touchpoints are kept only if they occurred within a journey or within 2 weeks preceding the start of the journey. Additional data cleaning was required for our analysis and is discussed hereafter.
Data Cleaning
In order to be able to draw a Markov Chain from the data, modifications needed to be made to run the appropriate algorithms and anomalies needed to be addressed.
First, the date was changed to the POSIXct format which recognizes both the date and the time. The date was later used to compute the time difference between touch points. A frequency table showed that half of the touchpoints happen within 33 seconds and 25% happen within 9 seconds. All were kept, as we believe that this makes sense. Outliers also existed in the time variables and number of touchpoints visited but were kept as they do not violate the restrictions formerly defined and they will not be included in the model as such, given the nature of Markov chains. The latter also accounts for similar touchpoints appearing in succession.
Second, since we are investigating the difference between journeys with FICs and those without, dummies were created for each FIC accordingly. A journey including a pre-roll will have a dummy of value 1 and one without will have a dummy of value 0. Separate subsets were later created according to this classification.
Finally, 2,955 journeys with only one touch point were removed as their weights in a multi-channel attribution model will be misleading since they count as 100% attribution in some journeys and much less in others (Anderl, Becker, von Wangenheim & Schumann, 2016). They also offer no insight into the effect of FICs on the journey since there is only one touch point. Additionally, 534 and 1,809 observations were removed because they respectively contained touchpoints 11 and 17 which no information is available on.
Table B: FICs and Number of Journeys With At Least One Appearance
Method
In order to analyze the customer journey, we used a graph-based Markovian model. The strong suit of Markov chains is their ability to represent dependencies between sequences of observations of random variables (Anderl, Becker, von Wangenheim & Schumann, 2016; Archak, Mirrokni & Muthukrishnan, 2010). In our case, the sequential touch points in a journey. They have previously been used many times in marketing research (Styan & Smith, 1964; Homburg, Steiner & Totzek, 2009; Pfeifer & Carraway, 2000; Bronnenberg, 1998; Che & Seetharaman, 2009).
A Markov model M is defined by a set of states S and a transition matrix W. The states, here, are the different touch points in the customer journey. They include all FICs and CICs, as well as the START of the journey, a CONVERSION at the end or NULL if no conversion occurred. The latter is included because by looking at conversion, we are also interested in non-conversion. All of the previous states are function of a complete probabilistic model (Anderl, Becker, von Wangenheim & Schumann, 2016).
S = {s1, …, sn}
The transition matrix lists all the possible states and the probabilities pij of transitioning
from one state to another, given the Markov chain {X0, X1, X2, . . .}. The transition matrix will help
us test the first hypothesis as it shows the direct impact on the probabilities of moving from one touch point to another, when FICs are added or removed.
pij = P(Xt+1 = j | Xt = i) for i,j ∈ S, t = 0,1,2,...
Xt+1: The TO states are the columns of the matrix
Our model is of first order, meaning it presumes that the present depends on the last observation only. This has been proven to offer a better balance between accuracy and stability (Anderl, Becker, von Wangenheim & Schumann, 2016). The reason is that when the order increases, the number of independent parameters (the lagging observations) increases exponentially and the model becomes too complex to run efficiently (Berchtold & Raftery, 2002). In order to find out the relative importance of the touch points in conversion, and their structural relationship, the principle of removal effects is used: touch points are removed from the Markov graph consecutively to see how much value is lost (Anderl, Becker, von Wangenheim & Schumann, 2016; Archak, Mirrokni & Muthukrishnan, 2010).. The idea is that if a number of N conversions is reached without a touchpoint, compared to T conversions when that touchpoint is included, the latter reflects the change T-N in conversions. Once all touchpoints go through this process, they need to be weighted because the total sum of T-N would be bigger than T. This will be represented in percentage in our graph. By splitting the dataset in a subset that includes FICs and one that doesn’t, we use this approach adapting a script by Sergey Bryl and the ChannelAttribution R library by Davide Altomare & David Loris (2019). This would be similar to an uncontrolled A/B test where some journeys were exposed to a firm advertisement and others were not.
RESULTS
Results
In this section, we reviewed the results of the models by looking at the transition matrices extracted, testing for significant differences using t-tests and later evaluating the relevance and validity of our model using well documented criteria and techniques.
We were interested in relevant changes in transition likelihood between journeys since this would indicate a systematic difference between journeys likely linked to the presence or absence of firm-initiated content. As there are 17 stages (15 CICs, conversion, and null) we only report the significant t-tests and the highest changes in probabilities. Since the t-tests offer a wider perspective (comparing the means of all the probabilities of transitioning to each stages), we also look at visual representations of the transition matrices for granular results. When doing the latter, we set a 5% margin before reporting a significant change between journeys, because we deemed lower changes not high enough to infer that the transition likelihood differs between journeys.
Average Differences Between Journeys
First, comparing the means of all the probabilities of the journeys with FICs and the one without any, results in only a significant difference for the journey with email (t(439.31) = -2.02, p = .04). This signifies that on average transition probabilities are higher in journeys with email (9.29%) than in journeys without (6.72%).
Diving into the stages (represented by CICs) of the journeys, starting with the one including affiliates, the only significant t-test was for the transition to a generic search (t(28.01) = 2.36, p = .03). The probabilities leading to a generic search are higher in journeys without any FIC (8.83%) than in journeys with affiliates (6.71%).
Journeys with banners have significantly different probabilities leading to accommodations website than journeys without any FIC (t(13.07) = 2.23; p = .04). On average, probabilities in the former are lower (0.10%) than in the latter (2.74%).
The journeys with pre-rolls do not contain specific stage transition significantly different from the control journey.
Last, journeys with retargeting show a significantly different likelihood of moving to a generic search compared to the control journey (t(28.39) = 4.76, p < .01). It is lower in journeys with retargeting (4.70) than in the control journey (8.83).
Unfortunately, for all journeys, comparing the likelihood to convert or not, using t-tests, was not possible due to the lack of observations containing conversions.
Specific Stages Differences Between Journeys
We describe in details the different journeys. Starting with the journeys without any FICs (n = 24,471), we notice that generic search (p = .73) and information/comparison apps have high a high probability of leading to an accommodation website (Figure 3). There is a very high probability of moving from the focus brand search to its website (p = .76). Similarly, there is a .78 probability of moving from an information/comparison search to an information/comparison website. Another notable point is that many touchpoints are likely--to different degrees--to lead to a competitor’s website. The more likely being competitor search (p = .64), the focus brand website (p = .46) and an accommodation website (p = .42). Most touchpoints seem to almost never lead to conversion. This can, again, be due to the original low proportion of conversion in the data (0.73%).
Figure 3: Transition Matrix in Journeys Without Any FIC
Figure 5: Transition Matrix in Journeys with Banner
For banner ads (Figure 5), the first striking difference with the control journey is the higher probability of moving from a focus brand search to their website (p = .84). The probabilities of leading to an accommodation website are higher from a competitor’s website (p = .50), an accommodation app (p = .86) and a flight tickets website (p = .53). Furthermore, moving from flight tickets search to flight tickets website is also more likely (p = .47), as well as flight tickets app to information/comparison app (p = .27). Finally, the probabilities of leading to a competitor’s website are higher from an information/comparison website (p = .38) and an accommodation website (p = .47). The banner ad itself has a low probability of leading to the focus brand\s website (p = .03).
As for emails (Figure 6), the probabilities leading to the focus brand’s website from an information/comparison website (p = .16) or a competitor’s website (p = .12) are higher. Moreover, transitions to accommodation website from an accommodation app (p = .17), an information/comparison website (p = .21), a competitor’s website (p = .36), and generic search (p = .26), are lower. The direct transition from email to the website has a probability of p = .23.
Comparing journeys with pre-roll videos (Figure 7) to those without any FIC, transitions to an accommodation website are more likely from a competitors website (p = .53), information/comparison app (p = .27) and flight tickets app (p = .55); but less likely from an accommodation app (p = .23) or the focus brand’s website (p = .33). Going to a competitor website is less likely from a competitor app (p = .04), the focus brand website (p = .33). In and of themselves, pre-rolls have a very small probability of leading to the focus brand’s website (p = .01).
Figure 7: Transition Matrix in Journeys with Pre-Roll
For all the previous FICs, there doesn’t seem to be any difference in leading to conversion. This is most likely due to the low variance of the conversion variable (s2 < 0.01) with only 189 instances over 26,057 journeys. This is congruent with the t-tests conducted above.
In sum, we see noticeable differences between journeys with particular FICs and journeys without any, at many stages but not all of them and not at an average level. This partially confirms our hypothesis. The differences and their direction are summarized in Table C. No difference in conversion could be truly tested given the very low proportion of conversions (0.73%). We could therefore not truly test our second hypothesis, however we discuss an alternative explanation in the discussion part of our study. We validate the model in the following section.
Evaluation Criteria
Many criteria exist to determine whether a model is appropriate and reliable. In this report, we will use the criteria listed by Anderl, Becker, von Wangenheim & Schumann, (2016). They extend and update the original criteria proposed by Little (1970) which state that a model should be simple, evolutionary, adaptive, complete and robust.
Objectivity
A channel attribution model should attribute credits to all touchpoints in accordance to the value they individually generate. This leads to objective and unbiased results. In our model, all the touchpoints available were used and the Markov Chain model assigned credits to all touchpoints. Furthermore, the data was cleaned in a manner that ensured less biases would affect the model.
Predictive Accuracy
An attribution model should also be able to classify conversions (Shao & Li, 2011). As the distribution of conversion in the data is unbalanced, measuring classification performance with classic metrics such as loglikelihood would not be appropriate (He & Garcia, 2009; Provost, Fawcett & Kohavi, 1998). A widely accepted alternative, the ROC curve measures classification performance by accounting for the distribution of the observations in groups (Baesens et al. 2002; Fawcett 2006). The ROC curve plots the model’s true positive rate against its true negative rate at different discrimination thresholds (Fawcett, 2006). A discrimination threshold is the probability at which an observation is either classified as 0 or 1. For instance, if the threshold is 0.4, any probability over that is classified as 1. The area under the curve (AUC) is then calculated to summarize the accuracy of the model given different thresholds. Furthermore, the AUC can be compared to that of a naïve model (which is 0.5) or any other model.
out-of-sample AUC that Anderl, Becker, von Wangenheim & Schumann (2016) got on their first dataset using a first-order attribution model.
Figure 9: ROC Curve
Robustness
The results of the model must be reproducible and stable for it to be considered robust. To test for the robustness, we ran the same model (with all journeys) with 10-fold cross-validation and the results were almost identical (Appendix A). As this is a classification algorithm, obtaining consistent results is a positive sign as it shows that the results are not volatile or dependent on any other extraneous explanation.
Interpretability
The specification and estimation of the model should be within the reach of practitioners’ understanding. In our case, the construction of the model, the variables used and their interpretation is straightforward. The attribution model takes each journey and maps the probabilities of moving from one step of the journey (channel) to another. The visualization of the transition matrix and the possibility to plot a Markov graph are means to increase interpretability.
Versatility
Algorithmic Efficiency
Finally, the efficiency of a model is the speed and complexity involved. Running the models from the data preparation to the visualization took less than 220 lines of code and around 93.62 seconds. Given the size of the dataset we used, this is pretty efficient. Evidently, this will depend on both the data and the device specifications however we do not suspect that it would vary significantly on any modern device.
All the evaluation criteria attest to the objectivity, predictive accuracy, robustness, interpretability, versatility and the overall validity of the channel attribution model used to answer our research questions. We discuss these results hereafter.
DISCUSSION
Discussion
It is also important to explain the answers we found, what the managerial and research implications are and finally how our study could be improved. In this section, we first discuss the significant average differences between journeys, then the specific transition differences and later the implications and limitations.
Average Differences between Journeys
third sends a personalized message to a previous visitor of the website (Lambrecht & Tucker, 2013). This would then reduce the effort and time put in generic search. Finally, banners showed a lower likelihood for a customer to visit an accommodation website. Some literature suggested that display advertisement could be seen as unwanted (Blattberg, Kim & Neslin, 2008) or even intrusive (Goldfarb & Tucker, 2011) which may explain the previous results. In that case, the banner would be either ignored or blocked and the consumer would not consider finding an accommodation related to that product/service.
Specific Stage Differences Between Journeys
We noticed that journeys with affiliates have higher likelihoods of leading from an accommodation app, an information app or a flight tickets app to a competitor’s website. We cannot see how this can be a direct impact of affiliate marketing as the purpose of the latter is to directly link to the focus brand’s website (Malaga, 2007) which our results showed it does significantly better than a journey without any FIC. These effects could be coincidental correlations or the consequence of other factors not included in the data. They could be isolated and studied in future research to answer this question.
For journeys with banners, a focus brand search is more likely to lead to the focus brand’s website. This would make sense if banner ads are located in the journey either before a customer searches the name of the brand or after. In the first case, it could be being exposed to the banner ad and to want to know more by searching for the brand. In the second, it could be searching for the brand and seeing a banner ad. Banners were shown to lead to higher clickstreams (Urban, Liberali, MacDonald, Bordley & Hauser, 2013). The low likelihood of getting on the focus brand’s website by directly clicking the banner ad may add to the argument that this form of advertisement may not always be wanted (Blattberg, Kim, & Neslin, 2008).
Journeys with emails showed a higher probability of leading from a competitor’s website or an information/comparison website to the focus brand’s website. Previous studies already argued that email can increase website visit (Ansari & Mela, 2003). We could suggest that, in this case, when the email appears after a customer visited a competitor’s website or information website, it serves as a recall or provider of information and the next step in the customer journey is to return to the focus brand’s website.
of advertising. Belanche, Flavián & Pérez-Rueda (2017) suggested that a metric for measuring video ad effectiveness is the attitude toward that video. This can be translated as a feeling due to the exposure to the ad (Shimp, 1981). This feeling is an important mediator of brand consideration and choice (MacKenzie, Lutz & Belch, 1986; Shimp, 1981). This could explain the decrease in likelihood to visit a competitor’s website after visiting the focus brand’s own website and being exposed to pre-roll advertising. There is also a higher chance of visiting an accommodation website after visiting an info/comparison app or flight tickets app, which may indicate that consumers’ interest has been triggered and they are looking into the costs of particular trips. Pre-rolls, however, show little direct impact on the visit of the focus brand’s website, in spite of their presence in 571 journeys (second highest frequency of all FICs in the data). As these may be used on all kinds of video platforms at the beginning of a video a customer intends to play (Campbell, Mattison Thompson, Grimm & Robson, 2017), their low efficacy may indicate that they need to have a narrower targeting, that the location and time are simply not adequate, or that they fail to create a positive attitude as mentioned earlier. More studies are needed to increase our understanding of this marketing tool.
Retargeting showed higher likelihoods to move from the start of the journey, the focus brand’s search and a competitor’s website to the focus brand’s website. As presented earlier, the point of retargeting is to personalize advertising to customer who previously interacted with the brand which helps drive traffic to the website (Lambrecht & Tucker, 2013); which is confirmed by the model results. This may mean that customers who search for the focus brand, or are at the beginning of their journey and have visited the firm’s website in a preceding journey, may be more likely to revisit this website when retargeted. In a similar vein, these customers may have visited a competitor’s website before being exposed to the retargeting efforts. It also seems that journeys with retargeting have lower probabilities of leading from any of the websites mentioned to the accommodation website which could either be explained by consumers not finding the need to visit that type of website after visiting the focus brand’s website, a competitor’s website. This could confirm the idea that retargeting is used to redirect consumers, who are in an advanced stage of their purchase journey, to a website (Lambrecht & Tucker, 2013).
(e.g. Bucklin & Sismeiro, 2009; Ngwe et al. 2019). And it would make sense: one needs to visit the brand’s website if one needs to purchase their product/service, given it offers an e-commerce channel. We can go further and suggest that a consumer is closer to the conversion square when visiting the focus brand’s website, than an information/comparison website or app, a generic search or a competitor’s website, search and app. The first two serving mainly for information collection as described in our literature review (Brown et al., 2007). Similarly, assuming any brand purchase will be made the accommodation, competitor, flight tickets and focus brand’s respective channels might, by definition, be part of the final route toward conversion. Our results on the specific stages (Table C) indicate that affiliates, emails and retargeting can be effective in driving traffic from some CICs (information/search channel or competitor website) to the focus brand’s website. They can, thus, be considered as bringing consumers closer to conversion and can be focused on in future research. We can also learn from competitors and accommodation websites. The first benefits from affiliates, banners, emails and retargeting, and the second benefits from banners and pre-rolls. How these FICs perform compared to specific CICs could be the subject of further research.
Research and Practical Contributions
automated newsletter or reminder email could have appealing results after a customer visits an information/comparison website or even a competitor’s website. New insights on the positive consequences of pre-roll advertising on accommodation websites is offered. We would recommend conducting studies solely focused on those as an initial way to grow data resources available and start inferring more significant relationships within the customer journey. Similar tracking could isolate the customers who can be more likely brought to the website using retargeting, given their previous journeys, their behavior on competitors’ websites and their search habits. Finally, in the cases conversion is not observed, marketeers can pay closer attention to stages that are closer to conversion and on which we showed FICs perform differently.
Overall, the model could also attribute revenue and cost, and by extension serve as a resource for budgeting decisions. The model being versatile and interpretable, adding the data and understanding it, is straightforward for practitioners. However, some improvements to our study could still be made.
Limitations & Further Research
As such, the field of research on the customer journey offers increasing exciting possibilities. We are glad to have contributed to that understanding and are as excited to follow the advancements in the field as we were researching it.
ACKNOWLEDGMENT
APPENDIX
REFERENCES
Agarwal, A., Hosanagar, K., & Smith, M. D. 2011. Location, location, location: An analysis of profitability of position in online advertising markets. Journal of Marketing Research. 48(6): 1057–1073
Alba, J., Lynch, J., Weitz, B., Janiszewski, C., Lutz, R., Sawyer, A., & Wood, S. 1997. Interactive home shopping: Consumer, retailer, and manufacturer incentive to participate in electronic marketplaces. Journal of Marketing. 61(3): 38–53.
Anderl, E., Becker, I., von Wangenheim, F., & Schumann, J. H. 2016. Mapping the customer journey: Lessons learned from graph-based online attribution modeling. Journal of Research in
Marketing. 33 (3): 457-74.
Animesh, A., Viswanathan, S., & Agarwal, R., 2011. Competing “creatively” in sponsored search markets: The impact of rank, unique selling proposition, and competition on the effectiveness of sponsored search listings. Information System Research. 22(1): 153–169.
Archak, N., Mirrokni, V., S. & Muthukrishnan, S. 2010. Mining advertiser-specific user behavior using adfactors. Proceedings of the 19th International Conference on World Wide Web. WWW 2010. Michael Rappa, Paul Jones, Juliana Freire.
Baesens, B., Viaene, A., van Den Poel, D., Vanthienen, J., & Dedene, G. 2002. Bayesian Neural Network Learning for Repeat Purchase Modelling in Direct Marketing. European Journal of
Operational Research. 138(1): 191-211.
Belanche, D., Flavián, C., & Pérez-Rueda, A. 2017. Understanding interactive online advertising: congruence and product involvement in highly and lowly arousing, skippable video ads. Journal
of Interactive Marketing. 37: 75-88.
Berchtold, A., & Raftery, A. E. 2002. The Mixture Transition Distribution Model for High-Order Markov Chains and Non-Gaussian Time Series. Statistical Science. 17(3): 328-356.
Blattberg, R. C., Kim, B.-D., & Neslin, S. A. 2008. Database marketing: analyzing and managing
customers. Heidelberg, Germany: Springer.
Blake, T., Nosko, C., Tadelis, S., & National Bureau of Economic Research. 2014. Consumer
heterogeneity and paid search effectiveness: A large scale field experiment. Cambridge, Mass.:
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.
Braun, M., & Moe, W. W. 2013. Online display advertising: Modeling the effects of multiple creatives and individual impression histories. Marketing Science. 32(5): 753–767.
Breuer, R., Brettel, M., & Engelen, A. 2011. Incorporating long-term effects in determining the effectiveness of different types of online advertising. Marketing Letters. 22: 327-340.
Bronnenberg, B. J. 1998. Advertising frequency decisions in a discrete Markov process under a budget constraint. Journal of Marketing Research. 35(3): 399-406.
Brown, G.J., Pitts, M.G., & Wetherbe, J.C. 2007. Cognitive stopping rules for terminating information search in online tasks. MIS Quarterly. 31(1): 89-104.
Bucklin, R. E., & Sismeiro, S. 2009. Click here for internet insight. Advances in clickstream data analysis in marketing. Journal of Interactive Marketing. 23(1): 35–48.
Cambra-Fierro, J., Melero-Polo, I., Sese, F. J., & van Doorn, J. 2018. Customer-Firm Interactions and the Path to Profitability: A Chain-of-Effects Model. Journal of Service Research, 21(2): 201-218. Campbell, C., Cohen, J., & Ma, J. 2014. Speaker's box: advertisements just aren't advertisements anymore: a new typology for evolving forms of online "advertising". Journal of Advertising
Research. 54(1): 7.
Campbell, C., Mattison Thompson, F., Grimm, P. E., & Robson, K. 2017. Understanding Why Consumers Don’t Skip Pre-Roll Video Ads. Journal of Advertising. 46(3): 411- 423.
Chatterjee, P., Hoffman, D. L., & Novak, T. P. 2003. Modeling the clickstream. implications for web-based advertising efforts. Marketing Science. 22(4): 520-541.
Che, H., & Seetharaman, P. B. 2009. “Speed of replacement” modeling brand loyalty using last-move data. Journal of Marketing Research. 46(4): 494-505.
Dias, M. B., Locher, L., Li, M., El-Deredy, W., & Lisboa P. J. G. 2008. The value of personalised recommender systems to e-business: a case study. In Proceedings of the 2008 ACM Conference
on Recommender Systems. New York: Association for Computing Machinery. 291–94.
Ellis-Chadwick, F., & Doherty, N. F. 2012. Web advertising: The role of e-mail marketing. Journal of
Business Research. 65(6): 843-848.
Fruchter, G. E., & Dou, W. 2005. Optimal budget allocation over time for keyword ads in Web portal.
Journal of Optimization Theory and Applications. 124(1): 157–174.
Ghose, A., & Todri-Adamopoulos, V. 2016. Toward a digital attribution model: Measuring the impact of display advertising on online consumer behavior. Mis Quarterly. 40(4): 889-910.
Ghose, A., & Yang, S. 2009. An empirical analysis of search engine advertising: Sponsored search in electronic markets. Management Science. 55(10): 1605-1622.
Goldfarb, A., & Tucker, C. 2011. Online Display Advertising: Targeting and Obtrusiveness. Marketing
Science. 30(3): 389-404.
Gopal, R., Li, X., & Sankaranarayanan, R. 2011. Online keyword based advertising: Impact of ad impressions on own-channel and cross-channel click-through rates. Decision Support Systems, 52(1): 1–8.
Gregori, N., Daniele, R., & Altinay, L. 2014. Affiliate marketing in tourism: determinants of consumer trust. Journal of Travel Research. 53(2): 196-210.
de Haan, E., Wiesel, T., & Pauwels, K. 2016. The effectiveness of different forms of online advertising for purchase conversion in a multiple-channel attribution framework/ International Journal of
Research in Marketing. 33(3): 491-507.
Hoban, P. R., & Bucklin, R. E. 2015. Effects of internet display advertising in the purchase funnel: model-based insights from a randomized field experiment. Journal of Marketing Research. 52(3): 375-393.
Homburg, C., Steiner, V. V., & Totzek, D. 2009. Managing dynamics in a customer portfolio. Journal
of Marketing. 73(5): 70-89.
Howard, J. A., & Sheth, J. 1969. The Theory of Buyer Behavior. New York: John Wiley & Sons. Jansen, B. J., & Schuster, S. 2011. Bidding on the buying funnel for sponsored search and keyword
advertising. Journal of Electronic Commerce Research. 12(1): 1.
Kitts, B., & LeBlanc, B. 2004. Optimal bidding on keyword auctions. Electronic Markets. 14(3): 186– 201.
Kohavi, R. 1995. A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection. Proceedings of the Fourteenth International Joint Conference on Artificial
Kumar, N., Lang, K. R., & Peng, Q. 2005. Consumer Search Behavior in Online Shopping Environments. Proceedings of the 38th Annual Hawaii International Conference on System
Sciences: Big Island, HI, USA. 175b.
Lambrecht, A., & Tucker, C. 2013. When does retargeting work? Information specificity in online advertising. Journal of Marketing Research, 50(5): 561-576.
Lavidge, R. J., & Steiner, G. A. 1961. A Model for Pre- dictive Measurements of Advertising Effectiveness. Journal of Marketing, 25(November): 59–62.
Leeflang, P. S. H., Wieringa, J. E., Bijmolt, T. H. A., & Pauwels, K. H. 2015. Modeling Markets:
Analyzing Marketing Phenomena and Improving Marketing Decision Making. NY: Springen
New York.
Linden, G., Smith, B., & York, J. 2003. Amazon.com recommendations: item-to-item collaborative filtering. IEEE Internet Computing. 7 (1): 76-80.
Little, J. D. C. 1970. Models and managers. the concept of a decision calculus. Management Science. 16(8): 466-485.
MacKenzie, S. B., Lutz, R. J., & Belch, G. E. 1986. The Role of Attitude Toward the Ad as a Mediator of Advertising Effectiveness: A Test of Competing Explanations. Journal of Marketing
Research. 23(2): 130-43.
Malaga, R. A. 2007. The new marketing intermediaries: a multiple case study of three new e-business models. Journal of Academy of Business and Economics. 7: 1.
Manchanda, P., Dubé, J., Goh, K., Chintagunta, P. 2006. The effect of banner advertising on Internet purchasing. Journal of Marketing Research. 43(1): 98–108.
Martin, A. 2009. The long road to conversion: the digital purchase funnel. White Paper. Microsoft’s
Atlas Institute.
Montgomery, A. L., Li, S., Srinivasan, K., & Liechty, J. C. 2004. Predicting online purchase conversion using web path analysis. Marketing Science. 23(4): 579-595.
Mulpuru, S., Van Boskirk, S., Freeman Evans, P., & Roberge, D. 2011. The purchase path of online buyers: a joint holiday shopping study with gsi commerce of 15 retail websites. Research Report,
Forrester Research.
conference on Internet and network economics (WINE). Verlag, Berlin, Heidelberg: Springer.
131–142.
Neslin, S. A., Grewal, D., Leghorn, R., Shankar, V., Teerling, M. L., Thomas, J. S., et al. 2006. Challenges and Opportunities in Multichannel Customer Management. Journal of Service
Research. 9(2): 95–112.
Neslin, S. A., & Shankar, V. 2009. Key issues in multichannel customer management: current knowledge and future directions. Journal of Interactive Marketing. 23(1): 70-81.
Pfeifer, P. E., & Carraway, R. L. 2000. Modeling customer relationships as Markov chains. Journal of
Interactive Marketing. 14(2): 43-55.
Raman, K., Mantrala, M. K., Sridhar, S., & Tang, Y. 2012. Optimal resource allocation with time-varying marketing effectiveness, margins and costs. Journal of Interactive Marketing. 26(1): 43– 52.
Rusmevichientong, P., & Williamson, D. P. 2006. An adaptive algorithm for selecting profitable keywords for search-based advertising services. Proc. 7th ACM Conf. Electronic Commerce, ACM, New York. 260–269.
Rutz, O. J., Trusov, M., & Bucklin, R. E. 2011. Modeling indirect effects of paid search advertising. which keywords lead to more future visits? Marketing Science. 30(4): 646-665.
Shankar, V., & Malthouse, E. C. 2007. The growth of interactions and dialogs in interactive marketing.
Journal of Interactive Marketing. 21(2): 2–4.
Shao, X., & Li, L. 2011. Data-Driven Multi-Touch Attribution Models. Proceedings of the 17th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining: ACM. 258-264.
Shimp, T. A. 1981. Attitude Toward the Ad as a Mediator of Consumer Brand Choice. Journal of
Advertising. 10(2): 9-48.
Skiera, B., & Nabout, N. A. 2013. PROSAD: A bidding decision support system for profit optimizing search engine advertising. Marketing Science, 32(2), 213–220.
Verhoef, P. C. 2012. Multichannel Customer Management Strategy. In Shankar, V. & Carpenter, G. S. (Eds): Handbook of Marketing Strategy, p. 135-150. Cheltenham, UK: Edward Elgar Publishing. Wu, J., Li, K. J., Liu, J. S. 2015. Bayesian inference for assessing effects of email marketing campaigns.
Journal of Business and Economic Statistics. 36(2): 253.
