ABSTRACT
Digital advertising campaigns often are launched across multiple channels, a selection of which may include search, display advertisements, social media, mobile, video, and e-mail. By exposing consumers to advertisement impressions, these channels help consumers make purchase decisions or sign up to an advertised service (Fulgoni, 2016; Yang and Ghose, 2010). To gauge the effectiveness of such advertising campaigns, one must know which media channels or advertising formats have contributed to a purchase conversion. This process is known as attribution. A better understanding of attribution, or assigning conversion credit to the various relevant channels, can serve a number of research and industry purposes. Marketing managers may use such attribution models to interpret the influence of advertisements on consumer behavior and to optimize their advertising campaigns.
MANAGEMENT SLANT
At each stage in a consumer's journey toward purchase, different online channels feature most prominently.
Existing credit-assignment methods, such as the last click, suffer from the problem of attribution—they do not take into account the impact of all those advertising formats that were visited by a consumer contemplating a purchase.
Four rule-based models can be used for measuring the performance of an advertising campaign—the last-click, time-decay, uniformly distributed, and position-based models.
Multichannel attribution models have evolved to reflect the growing complexity of attributing credit with each new advertisement format.
INTRODUCTION
Early academic research into attribution modeling mainly focused on click-through rates as an effective measure of performance (Ansari and Mela, 2003). Digital advertising soon faced a conundrum, however; several studies conducted during the early 2000s reported that click-through rates in fact were declining fast (Manchanda, Dube, and Goh, 2006). Researchers discovered that only a small proportion of visits translated into final purchase (Chatterjee, Hoffman, and Novak, 2003).
This finding inevitably cast doubt over digital advertising as an effective mode of reaching consumers and called into question whether it can substitute for offline advertising, such as television and newspapers. Researchers found that banner advertising was not an effective online media strategy (Goldfarb and Tucker, 2011). Against this background, some practitioners began advocating the use of traditional measures of advertising effectiveness, such as awareness and recall, in relation to the performance effects of digital-advertisement performance (Fulgoni, 2016; Manchanda et al., 2006). Academics also investigated the impact of display advertisements on long-term brand awareness and similar performance measures (Goldfarb and Tucker, 2011).
A related stream of research measured changes in brand attitudes, brand awareness, and purchase intentions as a function of advertising exposure (Fulgoni, 2016; Manchanda et al., 2006). Because these brand-choice models excluded purchase incidences from their analyses, however, they could not be used to evaluate the impact of display advertisements in the consumer journey. A potentially more interesting line of research since has examined the workings and institutional features of different advertising formats, such as display. One study investigated the potential of targeted digital advertising by considering how new media channels might create entirely new consumer markets (Athey and Gans, 2010).
Recent research using these and other, similar, measures of performance suggests that banner advertisements indeed can be an effective form of advertising (Blake, Nosko, and Tadelis, 2015; Kireyev, Pauwels, and Gupta, 2016). In a study of the relationship between banner advertisements and consumer purchase patterns, researchers found that banner advertisements can play a significant role in customer retention (Manchanda et al., 2006). By estimating a purchase-incidence advertising-response model with individual-level response parameters, the study showed that exposure to banner advertisements increased the purchase probabilities for current customers.
Because attribution modeling only has begun to receive increased attention in academia and practice alike, concerns remain over the development of an appropriate performance-measurement system for digital advertising. The last-click method of attribution is flawed, because it fails to take account of the influence of all touch points except the last one (Moe and Fader, 2004) and so does not capture the full value of digital advertising (eMarketer, 2015). Companies adopting the last-click model forfeit the chance to better optimize their advertising spend (Moe and Fader, 2004).
With increasing recognition of the role of digital advertising as an effective strategy, and in response to the above criticisms, the advertising industry proposed the alternative concept of multichannel attribution. This framework assumes that more than one channel impression or touch point can have a fraction of the credit for a sale, on the basis of the true influence each impression has on the conversion (Lovett, 2009). The underlying assumption is that individual advertising channels should not be evaluated in isolation, and credit must be assigned equitably with respect to the campaign goals on these channels. Industry analysts also claim that the multichannel attribution strategy calculates individual channel cost per acquisition figures that are much closer to reality (Lovett, 2009), providing a better understanding of sales-cycle length and the purchase funnel. For instance, one study recommended using the last-click model if “ads and campaigns are designed to attract people at the moment of purchase” (Google Analytics, 2012).
In the current article, using an online company's purchase-conversion data, the authors examined the nature and scope of these rule-based attribution models in measuring the performance of online channels in customer journeys. The study employed a statistics-based attribution model for online businesses, to shed light on how multichannel attribution models can be used to better measure advertising performance. The authors developed hypotheses that examine
at what stage in a consumer's journey different online channels feature most prominently for an online business;
the financial importance of these channels under last-click models;
the effects of moving to rule-based multichannel attribution models—time decay, uniformly distributed, and position based—and statistics-based multiattribute models.
The study investigated
whether multichannel attribution models give different channel valuations than last-click models;
whether these channel valuations vary significantly among the various multichannel models;
whether statistical multiattribute models have predictive validity.
The authors considered the convergent validity of multichannel models as well as the forecasting ability of the statistical model, as measured by prediction to a holdout sample. To date, the effect of changing attribution models for different online channels remains largely unstudied. An analysis of these models, therefore, will offer conclusions on whether an advertising format's revenue significantly differs depending on the model used.
Extant literature provides generalizable insights on individual channel effectiveness (Kireyev et al., 2016). By contrast, the current researchers used data to compare different models' abilities to predict to the holdout sample and then compared directly with the best model identified. Using these insights, the current research also shows how different online sales channels should be credited for the conversion event, and to what extent.
The article is organized as follows. The authors first examine various existing and proposed attribution models in relation to online advertising. They then assess at what stage in a consumer's journey a particular media channel features most prominently for an online business. Empirical results on the outcomes of different attribution models then are presented in the next section. The study concludes by considering implications for different online sale channels and attribution.
ATTRIBUTION IN DIGITAL ADVERTISING: A LITERATURE SURVEY
Multiple vendors, publishers, or search engines (herein called channels) serve advertisements, so there can be a number of touch points or channel impressions with digital advertising in a purchase funnel. If there is only one channel impression, such as an iTunes movie purchase, attribution models unlikely will vary significantly in their predictions. Attribution modeling becomes interesting only when one considers the impact of several channel impressions together.
In the case of a cosmetic product, for example, customers likely read blogs, click display advertisements, or search branded products before making a purchase decision. Their online journey may take them to various channels, where they make contact with advertising media and formats. Advertising campaigns may be run on a wide variety of online marketing channels, such as social media and search, each of which may be important in customer journeys.
To deal with these complex interactions, researchers use attribution models that can reduce some of these complexities to their basic components by focusing on each stage of a customer's journey that results in a purchase. Because different advertising formats may differ in their total revenue effect, their influence may vary depending on the particular stage of the purchase funnel. The process of attribution thus can be understood as the assignment of conversion credit when multiple advertising channels reach a given online user. One then may apply multichannel attribution models that aim to find the optimal mix of digital-advertising channels that provide the highest return on investment. By tracking all the clicks in the stream, these models distribute the transaction's sale value across all channel impressions in accordance with their added value.
The problem of consumer interactions with different advertising forms can be visualized (See Figure 1). Suppose a web surfer clicks a display advertisement and makes a purchase (Case 1). The question of whether the display advertisement affects the consumer's purchase decision is trivial and can be measured with a simple look at purchase conversions.
In the second scenario (Case 2), a consumer visits display advertisement, paid search, social media, and price-comparison sites before taking a buying decision. Do visits to all these channels have equal effects on a conversion? Or do they vary depending on various contextual factors?
If, indeed, it is the case that channel visits have differential effects, then how can one assign credit to each of those consumer interactions? If the credit is all given to display, there are negative downstream effects—money may be reallocated from other productive channels—due to the lack of the assist from the other channels, including paid search, social media, and price-comparison sites. Conversely, the less-valuable channels could attract more credit than they deserve if one applies an arbitrary rule to assign credit for conversion. This also might lead to inaccurate advertising-spend calculations by the marketing departments when they allocate budget across multiple forms of advertising.
The Role of Online Sales Channels In the Customer Journey
A customer can take different choice actions at different stages of a purchase funnel. There can be three roles in a customer journey—introduction, assist, and conversion (Chandler-Pepelnjak, 2010; Internet Advertising Bureau, 2011). Introduction plays the initial role in starting the process, often populated by search. Consumer behavior tends to be exploratory at this stage.
The second role is that of assist, which is any contributing website that enables a transaction after the introduction and before the conversion. At this stage, consumer behavior turns to a goal-directed search (Kireyev et al., 2016). The conversion is the final step before the purchase is made. The presence of more than one marketing channel raises the question of which of these channels features prominently at different stages of the purchase funnel (Chandler-Pepelnjak, 2010). Because digital advertisements are automated, these channels can monitor which links are catching readers' attention.
As noted, many existing credit-assignment methods, such as the last-click method, suffer from the fundamental problem of attribution. They do not take into account the impact of all those advertising formats that were visited by a consumer contemplating a purchase (Wiesel, Pauwels, and Arts, 2011). When a consumer enters the generic keyword “notebook” in a search engine, the search will return paid advertisements by electronic retailers with links to branded notebooks as well as organic results on the search results page. Now assume that the searcher already had seen a link in a displayed advertisement format, decides to click on it, and converts (as previously discussed in Case 2). It is clear that both the displayed advertisement and the paid advertisement were on the path that led to the actual purchase—the consumer moved down the conversion funnel from displayed advertisement to paid advertisement.
A rapidly growing body of literature thus examines the entire click-stream history of individual consumers in terms of whether visits to different advertisement formats have positive effects that accumulate toward a purchase, such as learning about a product that the shopper intends to buy (Moe and Fader, 2004). This strategy of modeling the purchases resulting from the accumulative effects of all previous interactions largely focuses on how non-purchase activities—advertisement clicks, website visits—affect the probability of purchasing. These models consequently cannot deal directly with the question of attributing credit for conversion to each individual advertising form.
It also is difficult to observe a cross-channel effect in an interactive situation. Paid search and organic search may complement each other in the purchase funnel (Yang and Ghose, 2010). There thus is scope for additional research to figure out whether such complementarity effects exist in other online mediums. A study measuring cross-effects of online and offline media in terms of how customers move through the purchase funnel found a very high sales elasticity for AdWords (4.35) and much lower sales elasticities (0.05 and 0.04) for flyers and faxes (i.e., other sales channels), respectively (Wiesel et al., 2011). Another study used a probit-based consideration and nested logit formulation for visit and purchase to attribute conversions (Li and Kannan, 2014).
This and other related models for mutually exciting point process models generally have emphasized the classification accuracy of different channel contributions, which means that they often omit the stability characteristic of the variable contribution estimate (Xu, Duan, and Whinston, 2014). These models do not study how higher exposure in one channel, such as display, may lead to a higher level of activity in another channel, such as search clicks (Anderl, Becker, Von Wangenheim, and Schumann, 2016; Kireyev et al., 2016). There also is a need to investigate whether online advertising effects differ by stage in the purchase funnel, given that earlier models did not treat online advertising effectiveness from this perspective (Li and Kannan, 2014).
ATTRIBUTION MODELS
The following section first describes the four rule-based models that can be used for measuring the performance of an advertising campaign—the last-click, time-decay, uniformly distributed, and position-based models. As can be seen, the last three models are multichannel attribution models. All these models are populated with both advertisers and search engines, given that many users' home pages are search engines (Google Analytics, 2012). Companies may use campaign data to assign the percentage of credit on the basis of one or the other model. The authors then evaluate the assumptions of a statistics-based attribution model in the context of an advertiser's goal of achieving the optimal conversions.
The Last-Click Model
As digital advertising grew in the 1990s, the last-click model emerged as the industry's main advertising-measurement tool. It ascribes 100 percent of the credit to the last advertisement the user clicked on before a purchase conversion (Xu et al., 2014). With regard to Case 2, display, paid search, and social media would get nothing at all under this model; instead, the model would attribute the entire conversion to price comparison.
The strength of the last-click model lies in its ability to help determine which channels best lead users to a buying decision or final conversion. It thus provides a simple and elegant way to determine the credit assignment of each positive user and quickly became the fastest way to infuse confidence into the efficacy of digital-advertising campaigns. A notable drawback of the last-click model is that it does not take into account many vital interactions and steps in a customer's journey to transaction.
From a complete-journey perspective, it can be seen that the last-click attribution model unfairly rewards the sites at the end of a customer's journey. This has prompted both marketing practitioners (Clearsaleing, 2014; Lovett, 2009) and researchers (Wiesel et al., 2011) to consider alternative approaches, such as multicampaign methods, that attribute credit to all online marketing exposures that resulted in a conversion. The multichannel attribution models assign credit to multiple channels when a number have been observed showing an advertisement to a converting user (Moe and Fader, 2004). Credit thus is assigned to more than just one advertisement impression for driving the user to take a desirable action, such as making a purchase. The idea is to allow an advertiser to share credit among the websites that influenced the transaction at any stage. The multichannel models and their underlying assumptions are described in the following sections (See Table 1).
The Time-Decay Model
As a multicampaign attribution framework, the time-decay model adjusts credit so that the closer an impression is to a conversion, the more credit it receives. As the credit progressively increases in value across the customer's journey, the model nicely distinguishes itself from the last-click model (Clearsaleing, 2014). Under the time-decay model, the rewards can be apportioned so that the last contributor achieves maximum credit.
For channels in Case 2, the time-decay model calculates value by progressively attributing more credit to the impressions closer to conversion—that is, price comparison. As the credit increases with time from initial discovery to final conversion, the model captures an important aspect of the consumers' online behavior: that conversions are associated with a short attention span. Similarly, the importance of the impressions may decrease with time. In these cases, the closer an impression is to a conversion, the less credit it receives.
This rule-based attribution model follows the triangular numbers ratio of 1:3:6:10. In order to calculate the ratio breakdown for different-length journeys, therefore, the authors used the following formula: (1) In the above formula, the use of n represents what step number it is; “Tn” is the weighting given to it, and the 100 percent commission is divided up depending on the ratio. A three-step journey, for example, will be divided up under the ratio of 1:3:6, with 10 percent, 30 percent, and 60 percent attributed to the steps in chronological order. For each marketing tool, the model then totals the revenue generated to show the revenue generated by that tool for the time-decay model. It can be envisaged that those marketing channels that close the deal as quickly as possible benefit more from this attribution strategy, because the value is weighted progressively higher for channels nearest to the last impression. Researchers have argued, therefore, that the time-decay model is particularly useful for short-lived deals or promotional offers (Lovett, 2009).
Uniformly Distributed Attribution Model
In a uniformly distributed multi-impression attribution model, the value of each conversion is distributed uniformly to all impressions. In Case 2, the uniformly distributed model would attribute 25 percent of the conversion to each of the four channels involved. Because the shares of each conversion are divided equally among all channels, the model does not consider where the touch points occur.
The model's assumption that each interaction in the customer journey has equal influence on the user's purchase decision thus is not valid, given that there may be varying influences of all such interactions. The model is motivated by the concern that the intrinsic value of each impression cannot be credited easily, even if the influence of each impression is significant enough in a purchase conversion. Therein lies the advantage of an equally weighted attribution model: it simplifies the attribution process by assuming that each impression contributes equal value to the conversion.
The Position-Based Model
A popular version of the position-based model uses a Pareto distribution rule, which attributes value to specific parts of the customer journey. The Pareto distribution model places more importance on the first and last touch points than on all the others in between. For example, under the 80/20 rule, 80 percent of the conversion value is credited to the first and last touch points, and the remaining 20 percent is distributed to the other channels in the customer journey. In relation to Case 2, display and price comparison would be assigned 80 percent of the value of conversion, and paid search and e-mail would get only 20 percent.
The model assumes that the first impression is important because it attracts the user's attention, and the last one is important because of the role it plays in concluding the transaction. The remaining impressions can be evaluated equally low in their impact. A major strength of the model is that each channel's importance is determined individually in relation to the specific goals of an advertising campaign.
A higher value might be assigned to the first interaction because, in some situations, numerous leads might need to be generated, and the campaigns are more focused on creating awareness. Because the middle contributors receive value as well as the first and last clicks, the method offers an adequate response to the criticism that rule-based models remove brand value (Lovett, 2009). Despite emphasizing the role of all agents in value creation, however, the model still is biased heavily toward the first- and last-click channels.
Statistics-Based Attribution Model
Many existing models are concerned mostly with calculating channel effectiveness in multichannel settings, whereas attribution models ideally should be able to predict conversion events correctly (Shao and Li, 2011). In this section, the authors describe a statistics-based multiattribute approach that is based on empirical observations rather than theoretical assumptions. In light of the increasing complexity of digital advertising, in recent years, researchers have endeavored to develop a true data-driven methodology to account for the influence of each user interaction on the final user decision.
A probabilistic model was developed on the basis of a combination of first- and second-order conditional probabilities (Shao and Li, 2011). There are two steps involved in generating this model. The empirical probability of the main factors (i.e., the probable use of different media channels) is calculated as follows: (2) and the pairwise conditional probabilities (3) for i ≠ j.
A conversion event (purchase or sign-up) is denoted as y, which is a binary outcome variable, and xi,i = 1,…, p denotes p different advertising channels. Npositive (xi) and Nnegative (xi) denote the number of positive or negative users exposed to channel i, respectively, and Npositive (xi, xj) and Nnegative (xi, xj) denote the number of positive or negative users exposed to both channels i and j. Customer journeys contain one or more touch points across a variety of channels. The channels actually visited by the consumers out of the many channels involved give information on the number of positive and negative users in a purchase funnel.
The contribution of channel i then is computed at each positive user level as (4) where Nj≠i denotes the total number of js not equal to i. In this case, it equals N − 1, or the total number of channels minus 1 (the channel i itself) for a particular user.
Because there is significant overlap among the influences of different touch points as a result of the user's exposure to multiple media channels, the model fully estimates the empirical probability with the second-order interactions. An advantage of using the above estimation is that it includes the second-order interaction terms in the model. After rescaling, under certain simplifying assumptions, this probability model is equivalent to a Shapley value formulation, a solution concept in cooperative game theory named for Lloyd Shapley, who introduced it in 1953 (Dalessandro, Stitelman, Perlich, and Provost, 2012).
In a typical Shapley value cooperative game, a group of players generates a shared value, such as wealth or cost, for a group as a whole (Osborne and Rubinstein, 1994). The Shapley value of a player in a game is calculated as his or her expected marginal contribution over the set of all permutations on the set of players. In other words, the Shapley value of an advertising medium is its expected marginal contribution over all possible sets of the interacting channels.
RESEARCH OBJECTIVES
An advertising campaign may be designed in such a way that it induces a customer to visit different online channels until he or she finally purchases a product or service. Prior literature, as discussed, indicates the differential effectiveness of different advertising forms, but the effect of changing attribution models for these various forms remains mostly unstudied. This largely is because many such enterprises use the last-click attribution model (Clearsaleing, 2014). The focal company of the current study also uses the last-click model as a default attribution strategy, which means that this study can focus on assigning value to a particular channel and then compare the effects of moving to time-decay, uniformly distributed, position-based, and statistics-based attribution models against the current last-click model. The study postulates three specific hypotheses:
H1: Multichannel attribution models give different channel valuations than last-click models.
H2: These channel valuations vary significantly among the different multichannel models.
H3: Statistical multiattribute models have predictive validity.
Among other things, the authors' findings shed light on whether last-click attribution models should be discarded in favor of multichannel attribution models. As part of these investigations, the study considers the convergent validity of multichannel models and discusses the forecasting ability of the statistical model as measured by prediction to a holdout sample.
To implement the research, the authors considered whether an advertising format such as display generates more revenue under the last-click model than it would under a multichannel attribution model, such as the time-decay model—or any other multichannel attribution models, for that matter. This was motivated by the observation that display may act as a converter in the purchase funnel; because the last-click model attributes 100 percent credit to a converter, all credit goes to this online channel. In a study of digital-channel effectiveness, one research group found that display advertisements increased search conversion (Kireyev et al., 2016).
To examine the differences between display and the other online marketing channels, the authors assessed whether both groups' means statistically differed, with reference to the average order value. They also looked for any significant difference between the revenues attributed to the online channels. In the data, the authors included information about several key aspects of digital-marketing tools. For example, display or banner advertisements are used for both direct response and branding styles of marketing. The retargeting company maximizes the purchases per dollar spent on advertisement placements by choosing the most cost-effective placements. E-mail campaigns and retargeting can be used effectively, depending on the specific goals of an advertising campaign and the performance metrics used to gauge the campaign's effectiveness.
METHODOLOGY
In this section, the authors utilized logs from a large-scale online sales platform to identify where different online channels featured in customer journeys. In total, 996,708 transactions were included in the analysis, with a total revenue of USD158,519,417, at an average order value of USD112.5. In terms of the customer journey lengths, 65.95 percent were one step, 14.58 percent were two steps, 8.78 percent were three steps, 3.86 percent were four steps, and 6.84 percent were five steps or more. The conversion data spanned 104 weeks, from January 1, 2012, to February 28, 2014.
Currently, the investigated company attributes revenue generated through online transactions to its various paid marketing tools on a last-click basis. Because the authors tested the multichannel attribution models that look at touch-point sequences, the study's data contain the full set of touch points; thus, it has information about complete consumer journeys.
Data on the following eight most common online marketing channels were collected (See the Appendix for examples of this data-collection process):
Display-advertisement transactions consisted of any visitor who originated (i.e., visitors whose prior click stream included that channel) from display advertisements posted on any third-party nonsearch web domain.
Organic search transactions consisted of any visitor who originated from an organic (nonpaid) search on a web search engine.
Paid search transactions consisted of any visitor who originated from a pay-per-click advertisement on a web search engine.
Price-comparison transactions consisted of any visitor who originated from a price-comparison site.
E-mail transactions consisted of any visitor who originated from e-mail.
Retargeting transactions consisted of any visitor who originated from retargeting; retargeting is never the only step in a sales journey.
Social medial transactions consisted of any visitor who originated from a social media website.
“Others” included transactions that consisted of any visitor who originated from a manual URL entry into a web browser.
It is important to acknowledge that, depending on the product category, offline channels, such as word-of-mouth or bricks-and-mortar store visits, also are
The authors first parameterized their models and then fitted them to two-thirds of the data and tested a holdout sample. They subsequently chose the model that fit and predicted best. They defined their conversion measures as follows:
Purchase conversions: Number of final sales transactions generated by the online sales-channel advertisements.
Purchase conversion rate: Percentage of purchase conversions on advertisement impressions out of the total number of times that advertisements were clicked.
RESULTS
Descriptive Statistics
The researchers measured the contributions of different online marketing tools to online revenue under the last-click model (See Table 2). As discussed, the model attributes all conversions to the last referring impression within a customer journey, which means it is the final interaction that matters from a marketing perspective. It can be seen that with the current last-click method, the highest revenue-generating online marketing tool was that of organic search, which brought 63 percent. Social media contributed the least with the current model, at 1 percent. In other words, organic search was the biggest contributor to company revenue, more than display and other media channels.
The data also shed light on social media's relatively small contribution to the company revenue; however, this was expected from the study's data, because social media is still an emerging new media channel in many sectors of the economy. Another important finding is that the mean order value for display was higher than any other of the marketing tools, at $159. In addition, the data show that display featured most prominently as the converter, given that 39.08 percent of display advertisements in the sample acted as a converter (Chandler-Pepelnjak, 2010), being the last step in a multistep customer journey.
In addition, 22.59 percent of all converters in online marketing were display advertisements. Display featured least prominently when undertaking the role of introducer (11.30 percent). The fact that any customer journey longer than five steps was shortened to five might ignore some display advertisement introducers. This is an area that requires further exploration in future research.
Main Results
To examine whether multichannel attribution models give different channel valuations than last-click models, the authors conducted a two-sample t test comparing average order values of different online marketing tools under different multichannel attribution models. They first examined the time-decay model (as developed by Google Analytics, 2012); it credits most of the sites along a customer journey but places increasing emphasis on the steps closest to the transaction.
Time-Decay Model
The authors measured the effects on revenue for online media channels if the time-decay model was introduced (See Figure 2). As with the last-click model (See Table 2), organic search led the way, contributing 61.27 percent. Display advertisements represented 14.68 percent of the revenue generated, down on the 20.30 percent revenue accumulated under the last-click model. E-mail and social media saw the largest change in value percentage: increases of 197.67 percent and 435 percent, respectively.
The authors also measured the revenue allocated to display using all five attribution models. The results did not change if any other online media channel was considered instead of display. Data included were the mean of revenue that each method attributed to display, and the difference between the last-click method and the four other attribution methods' revenue. There was a considerable difference in the mean revenue of the five models, with time decay trailing last click to a large extent. A contributing factor could be that the last-click model allocates 100 percent of the revenue to the last media channel in customer journeys. In contrast, the time-decay model offers varied percentages.
The authors conducted a two-sample t test comparing last-click and time-decay average display rewards. The t value of 25.397 was greater than the two-tailed critical value of 1.960, therefore indicating, with a 95 percent confidence level, a significant difference between the average display advertising rewards. Accordingly, one could conclude that the time-decay model, on average, attributes lower revenue to display. This is supported by the data, which show that displayed advertisements were allocated 14.55 percent lower revenue under the time-decay model (See Table 3). These results suggest that the time-decay model as a multichannel attribution model gives different channel valuations than the last-click model.
Uniformly Distributed Model
The uniformly distributed model follows the last-click and time-decay models, collating the revenue for each online media channel to sum their overall contribution to online marketing in the weighted-average format. Organic search continued to dominate the revenue streams, followed by display and paid search, when the authors looked at the effects on the attribution of revenue for online media channels using the uniformly distributed model (See Figure 3). A decrease in display revenue from the last-click model and the time-decay model was evident. Display contributed 14.12 percent of the revenue, a decrease of 23.34 percent from the last-click model. Price comparison, e-mail, and social media continued to see increases in revenue percentages from the last-click model, at 53.95 percent, 154.65 percent, and 415 percent, respectively.
The authors conducted a two-sample t test comparing last-click and uniformly distributed average display rewards. The t value of 30.804 was greater than the two-tailed critical value of 1.960, therefore indicating, with a 95 percent confidence level, a significant difference between the average display rewards. Consequently, one could conclude that the uniformly distributed model, on average, attributes lower revenue to display. This is supported by the data, which show that display advertisements were allocated 25.46 percent lower revenue under the uniformly distributed model (See Table 3). Employing linear regression, the authors further carried out hypothesis testing of the difference between means using the t test. As previously, these findings confirm that the uniformly distributed model as a multichannel attribution model gave different channel valuations than the last-click model.
Position-Based Model
The position-based (sometimes called Pareto distribution) model, a relatively new attribution approach, attributes credit for purchase conversion to specific parts of the customer journey. When the authors looked at revenue attribution for online media channels using the position-based model, organic search continued to dominate the online marketing tools, with 63.47 percent (See Figure 4). Display dominated among the paid marketing tools, being responsible for 13.65 percent, which represents a 25.89 percent decrease on the last-click model.
As seen in the other models, social media continued to contribute the least revenue; however, it still reflected a 295 percent increase on the revenue generated in the current last-click model. Retargeting was fairly consistent with the last-click model, slightly increasing, from 1.22 percent to 1.78 percent. Paid search saw a sizable increase in proportional revenue, from 10.92 percent to 13.88 percent.
The authors conducted a two-sample t test comparing last-click and position-based average affiliate rewards. The t value of 29.913 was greater than the two-tailed critical value of 1.960, therefore indicating, with a 95 percent confidence level, a significant difference between the average display rewards. The authors concluded that the position-based model, on average, attributed lower revenue to display. This is supported by the fact that display advertisements were allocated 23.73 percent lower revenue under the position-based model (See Table 3). The authors thus concluded that the position-based model as a multichannel attribution model gives different channel valuations than the last-click model.
Statistics-Based Model
When the authors looked at the effects on revenue attribution for online media channels using the statistics-based model, the results showed that display represented 14.34 percent of the revenue generated, down from the 18.42 percent revenue accumulated under the last-click model. Organic search, in contrast, registered a decrease, from 63.45 percent to 60.84 percent. Social media and e-mail recorded the largest changes in value percentage, as reflected in their revenue generation contributions of 3.67 percent and 3.38 percent, respectively. There was also a sizable increase in paid search, from 10.92 percent under the last-click model to 12.85 percent under the probability model.
The authors conducted a two-sample t test comparing last-click and probability-based display rewards. The t value of 28.435 was greater than the two-tailed critical value of 1.960, therefore indicating, with a 95 percent confidence level, a significant difference between the average display returns. The authors thus concluded that the statistics-based attribution model, on average, attributes lower revenue to display. This is also supported by the fact that display advertisements were allocated 24.86 percent lower revenue under the statistics-based attribution model (See Table 3).
Hypothesis Testing
Employing linear regression, the authors further carried out hypothesis testing of the difference between means using the t test. The R2 of .456 and adjusted R2 of .321 explained well the response variable (average order value; See Table 4). F = 808.264 in the regression analysis is equal to the square of the t value (28.435) from the t test, which is consistent with Property 1 of F distribution. All these results provide support to Hypothesis 1.
In order to examine Hypothesis 2, that channel valuations would differ significantly among the various multichannel models, the authors provide correlation data analysis (See Table 3). As can be seen from the data, there were higher correlations among the channel valuations of different multichannel attribution models. That is, the differences among multichannel models were not that great, except for the smaller channels. The results verify the convergent validity of the multichannel approach, but do not support Hypothesis 2.
In addition to the forecasting accuracy of models fitted to past data (See Table 4), the authors used a holdout sample to validate the study's results. As discussed above, they divided their sample into two subsamples: one for estimation of the model, and another for validation purposes. The procedure applied was to first parameterize the models and then fit them to two-thirds of the data and test the holdout sample. Consequently, the authors could compare different models' abilities to predict the holdout sample (See Table 5). The statistics-based attribution model provided a better fit to the data than the other models did, confirming the final hypothesis. This can be seen from the low mean absolute percentage error and mean absolute error for the statistics-based attribution model.
The logistic regression results for various online channels offered further insight (See Table 6). The regression predicts order or no order. As expected, all major online channels, including organic and search, showed greater contributions toward final conversions, providing further support to the earlier results. In summary, all the multichannel models showed that the last-click model overstated an online channel when it was acting as a converter. The differences among multi-channel models were not that great, except for the smaller channels. Also, the mean absolute percentage error and mean absolute error were quite low, which enhances the predictive validity of the statistical approach, and this, in turn, enhances the convergent validity of the multichannel approach. Hypothesis 3 was therefore supported.
DISCUSSION AND CONCLUSIONS
Evidence shows that multiple touch points or funnel stages typically are required before purchase and that these touch points have different effects on purchase likelihood (Xu et al., 2014). The authors examined the hypothesis that multiple digital channels or touch points generate more revenue under the last-click attribution model than under the uniformly distributed, time-decay, position-based, and statistics-based models. They compared the effects of moving to time-decay, uniformly distributed, position-based, and statistics-based attribution models against the current last-click model.
The results show that the last-click model generated the most revenue for the converter and delivered the highest average reward. When the authors compared the last-click model against each one of the other models, the t tests showed a significant difference in the average channel reward value, with the last-click model attributing significantly more to the display channel. Both the revenue and the average channel reward values showed a significant difference between the current model and the others explored.
The revenue attributed to display declined—on average, 22.15 percent—when the authors moved to the other attribution models from the current last-click model. The largest such decline was seen in the uniformly distributed model: 25.46 percent. Previous research failed to explain this phenomenon, except for the suggestion that a modeling approach such as the uniformly distributed one confirms the relevance of the entire sales cycle, as when one channel predominantly plays the role of the converter (Havas Digital, 2010). Any activity that takes credit away from the converter, therefore, likely will be to the detriment of that channel's revenue. The authors concluded from this analysis of the multichannel models that the last-click model overstates display, or any other online marketing tool, when it acts as a converter. The difference among multichannel models, moreover, were not that significant, except for the smaller channels.
The problem with any rule-based attribution model is that it is impossible to predict, on an individual basis, not only the ways customers across multiple backgrounds make purchases but also how the shared value can be allocated equitably among the players according to their individual contributions. Given these constraints, it is too complicated to develop an attribution model for each path to purchase (Anderl et al., 2016; Blake et al., 2015; Shao and Li, 2011). Consequently, one might have to consider the more rigorous variety of statistics-based attribution models as a preferable attribution strategy. These models allow one to provide more stable credit assignments to the digital channels in a purchase funnel; this is evidenced from the forecasting ability of the statistical model, as measured by prediction to a holdout sample.
In the present context, the findings enhance the predictive validity of the statistical approach, and this, in turn, enhances the convergent validity of the multichannel approach. Because the authors also value each medium for its contribution to the end purchase, however, the study understates the value of some key emergent media in the chain, particularly social media. It appears that the attribution models currently do not value fully social media, which often do not directly lead to purchase but can have a strong behavioral impact (e.g., by shaping the consideration set). Although the value of social media does improve as the sophistication of the attribution models increases, the consumer behavior implications need to be accounted for fully in future research.
IMPLICATIONS FOR MARKETING MANAGERS
The most frequently employed attribution strategy is the last-click model. This measurement strategy has significant consequences for all paid advertisers, particularly because it can be used readily to justify online advertising spend in comparison with the budget for offline media, such as television. A major drawback of this attribution model, however, is that it undervalues the consumer click activity that might have preceded the last click leading to a purchase conversion.
Because consumers typically search multiple times before making a purchase, the model attributes final search queries with more conversions than they deserve. This suggests that a better understanding of the purchase funnel will drive smarter marketing campaigns, generating a higher return on advertising spend. Multichannel attribution models have evolved to reflect the growing complexity of attributing credit with each new advertisement format. An advantage of a campaign-based or a rule-based attribution model is that it can be customized to the specific goals of a digital marketing campaign. Because these models allow each click in a stream, not just the last click, to receive credit for a conversion, understanding their true value allows an advertising manager to optimize online spend and establish a higher level of accountability.
As demonstrated in this study, however, these models are based on a number of weak and limiting assumptions. Although an advertising campaign may find that the simplicity of the time-decay multi-impression model or the last-click model has significant advantages over other models, the question remains whether they accurately attribute credit to all the channel impressions in a purchase funnel. This empirical study of a statistics-based attribution model suggests a direction that companies can take when attributing credit to different channels in a purchase funnel. In particular, the authors expound the way a statistics-based attribution model provides stable and accurate interpretations of the influence of each user interaction, although it must be recognized that attribution primarily takes a retrospective view (Kireyev et al., 2016; Shao and Li, 2011).
The problem of attribution modeling can be defined as aligning the incentives of the advertiser with those of the channels hired to run advertisements on behalf of the advertiser. Whereas an advertiser wants to drive as many conversion events as possible at as low a cost as possible, the channels would like the conversion events that the advertiser observes to generate the highest profit possible. In this context, the authors focus more on accurate and stable interpretations of the influence of each user interaction on the final user decision, rather than just on user classification. As they argue, one can achieve this by implementing a statistics-based attribution model that provides stable and accurate interpretations of the influence of each user interaction, a goal that cannot be achieved when one uses a rule-based attribution model.
For example, the authors have examined the assumptions of several attribution models in how they assign credit to different online channels in a purchase funnel. They consider the contribution of the convergent validity of the multichannel models as well as the predictive ability of the statistical model. The study's results show that display loses value under last-click models and that multichannel methods are consistent with each other in how they assign credit to different online channels. The authors thus are able to offer guidance on aggregate-level budget-allocation decisions across multiple forms of advertising.
ABOUT THE AUTHORS
Tahir M. Nisar is an associate professor in Southampton Business School at the University of Southampton, England. He has published numerous articles in distinguished academic journals, including Journal of Retailing, Computers in Human Behavior, and Technological Forecasting and Social Change. His current funded research is on digital and social media analytics, big data, and attribution modeling.
Man Yeung is a senior associate at Digital Lab, London and a research fellow at the University of Southampton. Prior to joining academia, Man spent a number of years working in international business at a senior level. He has provided his services for over a decade in leading EC- and BEIS-funded projects. Man's research interests are primarily in the area of digital networks, big data, and Internet security.
APPENDIX Data Collection (and Model-Based Revenue Distribution)
First, the authors considered the example of last-click model-based revenue distribution, using Microsoft Excel. As discussed, the last-click model is the most simplistic and straightforward method. With respect to Table A1, the data highlighted were rewarded with the revenue stated. Consequently, the display advertisement gained the credit for driving the sale. This always is the last step in the customer's journey, and in the case of a one-step journey it is the only step.
When this model is applied to all of the data, first they are separated into the total number of steps in the journey (e.g., Table A1 displays a four-step journey). With the sort function, the data are placed alphabetically from the last cell (the rewarded cell) to group into the various online-marketing tools, such as e-mail, social media, and price comparison. From here, the revenue for each online marketing tool can be calculated, and then all revenues are totaled for the length of customer journeys.
For instance, Table A2 shows a cell representation of a three-step journey according to the time-decay model (a three-step journey is divided up under the ratio of 1:3:6, with 10 percent, 30 percent, and 60 percent attributed to the steps in chronological order). In calculation of the total values, the revenue is multiplied by the required percentage, and then that amount is allocated to that step.
Table A3 provides an example of a time-decay three-step journey. For each marketing tool, the revenue generated then is totaled to show the revenue generated by that tool for the time-decay model.
- Received February 24, 2016.
- Received (in revised form) March 1, 2017.
- Accepted April 4, 2017.
- Copyright© 2018 ARF. All rights reserved.