The Implications of Improved Attribution and Measurability for Online Advertising Markets

Catherine Tucker November 20, 2012

Abstract Digital data has transformed the ability of advertisers to assess the performance of their online paid and display advertising campaigns. This has led to the launch of a new breed of technologies allowing cross-channel attribution. These technologies allow advertisers for the rst time to compare the performance of ads across dierent types of advertising channels and assess which campaigns are the most successful at achieving their goals. The theoretical literature in economics suggests that the improvements in measurability implied by the collection of digital data should both make switching between media platforms easier and reduce prices for advertisers. This implies positive welfare consequences. This paper explores whether there is empirical support for these theories. The paper nds evidence that advertisers are able to use these attribution technologies to substitute towards advertising campaigns across dierent platforms that lead to more conversions at a lower price. Specically, these technologies allow advertisers to develop more rened targeting criteria, which helps them to avoid paying high prices. This suggests that there may be unforeseen costs to policies that reduce the ability of advertisers to measure the comparative performance of their advertising.

Catherine Tucker is Associate Professor of Marketing at MIT Sloan School of Management, Cambridge, MA, and Research Associate at the NBER. All errors are my own. I thank Google for nancial support for this project.

Contents
1 Introduction 2 The evolution of attribution technologies 2.1 The Growth of Attribution Technologies Online . . . . . . . . . . . 2.1.1 Collection of Data on Advertising Exposure . . . . . . . . . 2.1.2 Collection of Data on Conversion . . . . . . . . . . . . . . . 2.1.3 Attributing Causality . . . . . . . . . . . . . . . . . . . . . . 2.1.4 Automation of Media Spend Based on Return on Investment 2.2 Consequences of Attribution . . . . . . . . . . . . . . . . . . . . . . 3 Theoretical Eects of Increased Measurability 3 4 5 7 8 8 10 11 15

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

4 Empirical Analysis of Advertising Outcomes in the Presence of Attribution Technologies 19 4.1 Analysis of a single-channel attribution technology . . . . . . . . . . . . . . . 19 4.2 Analysis of a cross-channel attribution technology . . . . . . . . . . . . . . . 25 5 Implications and Policy Discussion 33

Introduction

The digital revolution has often been heralded for the transformation it has implied for digital advertising. In particular, the ability to collect data about the individual, parse it automatically and then serve ads on that basis has transformed the ability of advertisers to target ads and show specic ads to specic groups of consumers. This targeting revolution has led to the emergence of paid search advertising, where search engines serve ads on the basis of billions of possible search terms. In online display advertising, it has led to contextually-targeted banner ads which accurately match an ad to the content the consumer is reading, and behaviorally-targeted banner ads where the advertiser can use past browsing behavior to target ads. Far less discussed in the academic literature is what this digital advertising revolution implies for measuring advertising eectiveness (Goldfarb and Tucker, 2011b). Tracking a users clickstream across websites allows far more accurate measurement of the performance of advertising and establishing what kinds of ads work and what kind of ads dont work. Though this revolution has not attracted much academic interest, the advertising industry has responded swiftly to this increased potential for accurately measuring advertising eectiveness. A new type of advertising rm and technology has emerged, specializing in the measurement and attribution of advertising performance. By 2009, 31% of internet rms were actively using cross-channel attribution technologies (Forrester, 2009). These technologies allow advertisers to assess and compare the relative performance of dierent online (and sometimes oine) advertising platforms, such as search advertising, display advertising, social media and direct mail campaigns. The existing theoretical literature on advertising and technology implies that there may be positive eects on consumer welfare when rms can accurately measure the relative return on investment of dierent advertising campaigns and channels. This paper attempts to

provide some initial empirical tests of this theory, by looking at the evolving behavior of advertisers after adopting an attribution technology. I use two dierent datasets that track advertiser behavior after the adoption of new technologies which facilitate attribution. One dataset is based on advertising allocation decisions within a single large platform. The other dataset is based on advertising allocation decisions across dierent paid search providers and display advertising campaigns. Analysis of this data suggest that the ability of rms to gauge ad performance allows them to use far more rened targeting criteria. This substantially reduces the cost of advertising by allocating campaign dollars to advertising platforms that give a higher return on investment.

The evolution of attribution technologies

The quote I know half my advertising is wasted, I just dont know which half has been variously attributed to Henry Ford and John Wanamaker. It illustrates that when it comes to advertising, the biggest problem advertisers face is that they know advertising is useful, but they do not know which particular advertising is useful. Substantively, there are three steps in measuring the eectiveness of advertising: 1. Observing whether or not a consumer is actually exposed to an ad. 2. Observing whether or not the consumer takes the action that the advertising is intended to promote. 3. Identifying which (if any) of the multiple potential ads that a consumer was exposed to was actually responsible for the consumer taking the action. Oine, all of these steps are hard. For traditional media, because advertisers do not track individuals and what ads they see, it is hard to know whether a particular person has been exposed to a television, print or radio ad (Assmus et al., 1984). A few traditional media techniques help, like embedding an identifying code in a Macys coupon in a newspaper, or 4

making a TV infomercial that uses an identiable phone number. In the oine world, many retailers do not observe exactly who purchases their products, let alone whether it is the same person who was exposed to their ads. Last, even if rms can observe a clear link between someone seeing an ad and then buying the product, it is not clear that there is a causal link between the two. This is the classic endogeneity problem of advertising. It could be merely that the kind of person who was chosen to be exposed to that kind of advertising is also more likely to purchase the product. For example, even with the coupon, Macys cannot observe whether the people who used the coupon would have bought anyway. These are all well-recognized as limitations of empirical studies in oine advertising. 2.1 The Growth of Attribution Technologies Online

By contrast the history of online advertising over the past decade has been the development of techniques and tools that specically address all these shortcomings of oine media. This has led to exponential growth in rms that specialize in what is often referred to as cross-channel attribution. Typically, cross-channel attribution technologies do four things: 1. They collate data on who has been exposed to what ads across a rms many advertising campaigns. 2. They match this data with whether or not there is a record that individual has converted. For online businesses, this conversion is usually a sale. For oine businesses, this may be registering on the website or using an electronic coupon. 3. They use a probability model to assess which combination of ads to attribute the successful outcome to. 4. (Sometimes) They automate and optimize the adjustment of media spend on behalf of the advertiser, to reect the relative return on investment of dierent advertising channels 5

Figure 1: Screenshot from typical cross-channel attribution technology

Source: Icrossing Figure 1 is a screenshot of a cross-channel attribution providers websites, showing that these technologies typically provide easy-to-read dashboards that allow easy attribution of conversions in dierent advertising channels. At the turn of the twenty-rst century, such technologies were unheard-of, but now they are widely used. According to Forrester (2009), of 275 website decision-makers surveyed in 2008, 52% agree that such cross-channel attribution technologies would enable them to spend marketing dollars more eectively. 31% report that they are actively using attribution technologies today. A more recent survey by Econsultancy (2012) suggests that 62% of marketers and 77% of agencies are using attribution technologies.1
1 This gure is taken from Econsultancy (2012). The results are based on a survey elded to marketers and agencies between September 26th and October 23rd of 2011, yielding 607 responses. The regional decomposition of advertisers was: 44% North America, 33% United Kingdom, 12% Europe (non-UK), 6% APAC, 5% Other.

2.1.1

Collection of Data on Advertising Exposure

Attribution technology evolved in part due to the relative ease of collecting data for the paid search and online display advertising channels. One of the attractive properties of search advertising is that advertisers can easily trace visitors to their website who navigated to their website because they clicked on an online search ad. Such arrivals are usually directed to a specic landing page or visit an identiable URL. Further, advertisers can use their internal server logs to track how someone behaves who reached their website after clicking on an ad associated with a particular search term (Chan et al., 2011). At rst, online banner ads were much like electronic billboards, and advertisers could not track exposure unless a consumer clicked on them directly (Chatterjee et al., 2003). However, as click rates fell and advertisers began to use display ads primarily for branding, what became crucial for understanding their eect was not clicks but instead impressions, or whether a consumer was directly exposed to an ad. In response to this need, advertisers developed systems of tracking exposure to ads via the use of pixel tags. Very simplistically, pixel tags work as follows. Each time a user visits a website with a pixel tag embedded in an ad, the pixel tag downloads from a remote server, allowing the advertiser or advertising network to record the time and page that person saw (with the associated IP address and/or cookie).2 Though data is easy to collect for display, paid search and organic search advertising channels, recently the ambitions of many attribution technology developers have broadened to encompass many new digital advertising channels. Figure 8 shows that attribution technologies are developing the ability to compare the relative performance of organic leads,
One issue in online display advertising has been that it has been hard for advertisers to observe whether a consumer was actually exposed to an ad, as sometimes banner ads can be at the bottom of a webpage but unseen by the browser. This has been addressed by the IAB in their recent Guidelines for the Conduct of Ad Verication, Feb 2012, and by the development of above and below the fold categories for ads.
2

online videos, Twitter, Facebook and mobile advertising. Another recent development has been the ability of these technologies to integrate oine media including television; catalog; direct mail; point of sale; and call center into their attribution models (Osur, 2012). This is pertinent because competition policy has often treated the online advertising market as separate from the oine advertising market (Goldfarb and Tucker, 2011e). However, the fact that these new technologies allow advertisers to substitute automatically between oine and online channels suggests that online and oine advertising channels may turn out to be far easier substitutes than previously supposed. Unlike online and oine advertising this is unlikely to be instantaneous, since there are more substantial lead times for buying radio and television ads or planning a direct mail campaign. However, with this caveat these advances do echo earlier evidence of substitution between oine and online advertising markets presented in Goldfarb and Tucker (2011a,d). 2.1.2 Collection of Data on Conversion

As long as a transaction is conducted or initiated online, it is relatively straightforward to link this activity to previous advertising exposure. For search advertising, tracking is relatively straightforward, as the search ad directs the consumer directly to the webpage where such transactions can be tracked.3 Over the past few years, display ads have developed a similar capacity whereby they can use a combination of cookies and pixel tags to match a conversion event on an advertisers website with an individual user prole that records the various display advertising campaigns that users have seen previously. 2.1.3 Attributing Causality

Attribution technologies do two things which depart from previous practice. First, they use large-scale probability models which attempt to disentangle patterns from variation in expo3 Search engines have even experimented with a cost-per-action model of payments which facilitated such linkages even more. However, as pointed out by Dellarocas (2012), such models can lead to coordination conicts.

Figure 2: Advertising Channels linked to Attribution

Source: Osur (2012)

sure to dierent ads and take into account correlations among the rates of ad impressions, website visits and conversions (Braun and Moe, 2012). In other words, the sheer scale of data collection which is possible in an online environment allows advertisers to use large-scale data to tease apart what advertising works and what doesnt. Even more novel are their attempts to recognize that typically it is not just a single advertising event which leads a consumer to ultimately convert. Instead, it is a combination of events. Figure 3 describes the dierent steps in a typical consumers path towards conversion for a consumer product. Echoing older marketing models of a funnel, these attribution funnels recognize that there are dierent stages in the purchasing process, and that dierent advertising media may play dierent roles. Prior to the evolution of advertising technologies, advertisers tended to focus on the last click and attributed the entire motivation for conversion to this last click. Such procedures, however, ignored the work done by other media that have built brand and product awareness. In particular, they tended to favor search advertising over display advertising as paid search ads are often the last ads that a consumer sees. Figure 4 summarizes the dierent attribution methods used by typical attribution technologies, and suggests that while last click is still prevalent as an attribution methodology, other methods are growing. Figure 5 shows that often these technologies allow clients the exibility to choose the parameters of their own attribution algorithm. 2.1.4 Automation of Media Spend Based on Return on Investment

One of the major changes in the past three years is that such technologies have begun to automate ad allocation across channels depending on relative return on investment. In other words, the advertiser does not have to actively change and manage their campaigns in response to information about which advertising platform is performing better. Instead, the attribution technology automatically allocates a larger advertising budget to betterperforming advertising.

10

Figure 3: Dierent Advertising Events That Can Lead to Conversion

Source: C3 Metrics, Inc. Osur (2012) suggests that half of all vendors oer such a capacity and 30% of reference clients were using this capacity. Figure 6 summarizes the current use of such attribution technologies by advertisers. It is noticeable that 64% of clients are already using technologies to evaluate the relative performance of dierent advertising channels and allocate advertising appropriately. 2.2 Consequences of Attribution

Attribution technologies, by brokering data across multiple advertising platforms, help to reduce fears that the importance of data for measurement and targeting might lead to a natural monopoly in online advertising (Clemons, 2010). By brokering data across multiple platforms, attribution technology agencies reduce the tendencies of the advertising-supported internet towards concentration and increase the ability of smaller advertising platforms to compete. Figure 7 summarizes survey responses on how technologies aected the behavior of in11

Figure 4: Range of Attribution Methods

Source: Econsultancy (2012)

12

Figure 5: Attribution technologies allow some exibility in how attribution is performed

Source: Ignition One

13

Figure 6: How Advertisers use Attribution Services

Source: Osur (2012)

14

Figure 7: Changes to Advertiser Behavior from Attribution

Source: Econsultancy (2012) dividual advertisers. It emphasizes that one of the major eects of such technologies is to facilitate transition between dierent advertising channels such as search and display. Figure 8 provides further evidence that one of the consequences of such technologies has been a shift away from print media towards more eective digital media. This again appears to support an interpretation of widespread substitution between oine and online advertising channels (Goldfarb and Tucker, 2011e). As this paper is interested as well in the potential consequences of attribution technologies on advertising market prices and outcomes, I turn to the theoretical literature on advertising technologies and market outcomes to build a theoretical framework before moving to more detailed data.

Theoretical Eects of Increased Measurability

The question this paper seeks to answer is what the likely implications are for consumers and advertisers of this burgeoning cross-channel attribution technology. Earlier work such as Hu (2004) points out that the increased measurability of online advertising means that 15

Figure 8: Changes to Channel Investments Resulting from Attribution

Source: Econsultancy (2012)

16

online advertising platforms can oer performance-based pricing in ways impossible with the traditional types of advertising discussed by Anderson and Coate (2005); Ferrando et al. (2008); Reisinger (2012) or in the more general two-sided market (Armstrong, 2006). This can lead to an increase in incentives for advertising platforms to perform well. However, such work does not address the question of what the increased comparability of dierent advertising channels implies for advertising and their customers in general. Since the work of Grossman and Shapiro (1984), there has been a slew of theoretical literature that has studied the potential eects of improvements in targeting technology on advertising markets, where targeting is often modeled as ensuring that ad exposures are not wasted. However, there has been little literature that directly models what happens to advertising platform competition when advertising becomes more measurable and conversions are better-attributed. The literature on the implications of targeting improvements may still be relevant. This is because practically, improved targeting is modeled as reducing the proportion of consumers who see ads that are not relevant for them. Cross-channel ad attribution technology should achieve a similar outcome, though the process is dierent. The dierence in process is simply that targeting technologies prospectively identify better eyeballs through a theory about who responsive customers may be. Attribution technologies retrospectively use better data to identify campaigns that were more successful and therefore presumably reached better eyeballs.4 However, the end result of facilitating ads reaching more receptive eyeballs is similar. This dierence in nuance may be absent in the theoretical literature because, for reasons of simplicity, theory models focus on a static rather than multiple-period model of advertising allocation. Using this insight enables me to revisit the original cost function A(; ) that Grossman
4 Another relevant branch of theory is Halaburda and Yehezkel (2011) who study how asymmetries in quality information can lead concentration of market power to be welfare-improving.

17

and Shapiro (1984) hypothesized as applying to advertising. In their model, captures the fraction of the target population that is exposed to a message and captures advertising technology which drives the total and marginal advertising costs. In their model, attribution technology and automated cross-channel attribution could lead to a reduction in if an ad can be delivered to the population with less waste and more accuracy. The next question is what the equilibrium eects on advertising prices and competition will be, due to these reduced costs? Earlier work in this area such as Chandra (2009) that studied prices of newspaper advertising, suggested that improved targeting could lead to higher prices in advertising markets due to advertising platforms being able to capitalize on delivering better performance to advertisers and the increased attractiveness of advertising leading advertisers to bid up prices. However, such ndings has been challenged by a recent literature when it comes to the digital era. For example, Bergemann and Bonatti (2011) study the impact of targeting on competition between advertising platforms. They point out that improvements in targeting technology, by reducing the number of advertisers competing for each consumer, actually reduce the pricing power of a dominant platform, meaning that advertisers pay lower prices. Similarly Athey and Gans (2010) suggests that improved targeting actually increases eective supply of advertising if advertising space is limited. Translating this result to the question of measurability implies that improved measurability may reduce the pricing of advertising in the presence of platform competition, because it encourages advertisers to jettison unfruitful campaigns and therefore increases the availability of space to display eective campaigns. Levin and Milgrom (2010) go even further and suggest that the thinness of markets implied by excessively ne targeting can lead to downward distortions in prices, and allow advertisers to game publishers by paying low prices for valuable inventory. Therefore, the consumer welfare eects of improved measurability would seem likely to be positive if the reduced prices of advertising for advertisers are passed on to consumers. 18

However, a key condition for this positive outcome is that rms actually behave in the manner predicted by the latter set of theories, and use this improved measurability to increase the eciency of advertising allocation by seeking out sets of target consumers for whom there is less of a price premium.

Empirical Analysis of Advertising Outcomes in the Presence of Attribution Technologies

This paper uses two dierent datasets to explore the predictions of the theoretical literature. The rst is data on advertising allocation decisions within a single media platform after the introduction of a technology which improved the measurability of advertising. The advantage of focusing initially on the eects of a single-channel attribution technology is that the eects are relatively clear-cut and easy to understand. However, of course, cross-channel attribution technologies probably have the most potential for transformative eects of online advertising markets. Therefore, I also use a second dataset that traces out the advertising allocation decisions of advertisers after adopting a media attribution system that allowed them to compare the performance of search advertising and display. In all cases, the data analysis is relatively simple and descriptive. I do not have access in this data to a counter-factual - that is, what would have happened if the advertiser had not adopted the technology, so I do not make strong causality claims. However, the simple comparative statistics do allow me to study whether the changes I observe in advertiser behavior are in line with what is described in the theoretical literature in Section 3. 4.1 Analysis of a single-channel attribution technology

This dataset stems from advertising allocations decisions made by a single rm on an individual media platform. Ads could be targeted on the basis of the keywords which were based on users characteristics that they had shared on the platform. Display advertising on this platform was priced by a second-priced position auction. This auction mechanism is 19

Table 1: Summary statistics for Single-Channel Attribution Technology data Mean Clicks 16.15 Impressions 29388.80 ClickThrough Rate 0.002 Cost Per Click (USD) 1.08 Number Targeting Criteria 8.19 Observations 21463 Std Dev 70.79 117621.16 0.05 0.53 4.51 Min 39 55497 0 0 0 Max 2691 4726149 1 4 21

the same as is used on major search engine platforms. The way it works is similar to the mechanism described by Edelman et al. (2007); Varian (2007); Athey and Ellison (2011) advertisers choose which keywords to advertise their products, and bid the maximum amount they would be willing to pay for a click. Very roughly, the advertiser pays the price of the second-highest bid for that ad slot, though this is adjusted to reward advertisers who achieve high click-through rates. The advantage of this type of price mechanism for advertisers is that prices reect demand for that particular ad slot. The advantage of this type of price mechanism for advertising platforms is that it allows the automation of price-setting. At the starting point of a dataset, the rm had adopted a tracking system that allowed them for the rst time to measure the relative eectiveness of dierent campaigns and automatically use this information to allocate their advertising more eciently. The rm shared data that spanned the three months after initial adoption. I compare ad performance for the 9,665 separate campaigns run in the rst week after using this new attribution technology, with the ad performance of the 11,798 campaigns that were run in the nal week of the data. Table 1 reports summary statistics for this data. An observation is a campaign launched on a particular day. On average each campaign was seen by nearly 30,000 dierent people, which in industry terminology is an impression. Clickthrough rates were quite low, in line with the rest of the display ad industry, at less than one-fth of one percent (Tucker, 2012). Figure 9 reports the change in click rates for the average campaign at the beginning and end of this period. Figure 10 reports the change in impressions. Figure 11 shows how this 20

Figure 9: Single Channel: Change in number of clicks for an average campaign

Figure 10: Single Channel: Change in number of exposures of ads for an average campaign

21

Figure 11: Single Channel: Change in clickthrough rate of ads for an average campaign

translates into click rates. Generally, these statistics suggest that there was little change in actual advertising and exposures from the adoption of this technology. Therefore, a natural question is what the technology actually achieved. This is identied by actually looking at the cost of each of these clicks. For this platform, the advertiser paid each time a consumer clicked on an ad, so this was the major driver of costs. Figure 12 shows how the cost per click changed. There is a striking reduction in how expensive each click was. In other words, one of the major benets of the adoption of this attribution technology may have been a large reduction in costs. The next question is how this happened, given the relative lack of change in click-through rates and ad exposures? Figure 13 gives some idea about the mechanism behind this change in costs. The rm (presumably as a result of adopting this attribution technology) substantially increased the number of criteria it used when selecting its target market. Similar to the predictions of Bergemann and Bonatti (2011), this increase in the thinness of the market

22

Figure 12: Single Channel: Change in cost per click of ads for an average campaign

and led to lower prices per click.

23

Figure 13: Single Channel: Change in number of targeting criteria for ads in an average campaign

24

4.2

Analysis of a cross-channel attribution technology

The second dataset was provided by a rm that allows cross-channel attribution of dierent forms of online advertising campaigns. It allows me to study the evolution over three quarters of data of advertising campaigns by a single advertiser who had adopted an attribution technology. This technology allowed the advertiser to easily compare the conversion performance of each dierent campaign. In this attribution platform, unlike the previous data, the cost information concerning advertising is stored on the local server of the advertiser. Therefore, initially I simply document patterns in the data suggestive of substitution across advertising platforms and towards advertising campaigns that were more successful. I then collect additional external data changes in pricing of the paid search campaigns implied by the changing substitution patterns.5 Table 2 reports summary statistics for this data. An observation in this data is a single user who has been exposed to a variety of advertising campaigns. Typically the user is observed for 40 days. It is immediately apparent that paid search ad exposures are a small proportion of total advertising exposures for the typical individual. This is presumably because anyone can potentially see a display ad for a new cellphone, but users are only exposed to a paid search ad for a new cellphone if they are actively in the market for a new cellphone, which is a potentially a very small proportion of the population. An observation in the data analysis is a campaign that either started in the rst, second or third quarter. I analyze data at the campaign rather than the user level because I am interested in changes in campaign level performance. I have data on 883 such campaigns that were initiated over three quarters. 58% of these campaigns were search campaigns, and the remaining campaigns were a mixture of untargeted display ads, behaviorally targeted display ads and social media ads.
The pricing of search campaigns is easily accessible via the Google keyword estimator tool. Unfortunately, the prices of display advertising campaigns are less transparent.
5

25

Table 2: Summary statistics for Cross-Channel Attribution Technology data Mean Average # Days User Exposed to Ads for Campaign 40.82 Average # of Search Ads Seen 0.02 Average # of Display Ads Seen 45.55 Average # of Untargeted Display 32.73 Average # of Behaviorally Targeted Display Ads Seen 11.78 Average # of Social Media Display Ads Seen 1.05 Proportion of Conversions 0.02 Observations 358776 Std Dev 41.77 0.27 137.25 105.87 70.38 13.11 0.15 Min Max 0 218 0 58 0 4954 0 4954 0 4549 0 2850 0 1

Figure 14 traces the improvement in conversion rates associated with campaigns across three quarters after the adoption of this cross-channel attribution technology. It suggests that over time, the conversion rates associated with the average advertising campaign increased. Of course, a natural question is whether the improvement can be attributed exclusively to the attribution technology. Since I do not observe a counterfactual, that is would have happened if the rm had not adopted a attribution technology, I do not make strong causal claims and simply present the data as correlational. However, it is worth noticing that there does appear to be an appreciable increase in terms of orders of magnitude from the baseline. Such a large change does not support an alternative explanation resting on small incremental changes within a rm. It also does not appear that there are any clear seasonal trends (such as the holiday season) which would otherwise explain the pattern I observe. I next decomposed these statistics by advertising channel. Figure 15 suggests that there was an improvement in conversion probabilities for online display advertising. Figure 16 suggests there was a large improvement for search. Both increases were large. The eect on search is larger in absolute terms, because search started o from a higher baseline conversion rate. This may be because people who are interested enough to search are also going to be more likely to buy. Figure 16 suggests that the improvement was reasonably equal across Google and its competitors. This is reassuring that these attribution technologies do not

26

Figure 14: Cross-Channel: Change in conversion rate associated with campaign over time

appear to favor one advertising platform over another. Figure 17 examines how the length simply in terms of number of characters of the search terms used for targeting campaigns over over time. It suggests that much like the example of the single-channel attribution technology, some of this improvement in conversions can be traced back to the use of more detailed criteria to target ads. The next question is whether this increase in conversion rates aects prices in these advertising markets. In particular, do these higher conversion rates come at a higher cost? I did not have access to direct pricing data, so I collected data using Googles Trac Predictor tool on the likely cost per click of each of the dierent paid search ads for this rm.6 In general, as shown by Figure 18, the price per click dropped somewhat for search ads. However, what was startling was how the eective cost per conversion fell far more rapidly, as shown
As noted in Goldfarb and Tucker (2011d), in expectation this predictor tool gives unbiased estimates of prices, though ex post the prices paid may dier depending on how the search engine rates that advertisers quality
6

27

Figure 15: Cross-Channel: Change in conversion rate associated with display campaigns over time

Figure 16: Cross-Channel: Change in conversion rate associated with dierent campaigns at dierent paid search providers over time

28

Figure 17: Cross-Channel: Length of targeting criteria for paid search ads over time

by Figure 19. The dierence in price decline between Figure 18 and Figure 19 makes sense, given that advertisers could already use the internal search engine metrics to improve cost per click, but that it is only with the evolution of attribution technologies that they could improve cost per conversion. Due to the fact there is less transparency in display advertising markets, I was not able to get likely cost estimates for the display campaigns in my data. I was, however, able to get useful data about the typical length of a display advertising campaign or how long it ran before it was pulled by the advertiser. This can act as a proxy for costs, since the price of display campaigns is typically determined by the number of impressions which in turn is determined by the length of the campaign. Of course making this comparison assumes that in each case the reach (or number of impressions) on each day was similar for each campaign.7 If less eective campaigns were running for less time after the adoption of the
A quick comparison of impressions data rules out alternative more mechanical explanations for Figure 20 such as the number of daily impressions for each campaign increasing over time.
7

29

Figure 18: Cross-Channel: Price paid per click for search ads over time

Figure 19: Cross-Channel: Price paid per conversion for search ads over time

30

Figure 20: Changing length of Campaign for Display Ads by Campaign Success

attribution technology, then this implies that the platform was able to reduce costs. Figure 20 suggests that indeed this was the case. By the nal quarter, underperforming campaigns were being run for a substantially shorter time. One general concern with the interpretation of these results is that conversions are so high, for paid search. This is a concern as there is always the potential for advertising exposure data to reect an endogeneity bias (Lewis et al., 2011). Specically, the concern is that consumers would have been exposed to these campaigns would have would have purchased anyway even in the absence of advertising for reasons I do not observe. Though I do not have the data to address this directly in this paper, it is relatively straightforward to address such endogeneity problems with a randomized test, sometimes referred to as a split-test or an a+b test. Under these tests, consumers are randomly allocated to either seeing the ad in question or a placebo ad for something like the Red Cross. The automation inherent in online advertising facilitates the use of such large-scale eld tests to evaluate what kind of

31

advertising has the largest incremental eect rather than being simply associated with the largest average eect (Lambrecht and Tucker, 2011; Goldfarb and Tucker, 2011c). The use of such eld experiments to improve ad performance has been discussed as the next wave of improvements by various cross-channel attribution providers (Chittilappilly, 2012), so such data to validate identication may be available in future studies.

32

Implications and Policy Discussion

There are two separate sets of policy implications that can be drawn from this analysis. The rst set of policy implications are the consequences of these new attribution technologies for understanding how advertising markets work. Theory suggests that attribution technologies could have two potential conicting implications. First, if they led advertisers to cluster and target just a few small groups of receptive user eyeballs, then this could lead to higher prices paid for advertising (Chandra, 2009). Alternatively, if attribution technologies lead advertisers instead to substitute away from higher-priced advertising options and target ever-thinner markets, this could lead to lower prices paid by advertisers for advertising (Bergemann and Bonatti, 2011). These competing hypotheses in the theoretical predictions make the consequences of cross-channel attributions an empirical question - one I seek to answer in this paper. In two dierent datasets, I nd repeated evidence that the use of attribution (and improved advertising measurement) technologies appear to be associated with lower prices for advertising. I present evidence that this appears to be associated with advertisers using these attribution technologies to identify subsets of consumers to advertise to. The advantage of these more nuanced subsets of consumers for the advertiser, is that the advertising platform is unable to charge a premium because there are fewer other advertisers bidding up prices. I also present evidence that in general these attribution technologies are being used by advertisers to allocate resources across online and display advertising, and facilitate substitution across these dierent media platforms. Theoretically, this is important for competition policy, both for understanding the right market denition, and also because it helps alleviate concerns that the economies of scope in the collection of data for measurement purposes might favor any one advertising platform (Clemons, 2010). Given this set of apparent benets, the second set of policy implications concern the po-

33

tential consequences of inhibiting the diusion and use of these attribution and measurement technologies. The most obvious of these are that underlying these attribution technologies is a great deal of anonymized data, commonly about an individual cookie. The scale and scope of data collection for ad analytics is currently under discussion by policymakers involved in privacy policy globally. As set out by Evans (2009); Lenard and Rubin (2009); Goldfarb and Tucker (2011f), there is a trade-o between the protection of consumer privacy online and the ability of rms to use prior clickstream data to target ads eectively. There are reasons to think that policy governing data used for advertising analytics may even be a more problematic policy question for policymakers than clickstream data used for targeting. Unlike the benets of targeted ads, which might include more relevance and less irritation for consumers, the benets of ads whose performance is better measured is not immediately obvious to consumers. Instead, the ndings of this paper suggests that the benets of better advertising analytics are indirect, since such technologies lead advertisers to pay lower prices for ads, and these lower costs should in competitive markets translate to lower end prices for consumers. Therefore, it seems likely that when deciding whether or not to share data for advertising analytics purposes, consumers may not internalize the general welfare gains of providing better information to advertising platforms, since they are not immediately apparent to the consumer. This may lead to extremely low opt-in rates for consumers when it comes to acceptance of third-party cookies designed to perform third-party advertising analytics such as used for the technologies studied in this paper. As of November 2012, privacy policy makers in the EU and the US disagree when it comes to privacy policies surrounding the use of third-party cookies for advertising measurement and attribution (EU, 2012). Specically, the W3C standards-setting organization currently contemplates exempting third-party cookies from opt-in requirements for the US Do Not Track standard if they enable Frequency Capping, Financial Logging and Auditing and 34

Aggregate Reporting (http://www.w3.org/TR/tracking-compliance). Such uses could conceivably allow attribution technologies such as the ones studied in this paper to persist. By contrast, the EU Working Party 29 takes the view that for this proposed Do Not Track standard to be in compliance with EU law for companies serving cookies to European citizens, Do Not Track must eectively mean Do Not Collect without exceptions (EU, 2012). These issues are going to be even more pronounced when it comes to new technologies that try to allow advertisers to do multi-channel attribution that involve mobile ads, or oine advertising such as direct mailing or television. Obviously, to ensure eective multichannel attribution requires a far greater scale of data and scope of data collection about an individual consumer. Policymakers will have to trade o the welfare benets of increasing the ability of advertisers to switch to lower-priced and more eective platforms more globally than they do now, with this increased scope and scale of data collection. In general, though, it is noticeable that in these discussions over privacy policy there has been scant attention paid to the potential implications for competition policy (Campbell et al., 2011), and in particular to the increased diculty for advertisers to easily seek lowerpriced alternatives in the absence of measurement. The ndings of the paper suggest that the benets of improved substitution between dierent channels and associated lower prices should be explicitly weighted against the benets of more stringent privacy regulation. There are of course limitations to the ndings detailed in this paper. The rst is that though the datasets analyzed are massive in terms of scale and scope in terms of the coverage of customers, they apply to two advertisers and two attribution technologies at a single point in time. This limits generalizability. The second is that because I lack a true counterfactual in the data, the evidence presented should be interpreted as merely correlational rather than causative. The third limitation is that there are many dierent ways that advertisers could potentially use cross-channel attribution technologies, such as for oine 35

media planning or as part of a more generalized and rigorous scheme of ad testing which I have not explored. Notwithstanding these limitations, I believe this paper to be a useful rst step in understanding how the ability to accurately measure advertising performance across advertising channels may aect advertising markets.

36

References
Anderson, S. P. and S. Coate (2005). Market Provision of Broadcasting: A Welfare Analysis. Review of Economic Studies 72 (4), 947972. Armstrong, M. (2006, Autumn). Competition in two-sided markets. RAND Journal of Economics 37 (3), 668691. Assmus, G., J. U. Farley, and D. R. Lehmann (1984). How advertising aects sales: Metaanalysis of econometric results. Journal of Marketing Research 21 (1), 6574. Athey, S. and G. Ellison (2011). Position auctions with consumer search. The Quarterly Journal of Economics 126 (3), 12131270. Athey, S. and J. S. Gans (2010, May). The impact of targeting technology on advertising markets and media competition. American Economic Review 100 (2), 60813. Bergemann, D. and A. Bonatti (2011). Targeting in advertising markets: implications for oine versus online media. The RAND Journal of Economics 42 (3), 417443. Braun, M. and W. W. Moe (2012). Online advertising response models: Incorporating multiple creatives and impression histories. Mimeo, MIT . Campbell, J. D., A. Goldfarb, and C. Tucker (2011). Privacy Regulation and Market Structure. Mimeo, University of Toronto . Chan, T. Y., C. Wu, and Y. Xie (2011). Measuring the lifetime value of customers acquired from Google search advertising. Marketing Science . Chandra, A. (2009). Targeted advertising: The role of subscriber characteristics in media markets. The Journal of Industrial Economics 57 (1), 5884.

37

Chatterjee, P., D. L. Homan, and T. P. Novak (2003). Modeling the clickstream: Implications for web-based advertising eorts. Marketing Science 22 (4), 520541. Chittilappilly, A. (2012, July 11). Using experiment design to build condence in your attribution model. Online Metrics Insider. Clemons, E. K. N. M. (2010). Regulation of digital businesses with natural monopolies or third party payment business models: Antitrust lessons from the analysis of Google. Mimeo, Wharton . Dellarocas, C. (2012). Double marginalization in performance-based advertising: Implications and solutions. Management Science . Econsultancy (2012). Marketing attribution: Valuing the customer journey. White Paper: In Conjunction with Google Analytics . Edelman, B., M. Ostrovsky, and M. Schwarz (2007, March). Internet advertising and the generalized second-price auction: Selling billions of dollars worth of keywords. American Economic Review 97 (1), 242259. EU (2012, June 7). Article 29 Data Protection Working Party: Opinion 04/2012 on cookie consent exemption. Evans, D. S. (2009). The online advertising industry: Economics, evolution, and privacy. The Journal of Economic Perspectives 23 (3), 3760. Ferrando, J., J. J. Gabszewicz, D. Laussel, and N. Sonnac (2008). Intermarket network externalities and competition: An application to the media industry. International Journal of Economic Theory 4 (3), 357379. Forrester (2009, February). A framework for multicampaign attribution measurement.

38

Goldfarb, A. and C. Tucker (2011a, April). Advertising Bans and the Substitutability of Online and Oine Advertising. Journal of Marketing Research 48 (2), 207227. Goldfarb, A. and C. Tucker (2011b). Chapter 6 - online advertising. Volume 81 of Advances in Computers, pp. 289 315. Elsevier. Goldfarb, A. and C. Tucker (2011c, May). Online display advertising: Targeting and obtrusiveness. Marketing Science 30, 389404. Goldfarb, A. and C. Tucker (2011d). Search engine advertising: Channel substitution when pricing ads to context. Management Science 57 (3), 458470. Goldfarb, A. and C. Tucker (2011e). Substitution between oine and online advertising markets. Journal of Competition Law & Economics 7 (1), 3744. Goldfarb, A. and C. E. Tucker (2011f). Privacy regulation and online advertising. Management Science 57 (1), 5771. Grossman, G. M. and C. Shapiro (1984). Informative advertising with dierentiated products. The Review of Economic Studies 51 (1), 6381. Halaburda, H. and Y. Yehezkel (2011, September). Platform competition under asymmetric information. Working Papers 11-05, NET Institute. Hu, Y. J. (2004). Performance-based Pricing Models in Online Advertising. Mimeo, Purdue University - Krannert School of Managemen . Lambrecht, A. and C. Tucker (2011). When does retargeting work? Timing information specicity. MSI Working Paper 11-105 . Lenard, T. M. and P. H. Rubin (2009). In Defense of Data: Information and the Costs of Privacy. Technology Policy Institute Working Paper . 39

Levin, J. and P. Milgrom (2010, May). Online advertising: Heterogeneity and conation in market design. American Economic Review 100 (2), 60307. Lewis, R. A., J. Rao, and D. Reiley (2011). Here, there, everywhere: Correlated online behaviors can lead to overestimates of the eects of advertising. In International World Wide Web Conference. Osur, A. (2012). The Forrester Wave: Interactive attribution vendors, q2 2012. Forrester White Paper. Reisinger, M. (2012). Platform competition for advertisers and users in media markets. International Journal of Industrial Organization 30 (2), 243252. Tucker, C. (2012). Social advertising. Mimeo, MIT . Varian, H. (2007). Position auctions. International Journal of Industrial Organization 25 (6), 11631178.

40

About the Author

Catherine Tucker is the Mark Hyman Jr. Career Development Professor and Associate Professor (with tenure) of Marketing at MIT Sloan. Her research interests lie in how technology allows rms to use digital data to improve their operations and marketing and in the challenges this poses for regulations designed to promote innovation. She has particular expertise in online advertising, digital health, social media and electronic privacy. Generally, most of her research lies in the interface between Marketing, Economics and Law. She has received an NSF CAREER award for her work on digital privacy and a Gareld Award for her work on electronic medical records. Dr. Tucker is Associate Editor at Management Science and a Research Associate at the National Bureau of Economic Research. She teaches MIT Sloans course on Pricing and the Executive MBA course Marketing Management for the Senior Executive. She holds a Ph.D in Economics from Stanford University, and a BA from Oxford University.