TARGETING

20.3.1 Why it is Different Online

Even though the methods, indices and technologies employed to measure media audi­ences have advanced considerably over the years, advertisers still find it difficult to assess the value delivered by the audiences they purchase.

The HIW dilemma has two sides, reflected separately in the ‘half is wasted’ and the ‘I don’t know which half’ components of the Wanamaker quote. On the ‘I don’t know’ side is advertisers’ uncertainty concerning the value of the audiences they purchase. This is a measurement problem and, because risky products typically sell for less than their risk-free counterparts, advertisers and media firms alike thus have a vested interest in reducing the uncertainty in measures of media audiences. Not surprisingly, there is a size­able industry devoted to reducing the measurement uncertainties that make ad purchases risky. The ‘half is wasted’ side of the dilemma is the fact that the audience delivered by a media service to an advertiser always includes some individuals for whom the likelihood of purchasing the advertiser’s product is slim or none along with the prospective cus­tomers the advertiser really wants to reach.

For online audiences, Internet 2.0 interactivity and cloud computing have made pos­sible new tools for measuring, segmenting, and delivering audiences to advertisers that can be used to address both sides of the HIW dilemma. Because an online publisher has a direct connection to every Internet user that accesses its content, various measures of audience can be collected at the level of the individual audience member, which mitigates long-standing concerns over the representativeness of the audience members sampled to create measures of offline audiences. Further, cookies and other tools can be used to track users’ online activities over time and create behavioral indices of individual audi­ence members’ interests in various products and services for which there are no close offline counterparts. This information can then be used to deliver an advertiser’s message directly to a constructed online audience composed of individuals selected (targeted) for their predicted receptivity to the advertiser’s message. This process, called behavioral targeting,³ addresses both the measurement uncertainty and the unsold audience sides of the HIW dilemma. It also raises a variety of privacy-related policy concerns (Goldfarb and Tucker, 2011).

With tracking a publisher can also keep track of the number of times individuals attracted to its content have been exposed to its advertiser clients’ ads during previous visits to its website, or, if the publisher belongs to an ad network, the network’s records can be used to determine how often members of the publisher’s audience have been exposed to specific advertisers’ ads when visiting the websites of other network members.

An additional advantage of audience measurements for online ads is that user activi­ties, such as clicks on links in online ads and consummated sales following clicks, can also be recorded as more direct measures of an audience member’s attention to and engage­ment with an ad. Most Internet advertising is now sold on a pay-per-action basis.

To summarize, three notable features distinguish the targeting capabilities of online media from those of their offline counterparts. (1) With offline media, all members of a publisher’s audience are delivered the same ads whether they are likely to be custom­ers for the products advertised or not. By contrast, the members of the online audience attracted by a publisher may be shown entirely different sets of ads, based on individual­ized assessments of their consumption interests. (2) Rather than accepting that members of an audience differ in the number of times they have been exposed to its ads, an online advertiser, working through an online publisher or an ad network, can select recipients for its ads based on the number of times they have been exposed to its ads in the past. (3) Activity-based assessments of Internet users’ consumption interests can be used to match ads with individual Internet users. Offline media have no comparable mechanism for matching ads to ad recipients. It is not surprising then that Goldfarb (2014) sees ‘the substantial reduction in the cost of targeting’ as the fundamental difference between online and offline media.

20.3.2 The Economics of Online Targeting

The earliest work on economic aspects of targeted advertising appeared first in manage­ment and marketing journals, where the focus was primarily on ways sellers of consumer goods might use targeted ads to increase profits, often as a vehicle for implementing price discrimination strategies or segmenting markets (see, e.g., Iyer et al., 2005).

Gal-Or and Gal-Or’s (2005) article in Marketing Science was the earliest article on targeting we found that based its analysis on a model of a media firm. Gal-Or and Gal-Or develop a model where consumers are distributed uniformly along a Hotelling line describing the range of variation for a differentiated product. The market is served by two firms, one at each end of the product line. A firm can advertise to increase con­sumers’ awareness of its product and, holding awareness of the other firm’s product constant, sales increase with the level of consumer awareness. Gal-Or and Gal-Or show that the media firm can profit by offering firms the opportunity to have their ads delivered primarily to those consumers located closer to their end of the product line than their competitor’s end. Firms choose this option because differential awareness created through targeted ads softens price competition and allows them to raise prices and profits.

Kim and Wildman (2006) were also early contributors to the targeted advertising lit­erature with a model of competing cable TV networks that, in addition to earnings on their ad sales, collected license fees from the cable systems that carried them. Because viewers disliked ads, a network’s audience size and its license fee earnings varied inversely with the amount of commercial time in its programs. Consumers in television programs’ audiences differed in the sets of products they consumed and advertisers paid only for access to the audience subsets that comprised their potential customers.

Targeting fully entered the mainstream of the economics literature on advertis­ing with AER Papers and Proceedings articles by Athey and Gans (AG) (2010) and Levin and Milgrom (2010). AG in particular have served as a point of reference for subsequent work. AG model a situation in which every consumer belongs to one of a fixed number (say n) of mutually exclusive sets of consumers. For each set of con­sumers there is a corresponding set of products that only they consume. In the story motivating the model, a consumer’s set membership is determined by the local area in which he or she resides and each local area is served by a publication distributed only to local residents. In addition, there is a single general publication that draws its audi­ence from all local areas. Consumers only buy products for which they have seen ads. Consumers are media single-homers and choose either the general publication or their local publication, so an advertiser would have to advertise in both its local publication and the general interest publication to reach all of its potential customers. Because the audience for each publication is assumed fixed, the general publication and the local publications do not compete for audience members. This means that the effects on all publishers of a switch to geographically targeted ads from ads delivered to its entire audience by the general publisher are manifest only on the advertising sides of their two-sided platforms.

Even with its rather severe restrictions, the AG analysis generates several insights that should apply in more realistic settings. One is that targeting can increase the general publication’s profits only if it lowers costs or increases advertising revenue relative what would be observed with untargeted ads. While this observation is trivially true, it also forces us to think carefully about what might make targeting the more profitable strategy. Suppose, for example, that ad clutter doesn’t divert audience members’ attention from their local merchants’ ads and that the publication’s delivery costs do not increase with the number of ads packaged with its content. Then the same ads that might have been delivered to subsets of its audience through targeting could be presented to all members of its audience and advertisers would value the total impressions delivered the same in either case, leaving profits the same whether ads were targeted or not. AG also show that when advertisers are constrained in their media buys, perhaps due to fixed marketing budgets, the general publisher and local publishers compete for the marginal impression purchased by each advertiser and this competition depresses per impression prices to the detriment of the local publishers.

While it may seem obvious that subdividing an audience into more homogeneous sub­groups of consumers for delivery to advertisers would cause price per impression to rise, Bergemann and Bonatti (2011) use a model with single-homing media consumers to show that this is not necessarily the case. The reason is that when the audience offered to adver­tisers contains a mix of consumers who differ in the products they consume, sellers of all the products consumed by audience members will want to reach this audience. However, as the mix of consumers in a targeted audience becomes less diverse, the number of advertisers bidding for access to that audience declines. Bergemann and Bonatti (2011) find that while price per impression increases initially as targeting becomes more precise, eventually thinning of the advertiser side of the market reduces the intensity of compe­tition for impressions and price per impression can fall, potentially quite dramatically, a possibility first noted by Gal-Or and Gal-Or (2005). This possibility is also highlighted in Levin and Milgrom’s (2010) broader discussion of the pricing implications of target­ing, where they further point out that some advertisers may prefer audiences with fairly diverse mixes of consumers over ones that are narrowly targeted.

The models of targeted advertising just reviewed are all single-period models that assume publishers know enough about their audience members’ preferences over con­sumer goods to implement a targeting strategy. However, the tracking technologies employed to develop that knowledge can also be used to record individuals’ exposures to online ads and then dynamically adjust the mix of ads delivered to individual Internet users over time. Athey et al. (ACG) (2013) represent the first significant effort to model this process and explore its implications for advertising market equilibria.

ACG present a two-period model of an advertising market that contains two pub­lishers and many advertisers and consumers. Each consumer visits one of the two publishers each period, with the choice determined stochastically. As a consequence, the two-period audience for each publisher consists of a mix of ‘loyals’ who are in the audience both periods and ‘switchers’ who visited a different publisher each period. The full effect of an ad on a prospective consumer is realized after a single expo­sure and, while any individual advertiser assigns the same value to a first impression regardless of the consumer who is impressed, advertisers differ in their valuations of first impressions. So the market demand for first impressions is downward sloping. Publishers set per impression ad prices and advertisers choose the number of periods they purchase ads from each publisher. ACG identify publishers’ profit-maximizing pricing strategies when tracking can only be used to ensure that a member of its audi­ence is not presented with the same ad twice within a period, when tracking makes it possible for a publisher to guarantee advertisers that their ads will not be presented to its loyals twice over the two periods, and when it is possible to track exposures within the publishers’ combined audience and the same no duplication guarantee is offered for switchers as well.

ACG’s results provide some theoretical support for the conventional wisdom that ability to use tracking to sell advertisers audiences with fewer duplicated exposures should increase publisher profits and, depending on how this influences per impres­sion prices, benefit advertisers as well. But they also show that there are conditions under which publisher profits are lower for a tracking equilibrium than for an equi­librium with no tracking. This is due to higher-value advertisers being willing to pay considerably more for a first impression than low-value advertisers. When all advertisers are charged the same per impression price, high-value advertisers will purchase ads for both periods to pick up the added first exposures, even though the audience purchased will include some individuals who have seen their ads before. Low- value advertisers will purchase ads for only one period to avoid paying for ‘wasted’ impressions. High-value advertisers are thus paying a higher effective price for first impressions and there are distributions of advertisers’ first impression valuations for which profits from this implicit form of price discrimination can be higher than what can be achieved through tracking. This could be at least part of the explanation for why a considerable portion of Internet display ads are still sold on an untargeted basis.

20.4