Context information increases revenue in ad auctions: Evidence from a policy change
CContext information increases revenue in ad auctions:Evidence from a policy change
Sıla Ada
WU Vienna [email protected]
Nadia Abou Nabout
WU Vienna [email protected]
Elea McDonnell Feit
Drexel University [email protected]
December 3, 2020
Working PaperPlease do not cite or distribute without permission. a r X i v : . [ ec on . GN ] D ec ontext information increases revenue in ad auctions:Evidence from a policy change Abstract
Ad exchanges, i.e., platforms where real-time auctions for ad impressions take place, havedeveloped sophisticated technology and data ecosystems to allow advertisers to target users ,yet advertisers may not know which sites their ads appear on, i.e., the ad context. In practice,ad exchanges can require publishers to provide accurate ad placement information to ad buyers prior to submitting their bids, allowing them to adjust their bids for ads at specific domains,subdomains or URLs. However, ad exchanges have historically been reluctant to disclose place-ment information due to fears that buyers will start buying ads only on the most desirablesites leaving inventory on other sites unsold and lowering average revenue. This paper exploresthe empirical effect of ad placement disclosure using a unique data set describing a change incontext information provided by a major private European ad exchange. Analyzing this as aquasi-experiment using diff-in-diff, we find that average revenue per impression rose when morecontext information was provided. This shows that ad context information is important to adbuyers and that providing more context information will not lead to deconflation. The exceptionto this are sites which had a low number of buyers prior to the policy change; consistent withtheory, these sites with thin markets do not show a rise in prices. Our analysis adds evidencethat ad exchanges with reputable publishers, particularly smaller volume, high quality sites,should provide ad buyers with site placement information, which can be done at almost no cost.
Keywords:
Online display advertising, Real-time bidding, Advertising auctions, Informationdisclosure, Ad context, Conflation, Bundling Introduction
Digital display advertising has rapidly become popular among advertisers due to new targetingoptions as well as lower transaction costs enabled by ad tech. A central component of the industryis real-time bidding (RTB) markets for ad impressions (Figure 1). When a user requests a pagefrom a website, an impression (i.e., an opportunity to advertise to that user) becomes available.The site publisher can sell this impression on an RTB market by submitting a bid request to thead exchange, which often includes a cookie ID identifying the user. The ad exchange subsequentlybroadcasts the bid request to potential ad buyers typically through intermediaries called demandside platforms (DSPs.) In response, ad buyers submit bids for the impression and the exchangesells the impression typically in a second-price auction. This entire process occurs within 400msso that the ad loads almost instantaneously for the user. The publisher is paid the winning priceless a commission to the exchange. An alternative to RTB is programmatic direct advertising,which allows ad buyers to pre-negotiate a guaranteed number of impressions at a particular siteand are often processed using the same technology infrastructure as RTB. US programmatic digitaldisplay ad spending (including RTB and programmatic direct) is projected to reach $ Ad ExchangeWebsite PublisherUser(cookie id) Ad Buyerpage request bid request bid context of the impression over and above the information they already haveabout users (Another 5% included a subdomain and the remaining 80% included the fullURL.) When the high-level domain is included in the bid request, it often does not reflect the specificsite where the ad will appear (e.g., when nbc.universal.com is listed as the domain in the bid requestfor any NBCU-owned website, even if the ad will appear at a different domain). To make mattersworse, on the open exchanges some publishers provide outright fraudulent URLs in a practiceknown as “domain spoofing” (Sluis 2020). Even though the authorized digital sellers (ads.txt)protocol from the IAB (Interactive Advertising Bureau) was touted to increase transparency andreduce fraud, publishers providing outright fraudulent URLs in the ads.txt file remains an issue(Fou 2020). Thus, the site where an ad will appear is far from transparent to advertisers buyingon the open exchanges today. Our study suggests that prices will rise when domain information ismore transparent to ad buyers, benefiting ad buyers, reliable publishers, and exchanges.It was not obvious a priori that context information would affect ad buyers’ valuations for ads.When RTB was introduced to display advertising, the industry touted their capability to allowad buyers to target specific users no matter what sites they were viewing (e.g., retargeting) andthe industry has largely dismissed ad context as a poor proxy for the richer user targeting options This data was provided by personal communication with a large meta exchange that aggregates about 400-600billion bid requests each day across more than 150 supply-side platforms.
The New York Times have moved to protect users bydecreasing the amount of user information collected on their sites (Berjon 2020) and some publishers5ave moved to eliminate user tracking all together (Edelman 2020). Thus, user targeting will be lessstraightforward in the future and industry experts predict that contextual targeting will becomemore relevant (Schiff 2019, Tan 2019). This study provides direct, empirical evidence that contextinformation is valuable to ad buyers.In the next section, we briefly review findings on information disclosure from the theoreticalliterature on auctions, which shows that prices may go up or down when more information isprovided to all bidders, depending on market thickness and the heterogeneity in bidders’ valuations.In the following section, we describe the institutional setting and policy change in more detail. Wethen analyze the empirical effect of information disclosure on ad prices using a diff-in-diff analysis,which shows that prices rose on average. Mechanism checks show that competition at each sitedecreased, but the right tail of the distribution of winning bids increased (i.e., winning bids startedto spread out more) and nearly all sites saw an increase in revenue per impression. Followingthat, we compare the behavior of one buyer that received early access to the context information,compared to a synthetic control made of up of buyers without this information and find that thisbuyer won more auctions. This provides convergent evidence that ad buyers bid higher when theyhave context information for each impression, and markets thin-out a bit, but remain sufficientlythick for prices to rise. We then proceed with evidence for heterogeneous treatment effects relatedto 1) competition and 2) the size and quality of a site and conclude with a summary of our findingsand a discussion of our study’s implications.
Theoretical predictions for how information disclosure affects ad auctions are mixed, with someresearchers arguing that ad prices achieved in the auction should go up when more informationis available about each impression (Hummel and McAfee 2016) while others argue they should godown (Levin and Milgrom 2010). These predictions depend critically on 1) the valuations of adbuyers for sites and 2) the number of bidders with positive valuations for each impression after theinformation is disclosed. If ad buyers all prefer the same sites, then prices will rise for the desirablesites and fall for the others. However, prices can rise for all sites if each ad buyer prefers differentsites, i.e., buyers have heterogeneous preferences (Hummel and McAfee 2016), and markets are6hick enough.For example, consider an ad exchange where there are two sites and bidder i ’s valuation for animpression j at site k is composed of the value for the site s ik plus their value for remaining featuresof the impression r ij such as the cookie, time-of-day, etc. When context information is disclosedto all bidders, their bids are based on their valuation for each site. When context information iswithheld, bidders are forced to base their bids on their average site valuation. For simplicity in thisexample, we assume that the bidders know that the two sites place equal numbers of impressionsin the auction, thus the valuations are: v ijk = r ij + s ik if the site is disclosed r ij + ( s i + s i ) / r ij ∼ N (0 , ω ) s i ∼ N ( µ − δ/ , σ ) s i ∼ N ( µ + δ/ , σ )Here, ω represents variation in bidder’s values for individual impressions, δ is the difference inaverage valuation between the two sites, and σ represents the variation across bidders in valuationsfor sites.When σ is large and δ small, we have what Tadelis and Zettelmeyer (2015) refer to as “hetero-geneous bidders” or “horizontal differentiation between sites”. When there are a sufficient number( N ) of heterogeneous bidders, disclosure will increase auction prices for both sites because bid-ders bid more for the sites they uniquely prefer (Palfrey 1983, Chakraborty 1999, Hummel andMcAfee 2016, Chen et al. 2018). Figure 2a illustrates this scenario showing simulated selling pricesfor 1,000 impressions where there are twenty-five bidders with heterogeneous preferences for sites( δ = 0 , σ = 1). As can be seen from the horizontal lines in Figure 2a, average prices rise for bothsites, although prices may rise more for one site than the other, depending on the realization of s i and s i . Under context disclosure, the distribution of winning prices also has greater varianceand a longer tail. As we show in the next section, the auction outcomes in the exchange we studyare consistent with this scenario. The potential for prices to rise across the board when ad buyers7igure 2: Effect of information disclosure on auction prices. Each dot represents the selling price ofan impression in a simulated second-price auction where valuations are defined as in Equation (1).Average selling price is shown with horizontal lines. Panel (a) shows that when context informationis disclosed to a large number of heterogeneous bidders, average price rises for both sites. Panel (b)shows that when there are only two bidders, disclosure causes prices to fall slightly. Panel (c) showsthat when bidders homogeneously prefer site 2, prices fall for site 1 and rise for site 2. Panel (d)shows that when the residual value of the impression (due to cookie, time-of-day, etc.) dominatesthe site value, the context disclosure has little effect on the distribution of winning prices.(a) Site 1 Site2 Site 1 Site2
Without Disclosure With DisclosureHeterogeneous Bidders, Thick MarketsN = m = −0.5 s = d = w = P r i c e ( H i ghe s t B i d ) . . . . . (b) Site 1 Site2 Site 1 Site2
Without Disclosure With DisclosureHeterogeneous Bidders, Thin MarketsN = m = s = d = w = P r i c e ( H i ghe s t B i d ) . . . . . . (c) Site 1 Site2 Site 1 Site2
Without Disclosure With DisclosureHomogeneous Bidders, Cherry PickingN = m = s = d = w = P r i c e ( H i ghe s t B i d ) . . . . . (d) Site 1 Site2 Site 1 Site2
Without Disclosure With DisclosureCookie Values DominateN = m = −0.5 s = d = w = P r i c e ( H i ghe s t B i d ) . . . . . . . σ large and δ small), ifthere are an insufficient number of bidders, prices may fall. This is illustrated in Figure 2b. whichshows that average prices fall under information disclosure when there are N = 2 bidders. Becausethere are so few bidders, some impressions have a second-highest bid of zero, shown by the largenumber of points with prices at zero for Site 2. (This does not happen for Site 1 because realizationsof s and s are both high by chance.) Levin and Milgrom (2010) describe phenomena like thisin the context of user targeting and refer to it as an “orphaned category”. They argue that adplatforms should conflate markets so that “similar but distinct products are treated as identicalin order to make markets thick or reduce cherry-picking.” Not disclosing site information is onestrategy for conflation, when markets are thin.Thus, Figures 2a and 2b show prices may rise or fall when relevant information is disclosedto bidders in an auction. For a given set of bidders and preferences, the literature on ad auctionsconcludes that the relationship between auction outcomes and information disclosed is concave withan intermediate amount of information (or equivalently bundling) producing the highest revenue(Rafieian and Yoganarasimhan 2020).Context disclosure can lead to winners and losers among sites when ad buyers have homoge-neous preferences for sites. We can simulate this scenario by setting δ higher and σ lower. AsFigure 2c shows, when bidders have homogeneous preferences for one site over the other, prices risedramatically for the preferred site and fall for the other. This outcome may be revenue-neutral forthe auction platform (depending on the mix of sites it represents), but it has a substantial effect9n revenue for individual sites and may make publishers of less-preferred sites dissatisfied with theplatform. The potential for this type of “cherry picking” by ad buyers was a serious concern to theplatform operator we study. However, the data from before and after the policy change suggestthat ad buyers preferences for sites are largely heterogeneous (with some limitations, which we willdiscuss).Finally, because it is possible that ad buyers do not value context information above-and-beyondcookie information, in Figure 2d, we simulate a case where bidders show more variation in theirvalue for individual impressions and their cookies versus sites (i.e., σ is small and ω is large). Asthe graph shows, when bidders place more importance on the cookie and other features of theimpression than the context, then context disclosure has a modest effect on ad markets.Collectively, the simulations in Figure 2 show that context disclosure may 1) uniformly raiseprices for all sites, 2) uniformly lower prices, or 3) raise prices for some sites but not others dependingon the number of bidders and their valuations for sites. It was also possible ex ante that this policychange would have no effect at all because ad buyers may not behave rationally. The theoretical andstructural literature on auctions relies on the assumption that bidders will maximize their expectedvalue given available information. However, research on managerial decision making shows thatmanagers are often risk-averse (Amihud and Lev 1981) relying merely on historical performancepatterns (Busenitz and Barney 1997, Little 1970), such that they do not change their investmentdecisions when receiving better information (Lambert et al. 2007). Given the many potentialtargeting options available to online ad buyers, they may not have the time or incentive to adjusttheir bidding strategies for each site. Consequently, it is unclear whether ad buyers will put contextinformation to use at all.To summarize, it is difficult to predict whether ad context information will affect auction out-comes for three reasons: 1) if site placement is not valued by ad buyers then the change will haveno effect, 2) even if buyers value the context, they may not change their bidding strategy due tothe complexity of the advertising environment, and 3) even when ad buyers are behaving optimally,the effect of information disclosure on auction outcomes is a complex function of buyer valuationsand market thickness, and prices may fall or rise for particular sites or overall. Thus, it remainsan empirical question how context information will affect RTB ad auction outcomes. Next, wedescribe the institutional setting where we study the effect of a change in context disclosure.10 Institutional setting
This paper investigates a change in the ad context disclosure policy at a major private ad exchangein Europe. In a private exchange, a relatively small number of digital publishers agree to offerimpressions to a pre-approved group of ad buyers through RTB. This makes private exchangesdistinct from open
RTB exchanges like Google Ad Manager (formerly DoubleClick) where any adbuyer or publisher may participate and thousands do. While RTB began with the open exchanges,as concerns about transparency, fraud, and brand safety have grown, premium publishers includingHearst, Technorati, Conde Nast, CBS, NBCUniversal, IDG TechNetwork, The Weather Channel,and Vox have created private exchanges, which offer ads at a smaller set of reputable websites.Participating ad buyers and sites are vetted prior to bidding and the relatively small number ofparticipants increases transparency and brand safety for both ad buyers and publishers. Softwareplatforms for running ad exchanges such as Google AdMeld make it easy for small groups ofpublishers to build the necessary infrastructure to run an exchange. Sales on private ad exchangesare expected to exceed those on open exchanges in 2020 (Fisher 2020).
This private exchange offers us a unique opportunity to study the effect of a change in contextinformation provided to ad buyers. Prior to April 2016, buyers (including advertisers themselvesand intermediaries acting on behalf of multiple advertisers) on the exchange we study purchased adswithout any knowledge of where the impression would appear. The only form of context-targetingavailable to ad buyers was buying ads on a “channel”, where channels represented broad contentcategories like, “news,” “automotive,” or “finance.” Many ad buyers chose to place their ads on“run of network,” meaning their ads may appear on any of the participating publishers’ websites,while still following any user targeting criteria the buyer has established.In April 2016, a single large buyer was given access to context information in the bid request.Specifically, this buyer was given information about the subdomain where the ad would appear,e.g., nytimes.com/business. Throughout the analysis we use the term “site” to refer to thesesubdomains, recognizing that some of these “sites” are subdomains belonging to the same domain.In May 2016, after observing an aggregate rise in revenue when one buyer had site information,11he exchange made the site for each impression available to all bidders. When the policy changed,ad buyers were notified by the ad exchange, often through personal phone calls. The specifics ofhow users determined bids specifically for sites varied by demand side platform (DSP); Figure A.1in the Appendix shows an example of how buyers were able to restrict their bids to particularsubdomains in their bidding criteria. In addition to whitelisting or blacklisting sites, they couldalso use programmatic strategies to bid differently for different sites.The private exchange provides us with a well-controlled setting to study the effect of contextinformation on ad buyers’ valuations for ads. The participating publishers contractually agreedto sell all their digital ad inventory exclusively through this exchange. Publishers could choosebetween RTB and programmatic direct sales and we observe all sales in both formats. Before thechange, ad buyers knew that their ads would appear on one of the participating reputable sites,and not “anywhere” as in open exchanges. Fraud by the publishers is not an issue in this setting.Finally, the policy change happened all-at-once, buyers could easily manage their buys to targetspecific sites, and all sales are fully-observed.During this entire period, which was prior to the General Data Protection Policy (GDPR)coming into effect, ad buyers had access to the user cookie id for each impression and there was anactive market for third-party data on past cookie behavior, so the policy change gives us insightinto how ad buyers value context information over-and-above the rich user behavior data availableat the time.The exchange hoped that this change toward greater transparency would make the exchangeas a whole even more appealing to ad buyers and earn them certification as a brand-safe plat-form. However, the exchange had lengthy debates internally about the change. While some at thecompany were confident that revenue would rise for most or all sites, others were concerned thatcherry-picking of the most desirable sites would lower revenues for less desirable sites, and poten-tially overall revenue for the platform (which gets a 2.5% commission on all sales). As discussedin the previous section, these outcomes depend theoretically on the distribution of preferences forsites among the buyers, which was ex ante unobservable to the exchange.12 .2 Participating websites
As discussed in the previous section, the theoretical effect of context disclosure in an ad auctiondepends critically on how the buyers value the websites where ads appear. The participating sitesvary substantially in the types of content they provide and include one of the top 3 news sitesin the country (according to SimilarWeb), one of the top 3 sports sites, and a variety of specialinterest and community sites similar to quora.com, webmd.com, allrecipes.com or zillow.com in theUS. However, while some of the sites might be considered niche content, all of them are reputableand none would be considered extremist content or “clickbait” (as you might find in the openexchanges).The diff-in-diff analysis focuses on change in average revenue per impression for the 57 sitesthat participated in the market in both 2015 and 2016. To characterize the cross-sectional variationbetween sites, Table 1 shows summary statistics on the supply of impressions and average revenueper impression for these sites during one week before the policy change. Across all sites, the meanrevenue per thousand impressions (CPM) was e e e Our agreement with the exchange precludes us from naming the sites or the country they operate in. .
88 0 .
39 0 .
27 0 .
81 2 . .
66 37 .
79 7 .
57 50 .
57 159 . Our goal is to identify how the policy change affected the average selling prices for impressions, i.e.,the revenue per impression. As discussed in Section 2, if the buyers have heterogeneous preferencesfor the sites and there are a sufficient number of buyers for each site, then prices should rise overall.If there are sites that are undesirable to most buyers, such that the markets thin out, then pricesmay fall for some sites or overall. It is also possible that the market might not be affected at all,if buyers don’t value the ad context, or if the transaction costs of customizing bids to specific sitesare too high, or if buyers do not behave rationally.
As an initial investigation of the data, Figure 3a plots the overall daily average revenue per impres-sion (reported in e per thousand impressions, CPM) before and after the policy change. Figure3b plots the total supply of impressions (sold and unsold) over the same period. The day whenone buyer was given access to the subdomain for each impression in 2016 is indicated by the firstvertical red dotted line and the day when all buyers were granted this context information is indi-cated by the second vertical dotted red line. Figure 3a shows that the average price per impression Note that we do not have access to bids or selling prices for individual impressions. e e e e More importantly, the full disclosure of context information coincides with the Spring andSummer where there is a lower supply of impressions and an increase in prices in both years.This observed decline in supply of impressions (and increase in prices) is consistent with seasonalpatterns in web traffic, which tend to decline in the Summer. Thus, the policy change is confoundedby the seasonal decrease in supply. In the next section, we do a formal diff-in-diff analysis withadditional controls to account for this.
To show that the policy change increased revenue per impression for the platform, we regress theaverage revenue per impression for each site in each week on a 0-1 indicator for the policy change,a 0-1 indicator for the year and the interaction between those two, following the standard diff-in-diff approach. This regression is estimated using weekly data from January to July for both 2015and 2016 and includes additional controls and site fixed effects. Thus our estimate of the effect ofthe policy change is the change in revenue per impression observed in 2016 above-and-beyond thechange in revenue observed during the same period in 2015, i.e., the coefficient on the interactionbetween year 2016 and the policy change.While the raw data is at the daily level, we summarized it at the weekly level to avoid havingperiods where a site did not sell any impressions. The dependant variable (average weekly revenueper impression for each website) is computed by dividing the total weekly revenue for each site bythe total impressions that the site submitted to the RTB platform in that week including sold andunsold impressions. As our estimate of the treatment effect relies on a comparison of both years,we include only the 57 sites that sold impressions in both 2015 and 2016. This results in 3,058 As a robustness check, we re-do the diff-in-diff analysis excluding these observations and find that our substantiveconclusions are unchanged. e ) and supply of impressions (millions)from January to July for 2015 and 2016. The first (second) red dashed line indicates the change topartial (full) disclosure. (a) Weekly average revenue per impressions(b) Weekly supply of impressions16ite-week-year level observations, which is slightly fewer than 57 sites ×
27 weeks × Theinteraction terms
Partial disclosure x Year16 and
Full disclosure x Year16 are the key coefficientsof interest, which show that prices increased by 10.8 EUR cents per thousand impressions duringthe period when one buyer had access to the subdomain information for each impression and by15.4 EUR cents when all buyers had access to the subdomain information. This increase of 15.4EUR cents for full disclosure is above-and-beyond the increase in average revenue per impressionseen for the same months in 2015 (see the coefficient for
Full Disclosure ) and the other controls.This substantial increase in prices represents the effect of moving this auction from “channel” levelcontext disclosure to subdomain level disclosure. For those interested in how the treatment effect developed over time, we report effects separately for each monthin the Appendix. Those results indicate that there was no “learning” period for participants in the exchange: Theobserved treatment effect seems to have set in immediately. e per thousand) dueto context disclosure. (1) (2) (3)Average revenue Without controls Placebo testper impression(CPM in e ) Effect of policy change
Partial disclosure x Year16 0 . ∗∗∗ . ∗∗∗ (0 . . . ∗∗∗ . ∗∗∗ (0 . . − . . Baselines
Constant 0 . ∗∗∗ . ∗∗∗ . ∗∗∗ (0 . . . − .
006 0 . . . . ∗∗∗ . ∗∗∗ (0 . . . . . ∗∗ . ∗∗∗ . ∗∗∗ (0 . . . Controls
Supply of impressions − . ∗∗ − . ∗∗∗ (millions) (0 . . . − . ∗∗∗ (pre-treatment) (0 . . . − . . . Notes: Standard errors, in parentheses, are clustered at the week level. ∗∗∗
Significant at the1 percent level. ∗∗ Significant at the 5 percent level. ∗ Significant at the 10 percent level. Arobustness check filtering out calendar week 6, 2015 when prices spiked showed substantivelysimilar results. not due to ad buyers shifting budgetsinto the RTB market
Prior to the policy change, ad buyers could not buy impressions at specific sites through the RTBmarket, but they could buy impressions at a specific site by making a programmatic direct dealwith a specific site. These programmatic direct deals are often more expensive than prevailingTable 4: Average monthly impressions (in millions) sold through RTB and programmatic direct in2016 before and after the policy changeNo disclosure Partial disclosure Full disclosureImp. % Imp. % Imp. %Programmatic direct 1 , . .
4% 1 , . .
4% 1 , . . . .
3% 744 . .
6% 741 . . . .
3% 19 . .
0% 40 . . Theoretically, information disclosure increases prices in an auction when there are a sufficientnumber of bidders who have heterogeneous preferences for the items. In this section, we provideadditional evidence that this is the mechanism at play when context information was disclosed inthis private exchange. Specifically, we show that 1) the market for each site was sufficiently thickafter the policy change, 2) buyers appear to have been bidding higher for their preferred sites, and3) most individual sites experienced an increase in revenue.
A necessary condition for prices to rise in the auction is that the number of bidders for eachimpression does not fall too low. To provide evidence that markets do not thin out, we investigatehow the policy change affected the daily buyers for each site (i.e., the average number of winningbidders each day). Since we do not observe the individual bids, we can not say how many bidderswere bidding on each impression, but the average number of daily buyers for each site gives us aproxy for the number of ad buyers who were submitting competitive bids for impressions at a givensite. (The average daily buyers prior to the policy change was used as a pre-treatment control21ariable in the diff-in-diff analysis; here we look at whether the average daily buyers changed inresponse to the policy change.)Figure 4: Competitiveness of the auctions for each site as measured by the average daily buyersbefore and after the policy change (in 2016).
Note: Sites are sorted in order of weekly supply of impressions in the auction (highest to lowest).
Figure 4 plots the average number of daily buyers for each site before and after the policychange and shows that there was little change in the number of buyers for each site. Thus themarket for impressions at each site remained similarly competitive after the policy change. If atall, sites with thick markets seem to become slightly thicker, while sites with thin markets becomeslightly thinner. Yet, even the site with the fewest average daily buyers has an average of morethan 8 buyers each day after the policy change. Therefore, the data suggests that most marketswere sufficiently thick for prices to rise.To provide additional evidence that markets remain thick after the policy change, we fit aregression for the total average daily buyers at each site as a function of the year and the policychange, using the same controls and site fixed effects as used in our diff-in-diff analysis. Table5 shows a small, insignificant decrease in the average daily buyers when one buyer had access tothe context information for each impression (
Partial disclosure x Year16 ). In the period when allad buyers had access to this information, there is a slightly higher, but still insignificant decrease22able 5: Diff-in-diff analysis of the change in average daily buyers.Average daily buyers
Effect of policy change
Partial disclosure x Year16 − .
376 (2 . − .
932 (7 . Baselines
Constant 43 . ∗∗∗ (7 . .
923 (1 . . ∗∗∗ (6 . . ∗∗∗ (2 . Controls
Supply of impressions (millions) − .
028 (0 . .
024 (0 . Notes: Standard errors, in parentheses, are clustered at the week level. ∗∗∗
Significant at the 1 percent level. ∗∗ Significant at the 5 percent level. ∗ Significant at the 10 percent level. in the average daily buyers by -8.932 (
Full disclosure x Year16 ). Given the baseline of about 45average daily buyers for each site, the number of buyers for each site decreased by about 18% wheneveryone has context information, which is consistent with buyers bidding higher for the sites theyeach prefer leading to fewer average daily buyers for each site. However, markets remain sufficientlythick so that deconflation is not a concern.
If buyers have heterogeneous preferences and are bidding more for their preferred sites when theyhave context information, then we should also see the right tail of the distribution of winning bidsincrease, consistent with the simulation results in Figure 2a. Indeed, this is the case. Figure 5 plotsthe distribution of average prices paid by each buyer for each site and shows a distinct increase inthe proportion of impressions selling for e If preferences are heterogeneous and markets remain thick, then most sites should see a rise inprices. Figure 6 plots the estimated effect of full context disclosure for individual sites and showsthat revenue per impression rose for the majority of sites. These site-specific estimates are basedon a regression with the same specification as our main diff-in-diff analysis in Table 2, except thatsites are interacted with the treatment indicators. There are few “orphaned” sites; only one siteshows a significant drop in revenue per impression. The effect of context disclosure for most sites iseither neutral or positive, with a few sites that gain substantially. This is consistent with some adbuyers having strong preferences for impressions at a particular site (above the information theyalready had about the user from the cookie.)Taken together, Sections 5.1, 5.2, and 5.3 paint a picture that is consistent with the changesexpected for information disclosure in a competitive market where buyers have heterogeneous pref-erences. If each buyer is raising their bids for a different subset of sites, then average daily buyersfor each site should fall slightly (as shown in Table 5), the winning bids should have a longer righttail (Figure 5), and prices should rise for most sites (Figure 6).24igure 6: Estimated change in revenue for individual sites, with 95% confidence intervals.
Note: The estimated change in revenue for an individual site is often estimated very imprecisely and in some casesthe 95% confidence intervals run off the plot.
As a final mechanism check, we investigate the effect of the policy change for the period where onlythe buyer who used a particular DSP was provided with ad placement information (April 2016).Theoretical research on information disclosure and bundling has focused on the cases like thoseillustrated in Figure 2 where all bidders have access to the same information and product offerings(Milgrom and Weber 1982, Eaton 2005, Tadelis and Zettelmeyer 2015, Hummel and McAfee 2016).However, during a one-month period, the auction platform initially provided site information to onebuyer only. To understand the expected effect of this partial disclosure, we make a brief departureto review another simulation showing the effect of disclosure to a single bidder.Specifically, we assume that one buyer has the site information and will bid their valuation underdisclosure, while the other twenty-four buyers bid their valuation without disclosure (see equation1). The simulation otherwise follows the assumptions in Section 2. Figure 7 shows simulatedwinning prices for no disclosure versus partial disclosure with the auctions won by the first biddercolored red. In this scenario, when information is disclosed to just one bidder, that buyer’s bids are25igher for their preferred site, resulting in the treated bidder winning more often. In this example,the treated bidder wins 6.3% of impressions without disclosure and 7.2% when they have contextinformation. The amount by which the treated bidder wins more depends on the bidders individualpreferences for sites; in this simulation, the treated bidder had a fairly high valuation for Site 2.However, the average winning prices do not necessarily change substantially under partial dis-closure; whether that bidder pays more or less depends on the mix of sites that the treated buyerpurchases, which depends on the valuations of all the bidders. Thus, when there is a large numberof bidders who are heterogeneous in their valuations, partial disclosure theoretically results in the treated bidder winning more .Figure 7: Effect of information disclosure to a single bidder on auction prices. Each dot representsthe outcome of a simulated auction and is colored red if the auction was won by the bidder whoreceives the additional information. Average resulting price is shown with horizontal lines. lll lllll lllllll ll lll lll llll ll llllll ll ll ll llllllll lll lll lll lll lll ll lllll ll l ll ll ll ll l ll lll lll llll ll ll ll ll ll ll ll ll ll lll ll ll lllll lll lll ll lll ll llllll ll lll ll ll llllll llll l ll llll ll ll l lll ll lll lll lll ll l lll lll ll l l lllll lll l ll ll l ll ll ll lll lll ll l ll ll ll llll ll ll lll llll lll l ll llll lll ll l lll lll lllll ll ll ll ll llllll llll l lll l l lll ll llll llll lll lll ll l ll lll
Site 1 Site2 Site 1 Site2
Without Disclosure With DisclosureDisclosure to 1 BidderN = m = −0.5 s = d = w = P r i c e ( H i ghe s t B i d ) . . . . To understand what happened in practice at the private exchange, we analyze the behavior ofthe bidder who obtained exclusive access to site placement information in April 2016 relative to26ther bidders. This buyer was a DSP bidding on behalf of several advertisers. Prior to the policychange, the treated bidder paid higher prices than the average of all other bidders (compare thefirst column in Table 6 to Column 4).To construct a counterfactual for what would have happened if this buyer had not gottenplacement information, we use a synthetic control analysis (Abadie and Gardeazabal 2003, Abadieet al. 2010) to construct a counterfactual buyer that resembles the treated buyer during the pre-intervention period. The counterfactual synthetic control is a convex combination of untreatedbuyers that matches as closely as possible on several pre-treatment covariates known as “predictors”in the synthetic control literature. The weights that define the control buyer are chosen such thatthe counterfactual buyer’s predictors approximate the treated buyer’s during the period prior tothe policy change, week-by-week. Then, the constructed synthetic buyer is used to estimate acausal counterfactual for how the treated buyer would have behaved if not provided with placementinformation. We analyze both the winning price and impressions won as dependent variables andconstruct separate synthetic controls for each. (Note that average winning price is the same asaverage revenue per impression but at the buyer-week level.) Technically, the impressions wonand prices paid by buyers who did not have access to this information may have been affectedsomewhat since they are participating in the auction with the treated bidder. However, since ourtreated buyer represents less than 2% of impressions sold, the control buyers would have only beenaffected by a small amount.The synthetic control is matched on the following predictors: (1) number of impressions wonin each genre in each pre-treatment week, (2) average of price paid in each genre in each week,(3) total number of impressions won in each week and (4) average price paid in each week. Gen-res were utilized to create the predictors, since ad buyers were able to target channels or usersbased on behavioral information prior to the policy change. The covariates are created based onFebruary-July in 2015, as well as February-March 2016. In constructing the synthetic control, dailyobservations with substantially higher prices (e.g., e
10 CPM) or lower volumes ( <
500 impressionsin a day) were filtered out. These unusual observations are likely due to highly targeted buys thatare not representative of the types of prices paid by the treated buyer. They represent only 0.8% ofimpressions. The core identifying assumption is that these pre-treatment covariates represent thekey ways in which the treated buyer is different than the untreated.27able 6 reports the summary of the covariates used in the construction of the synthetic buyerand compares them to the treated buyer, which are by construction largely similar. Furthermore,buyers in the control group that are picked by the algorithm are mainly the same for both inde-pendent variables and the highest weights are assigned to other intermediaries bidding on behalfof multiple advertisers (similar to the treated buyer).Table 6: Descriptive statistics for pre-treatment behaviors used to construct the synthetic control(average of predictors over pre-treatment weeks).Synthetic Synthetic MeanTreated Control Control Other Buyersfor Price for ImpressionsImpressions purchasedOverall 1,126,804 1,412,513 1,366,452 1,176,408on Community & Forums 70,926 49,306 41,962 43,349on General interest 992,351 1,266,406 1,231,817 1,065,867on Health 4,147 5,328 4,409 4,834on Special interest 3,603 1,714 1,687 6,771on Sports 55,778 89,759 86,577 55,587Price paidOverall 3.46 3.22 3.05 0.40on Community & Forums 3.86 3.62 3.09 0.33on General interest 3.43 2.59 2.83 0.35on Health 2.30 2.39 1.44 0.14on Special interest 1.14 0.94 0.75 0.13on Sports 3.69 3.42 2.96 0.32Figure 8a shows that the trajectory of the synthetic buyer’s average winning price closelyfollows the treated buyer’s price, which suggests that the synthetic buyer nicely mimics the treatedbuyer prior to policy change. Consistent with the simulation in Figure 7, the additional placementinformation does not affect the average winning price for the treated bidder, as can be seen bycomparing the treated bidder to the synthetic control in Figure 8a after the policy change.However, the simulation in Figure 7 suggests that the number of impressions won by the treatedbidder should be higher when provided with placement information, so we compare the numberof impressions won for the treated and synthetic control bidders in Figure 8b. To the left of thevertical line, the number of impressions won is similar for the treated and synthetic control bidder The weight associated with the predictor “number of impressions won on Community & Forums sites” is verysmall, which indicates that it does not have predicting power for either dependent variables. e )(b) Number of impressions won Note: Pre-treatment time series includes February - July 2015 and February - March 2016 and is thus discontinuousin time.
Up to this point, the analysis has focused on how providing context transparency affected theprices for impressions on average. The analysis suggests that buyers value context information andbuyers have heterogeneous preferences for sites, which leads to revenue gains for all sites. Nextwe turn to the practical question of whether certain types of sites benefited more from this policychange. First, we show that sites with more buyers prior to the policy change, i.e., those withthicker markets, saw a greater increase in revenue after the policy change. Second, we show thatsmaller volume, high quality sites benefited the most from the policy change.
Theoretically, the effect of context disclosure on site placement should be moderated by the compet-itiveness of the market (see Figures 2a and 2b). This motivates an investigation of heterogeneoustreatment effects across sites with stronger or weaker competition prior to the policy change. Figure4 shows the average daily buyers for each site prior to the policy change. From this we create adummy variable for sites that had fewer buyers (thin markets) prior to the policy change. Weset the cutoff point at the first quartile, which is 28 average daily buyers. This variable capturesmarket thinness before the policy change, and we assume that sites with thinner markets before thepolicy change were likely to have thinner markets after the change. Figure 4 indicates that this is areasonable assumption since market competitiveness was similar before and after treatment. Thispre-treatment covariate is also not contaminated by the treatment. Based on our simulations in30ection 2 and the literature on auctions, we expect sites with thick markets to experience a greaterincrease in average revenue per impression, because it is more likely that there are several buyerswho will value those impressions more when provided with context information.Table 7: Heterogeneous treatment effects for sites with thin markets - Diff-in-diff analysis of thechange in average revenue per impression ( e per thousand) due to context disclosure. Average revenueper impression
Treatment effects
Partial disclosure x Year16 0 . ∗∗∗ (0 . .
028 (0 . . ∗∗∗ (0 . − .
150 (0 . Baselines
Constant 1 . ∗∗∗ (0 . − .
006 (0 . . ∗∗∗ (0 . . ∗∗ (0 . − .
110 (0 . . ∗∗ (0 . .
069 (0 . − . ∗∗∗ (0 . Controls
Supply in millions − . ∗∗ (0 . − . ∗∗∗ (0 . . . ∗∗∗ Significant at the 1 percent level. ∗∗ Significant at the 5percent level. ∗ Significant at the 10 percent level. As a robustnesscheck these regressions were also run excluding the week in 2015when prices spiked (calendar week 6, 2015) and substantive resultsremain the same.
The model reported in Table 7 shows the moderating effect of competition on average revenueper impression. The estimated increase in revenue per impression was 15.4 EUR cents for siteswith thick markets (see
Full disclosure x Year16 ), while it was nearly zero for thin markets. Theestimated effect for sites with thin markets is the sum of
Full disclosure x Year16 and
Full disclosure Note that sites with more buyers tend to sell more impressions, thus the effects sizes for thick markets here arevery close the the overall averages from the volume-weighted regression reported in Table 2. x Year16 x Thin which is 15.4 - 15.0 = 0.4 EUR cents. The finding that prices did not increasefor sites with thinner markets is consistent with the literature on auctions and the simulations inSection 2, thus serving as an additional mechanism check. For ease of interpretation, we plot theestimated effects for sites with thin versus thick markets in Figure 9.
For our final analysis, we look at heterogeneous treatment effects for sites of different quality andsize. While slightly less theoretically motivated, it gives us an answer to the practical questionof “Which sites benefit the most from context transparency?” Even though we study a privateexchange with generally brand-safe sites, not all of them are – what advertisers may consider –the highest-quality advertising outlets. To categorize sites according to their quality we askedthree industry experts (a head of media planning, a media planner, and a trader in RTB auctions)to classify the sites into those that provide premium and non-premium advertising environments,which is a common industry categorization. All experts were familiar with the sites and hadpurchased media from the ad exchange in the past. In addition, we categorized sites into smallversus large based on the number of impressions they provided to the RTB market. That leads Sites are categorized as “large” if their daily supply of impressions on the RTB platform was larger than 80,000 inMarch 2016, otherwise they are coded as “small.” Considering the distribution of daily average supply of impressions,80,000 daily impressions per site is a clear cut-off point which separates the sites with high and low supply of
32o four categories of sites: Premium large, Premium small, Non-premium large, and Non-premiumsmall.Table 8: Heterogeneous treatment effects for high quality and large sites - Diff-in-diff analysis ofthe change in average revenue per impression ( e per thousand) due to context disclosure. Average revenue Std.per impression Error
Treatment effects
Partial disclosure x Year16 0 . ∗∗∗ (0.032)Partial disclosure x Year16 x Non-Premium+Large − .
023 (0.037)Partial disclosure x Year16 x Non-Premium+Small − .
014 (0.069)Partial disclosure x Year16 x Premium+Small − .
079 (0.072)Full disclosure x Year16 0 . ∗∗∗ (0.044)Full disclosure x Year16 x Non-Premium+Large − .
021 (0.063)Full disclosure x Year16 x Non-Premium+Small 0 . ∗ (0.091)Full disclosure x Year16 x Premium+Small 0 . ∗∗ (0.149) Baselines
Constant − . ∗∗∗ (0.321)Partial disclosure 0 .
003 (0.019)Full disclosure 0 . ∗∗∗ (0.038)Year16 0 . ∗∗∗ (0.028)Non-Premium+Large 0 . ∗∗∗ (0.061)Non-Premium+Small 2 .
428 (0.212)Premium+Small 2 . ∗∗∗ (0.213)Partial disclosure x Non-premium large − . ∗ (0.029)Partial disclosure x Non-premium small 0 .
002 (0.055)Partial disclosure x Premium small − .
032 (0.048)Full disclosure x Non-premium large 0 . ∗∗ (0.041)Full disclosure x Non-premium small − .
067 (0.077)Full disclosure x Premium small 0 . ∗∗∗ (0.070)Year16 x Non-premium large − .
047 (0.030)Year16 x Non-premium small − .
094 (0.058)Year16 x Premium small − . ∗∗∗ (0.056) Controls
Supply in millions − . ∗∗ (0.0004)Daily average buyers 0 . ∗∗∗ (0.002)Monthly ad spending 0 . ∗∗∗ Significant at the 1 percent level. ∗∗ Significant at the 5 percent level. ∗ Significantat the 10 percent level. A robustness check filtering out calendar week 6, 2015 whenprices spiked showed substantively similar results. impressions.
Partial disclosure x Year16 and
Full disclosurex Year16 ; rows
Partial disclosure x Year16 x Non-premium large and
Full disclosure x Year16 xNon-premium large show the incremental effect for non-premium/large sites. Similarly, the effectsfor Non-premium small and Premium small sites are shown in the table. Because the total effect forall four types of sites is difficult to determine from the regression table, we illustrate the estimatedeffects in Figure 10. Premium small sites benefit the most from the policy change with an estimatedeffect more than three times that of the average (15.3 + 31.6 = 46.9 EUR cents). These small,premium sites typically serve a “niche audience” with a very specific topical interest (e.g., a websitethat provides content targeted at physicians) and once ad buyers know what they are bidding for,some of them value these sites far more than had they been in a bundle of unknown sites. Thus,the data suggests that such sites have the most to gain by increasing context transparency.
This study investigated a specific change in context information where an exchange moved fromproviding only the level of the “channel” to providing the subdomain associated with each ad34mpression in the bid request. Our analysis of the policy change shows that buyers value contextinformation above-and-beyond user information and act on it as soon as context information isavailable. Consequently, average revenue per impression rose after the policy change relative to theprevious year. As we illustrate with a simulation reported in Figure 2a, these effects are consistentwith a scenario where ad buyers prefer different sites. Such heterogeneous preferences lead to anincrease in prices with context disclosure (Tadelis and Zettelmeyer 2015), so long as the marketdoes not become thin (Levin and Milgrom 2010, Hummel and McAfee 2016). Under this scenario,ad buyers bid higher for the sites they each prefer after the policy change.Several mechanism checks provide convergent evidence that this is indeed what happened: 1)markets thinned out slightly suggesting fewer bidders were bidding on each site, but remainedsufficiently thick for prices to rise; 2) the distribution of winning bids shifted to the right meaningthat winning bids were more dispersed when ad buyers were provided with context information;3) most individual sites saw an increase in average revenue per impression; 4) consistent with oursimulations in 7, partial disclosure of information to a single buyer resulted in more auctions beingwon by this buyer. We are also able to rule out some alternative explanations. For one, the changein average revenue per impression was not due to ad buyers shifting budgets from programmaticdirect to RTB. In addition, ad buyer and site turnover was not responsible for the observed increasein revenue per impression. Our evidence therefore points towards buyers increasing their bids inresponse to the policy change.Finally, to answer the practical question of whether certain types of sites benefited more fromthe policy change, we also investigate heterogeneous treatments effects: The increase in averagerevenue per impression was most pronounced for sites with a large number of average daily buyersprior to the policy change (again, consistent with our simulations in 2a). From a more managerialstandpoint, we show that small, premium sites benefited the most from context disclosure. Theytypically serve a “niche audience” with a very specific topical interest and once ad buyers knowwhat they are bidding for, some of them value these sites far more than had they been in a bundleof unknown sites. Yet, there are almost no losers of the policy change in this market – we mainlysee sites that benefit more and sites that benefit less.These changes in prices were economically meaningful for this private exchange. The averageweekly supply for a site in our sample is roughly 3.5M impressions (see Table 1), sold for an average35PM of 88 EUR cents. According to our analysis, average CPM rises by about 15.4 EUR centswhen all ad buyers are provided with placement information (see Table 2, Column 1, Full disclosurex Year16). Therefore, on average each site generates an additional yearly revenue of (3.5M/1000)x 15.8 EUR cents x 52 weeks = e e e e e The extensibility of our findings to other markets depends critically on the sites and ad buyersthat participate in the market. Some of the sites participating in the exchange we study provideniche content, but even the smaller websites were brand-safe, reputable media outlets that hadbeen vetted by the exchange. This mix of sites is typical of private exchanges, and so our resultsstrongly suggest that private exchanges should provide information at the subdomain in the bidrequest. While we do not see buyer entry due to the policy change within our observation window36three months post-disclosure), it might be possible that buyers will want to join this specificprivate market if others do not provide the same level of context disclosure. Such buyer entrymight increase auction revenue even further.While our results translate fairly directly to other private exchanges, which are a growing shareof the display advertising market (Fisher 2020), is is more difficult to say how context disclosuremight affect open exchanges. Open markets attract a much wider range of sites (including “click-bait” websites, fake news, and other sites with low quality advertising environments) and it iscurrently left to individual publishers to decide whether to truthfully disclose the context for adimpressions they sell. There is certainly a higher risk of deconflation for sites that produce extremecontent and some may become “orphaned” by ad buyers, as evidenced by the recent drop in demandfor advertising at the alt-right site Breitbart.com in the US (Bhattarai 2017) when buyers becameaware that their programmatic ads were appearing on the site. Increased context transparencymay force them to leave the exchange if their revenues decrease substantially over time. Yet effortsby the IAB to increase transparency in the open markets (Sluis 2020) will, if successful, improveoutcomes for ad buyers, reputable publishers, exchange operators, and the industry as a whole.It also means that reliable publishers may want to avoid selling impressions on the open ex-changes when they are not transparent enough. Instead they may want to form their own privatemarkets as they can expect auction revenues to rise when providing more context transparency.Some publishers in the Netherlands have already done so, eliminating user tracking all together(Edelman 2020). Interestingly, digital revenue rose for those publishers. Private exchanges areincreasingly popular and are expected to process the majority of display advertising impressions in2020 (Fisher 2020).Our analysis shows that site placement information provides ad buyers with additional informa-tion about the value of an impression, above-and-beyond the rich cookie information available toEuropean ad buyers in 2016. That is, context information is complementary to user information. Ifcontext information is also a partial substitute for user information, then context information willbecome even more important as ad buyers’ access to user information becomes more limited. Reg-ulations like GDPR already limit the amount of user-level targeting that is possible. In addition,Google recently announced efforts to limit the use of third-party cookies in their Chrome browserby making “disable third-party cookies” the standard setting (AdExchanger 2020a,b). Thereby,37he market share of browsers (including Mozilla, Safari, and Chrome) that inhibit tracking willgrow to more than 80% in many countries in the next two years. Thus, user targeting will be lessstraightforward in the future and industry experts predict that contextual targeting will becomemore relevant (Schiff 2019, Tan 2019).There are few empirical studies that investigate the effect of reduced access to user informa-tion on outcomes of ad auctions. A recent working paper by Marotta et al. (2019) shows thatwhen the user’s cookie is available, publisher’s revenue increases, but the increase is just about4% corresponding to an average increase of $ The New York Times have moved to protect users by decreasing the amount ofuser information collected on their sites (Berjon 2020). In contrast, providing ad buyers with morecontext information is nearly cost less to publishers; we provide convergent evidence that doing soresults in a substantial revenue increases.While it is speculating beyond our data, we expect full URL disclosure to be very attractiveto ad buyers who can then target ads against the specific content in an article. In fact, startupssuch as Grapeshot (acquired by Oracle in 2018), Peer39, and Leiki (acquired by DoubleVerify)have been building machine learning tools to help ad buyers determine which URLs are mostattractive, based on the text on the page. However, finer levels of context disclosure may lead tothin markets and deconflation for specific URLs or particular content topics. At the same time, itmay encourage content creators and publishers to focus on content that is appealing to consumersand ad buyers rather than simply attracting an audience to generate impressions (Gal-Or et al.2012). We encourage future theoretical research that investigates how such fine-grained contextdisclosure could impact publishers’ incentives to produce content and the welfare of advertisers,publishers, and content consumers. 38 eferences
Abadie, A., A. Diamond, and J. Hainmueller (2010). Synthetic control methods for comparative case studies:Estimating the effect of California’s tobacco control program.
Journal of the American StatisticalAssociation 105 (490), 493–505.Abadie, A. and J. Gardeazabal (2003). The economic costs of conflict: A case study of the Basque Country.
American Economic Review 93 (1), 113–132.AdExchanger (2020a). After Google finally really did it, a necessary sys-tem cleanup. . [Last accessed: July2020].AdExchanger (2020b). How will publishers fare as Google moves tokill cookies in Chrome? https://adexchanger.com/the-sell-sider/how-will-publishers-fare-as-google-moves-to-kill-cookies-in-chrome/ . [Last accessed:July 2020].Amihud, Y. and B. Lev (1981). Risk reduction as a managerial motive for conglomerate mergers.
The BellJournal of Economics 12 (2), 605–617.Berjon, R. (2020). How the New York Times thinks about your privacy. https://open.nytimes.com/how-the-new-york-times-thinks-about-your-privacy-bc07d2171531 . [Last accessed: July 2020].Berman, R. (2018). Beyond the last touch: Attribution in online advertising.
Marketing Science 37 (5),771–792.Bhattarai, A. (2017). Breitbart lost 90 percent of its advertisers in two months;who’s still there. . [Last accessed: May 2018].Budak, C., S. Goel, J. M. Rao, and G. Zervas (2014). Do-not-track and the economics of third-partyadvertising. Working paper.Busenitz, L. W. and J. B. Barney (1997). Differences between entrepreneurs and managers in large orga-nizations: Biases and heuristics in strategic decision-making.
Journal of Business Venturing 12 (1),9–30.Chakraborty, I. (1999). Bundling decisions for selling multiple objects.
Economic Theory 13 (3), 723–733.Chen, Y., S. Li, and J. Yu (2018). Bundling decisions in multi-objects auctions with optimal reserve prices.
Economics Letters 163 , 141–145.Choi, H. and C. F. Mela (2018). Display advertising pricing in exchange markets. Working paper.Datta, H., G. Knox, and B. J. Bronnenberg (2018). Changing their tune: How consumers’ adoption of onlinestreaming affects music consumption and discovery.
Marketing Science 37 (1), 5–21.Eaton, D. H. (2005). Valuing information: Evidence from guitar auctions on ebay.
Journal of AppliedEconomics & Policy 24 (1), 1.Edelman, G. (2020). Can killing cookies save journalism? . [Last accessed: July 2020].Fisher, L. (2019). US Programmatic Digital Display Ad Spending. . [Last accessed: July 2020].Fisher, L. (2020). .[Last accessed: July 2020].Fou, A. (2020). How to avoid the pitfalls and loopholes of ads.txt. ites/augustinefou/2020/10/15/how-to-avoid-the-pitfalls-and-loopholes-of-adstxt/ . [Last accessed: October 2020].Gal-Or, E., T. Geylani, and P. Yildirim (2012, 08). The impact of advertising on media bias. Journal ofMarketing Research 49 , 92–99.Goldfarb, A. and C. Tucker (2011a). Online display advertising: Targeting and obtrusiveness.
MarketingScience 30 (3), 389–404.Goldfarb, A. and C. Tucker (2011b). Privacy regulation and online advertising.
Management Science 57 (1),57–71.Goldfarb, A. and C. E. Tucker (2014). Conducting research with quasi-experiments: A guide for marketers.Working paper.Hummel, P. and R. P. McAfee (2016). When does improved targeting increase revenue?
ACM Transactionson Economics and Computation (TEAC) 5 (1), 4.IAB Tech Lab (2020). OpenRTB 3.0. https://github.com/InteractiveAdvertisingBureau/openrtb .[Last accessed: July 2020].Johnson, G. A. (2013). The impact of privacy policy on the auction market for online display advertising.Working paper.Johnson, G. A., S. Shriver, and S. Du (2017). Consumer privacy choice in online advertising: Who opts outand at what cost to industry? Working paper.Johnson, G. A., S. K. Shriver, and S. Du (2020). Consumer privacy choice in online advertising: Who optsout and at what cost to industry?
Marketing Science 39 (1), 33–51.Lambert, R., C. Leuz, and R. E. Verrecchia (2007). Accounting information, disclosure, and the cost ofcapital.
Journal of Accounting Research 45 (2), 385–420.Lambrecht, A. and C. Tucker (2013). When does retargeting work? Information specificity in online adver-tising.
Journal of Marketing Research 50 (5), 561–576.Levin, J. and P. Milgrom (2010). Online advertising: Heterogeneity and conflation in market design.
TheAmerican Economic Review 100 (2), 603–607.Little, J. D. (1970). Models and managers: The concept of a decision calculus.
Management Science 16 (8),B–466.Lu, S. and S. Yang (2020). Targeting breadth in behaviorally targeted display advertising. Working paper.Marotta, V., V. Abhishek, and A. Acquisti (2019). Online tracking and publishers’ revenues: An empiricalanalysis. Working paper.Milgrom, P. R. and R. J. Weber (1982). A theory of auctions and competitive bidding.
Econometrica:Journal of the Econometric Society 50 (5), 1089–1122.Palfrey, T. R. (1983). Bundling decisions by a multiproduct monopolist with incomplete information.
Econo-metrica: Journal of the Econometric Society 51 (2), 463–483.Rafieian, O. and H. Yoganarasimhan (2020). Targeting and privacy in mobile advertising.
Marketing Science .forthcoming.Reiley, D., R. A. Lewis, and T. A. Schreiner (2012). Ad attributes and attribution: Large-scale fieldexperiments measure online customer acquisition. Working paper.Schiff, A. (2019). Can contextual targeting replace third-party cookies? . [Last ac-cessed: July 2020].Shamdasani, P. N., A. J. Stanaland, and J. Tan (2001). Location, location, location: Insights for advertisingplacement on the web.
Journal of Advertising Research 41 (4), 7–21. hehu, E., N. Abou Nabout, and M. Clement (2020). The risk of programmatic advertising: Effects of websitequality on advertising effectiveness. International Journal of Research in Marketing . forthcoming.Sluis, S. (2020). Domain spoofing be gone: Ads.txt will filter out imposter sites. . [Lastaccessed: July 2020].Tadelis, S. and F. Zettelmeyer (2015). Information disclosure as a matching mechanism: Theory and evidencefrom a field experiment.
American Economic Review 105 (2), 886–905.Tan, J. (2019). Can contextual targeting replace third-party cookies? . [Last ac-cessed: July 2020].Tirunillai, S. and G. J. Tellis (2017). Does offline TV advertising affect online chatter? Quasi-experimentalanalysis using synthetic control.
Marketing Science 36 (6), 862–878. eb AppendixA Interface of demand-side platform Figure A.1: Example of a demand-side platform (DSP) interface where users can enter the URLsof the sites where they want their ads to appear.42
Effects of context disclosure by month
Table A.1: Treatment effects month by month.Average revenue Average revenueper impression per impression
Treatment effects
April (Partial disclosure) x Year16 0 . ∗∗∗ . ∗∗∗ (0 . . . ∗∗∗ . ∗∗∗ (0 . . . ∗ . . . . ∗∗∗ . ∗ (0 . . Baselines
Constant (No disclosure) 1 . ∗∗∗ . ∗∗∗ (0 . . − .
007 0 . . . . ∗∗∗ . ∗∗∗ (0 . . . ∗∗∗ . ∗∗∗ (0 . . . ∗∗∗ . ∗∗∗ (0 . . . ∗∗∗ . ∗ ∗∗ (0 . . Controls
Supply in millions − . ∗∗∗ (0 . . ∗∗∗ (0 . . . R Notes: Standard errors, in parentheses, are clustered at the week level. ∗∗∗
Significant at the 1 percent level. ∗∗ Significant at the 5 percent level. ∗ Significant at the 10 percent level. A robustness check filtering out the spikesin price in calendar week 6, 2015 showed substantively similar results. Distribution of winning bids in 2015
Figure A.2: Density plot of average price paid per impression for 2015.44
Placebo tests for synthetic control
To assess statistical significance of the synthetic control results, we conduct a series of placebo testsby applying the synthetic control method to the ad buyers who were not provided with placementinformation. By doing so, we produce a distribution of weekly estimated gaps between each adbuyer and its optimal synthetic control (see Figure A.3). The quality of fit of the synthetic controlcan be assessed by using the mean squared prediction error (MSPE) prior to the policy change.Following Abadie and Gardeazabal (2003), Abadie et al. (2010) and Tirunillai and Tellis (2017),Figure A.3 visualizes the placebo buyers having a pre-intervention MSPE of less than 5 times theMSPE of the treated buyer which results in 131 control buyers in Figure A.3a and 121 control buyersin Figure A.3b. As denoted by the thick black lines in Figure A.3a and A.3b, the synthetic controlmethod provides a very good fit for the treated buyers. The estimated number of impressions wonhas a p-value of 0.016. , suggesting the theoretically-predicted, significant increase in impressionswon for the treated bidder. Also consistent with theory, we do not observe a significant effect onaverage winning price in the same period (p-value = 0.932). This is the standard method of assessing significance for synthetic controls (cf. Tirunillai and Tellis 2017). p-values are calculated by means of the ratio of post - pre intervention MSPE. If an ad buyer were randomlytreated, the probability of obtaining a post - pre intervention MSPE ratio as large as the one for the treated buyerwould be 2 (number of ad buyers exceeding the treated buyer’s MSPE ratio) over 127 (number of ad buyers).p-values are calculated by means of the ratio of post - pre intervention MSPE. If an ad buyer were randomlytreated, the probability of obtaining a post - pre intervention MSPE ratio as large as the one for the treated buyerwould be 2 (number of ad buyers exceeding the treated buyer’s MSPE ratio) over 127 (number of ad buyers).