Executive Summary<\/b><\/h2>\n
Increasing amounts of programmatic inventory are being sold in first-price auctions (\u201c1PA,\u201d winner pays their bid) in many of the major SSPs, a departure from the second-price auction model (\u201c2PA,\u201d winner pays second-highest bid) that has historically prevailed in programmatic ad sales. SSPs moving to 1PA have claimed that this change should be beneficial for buyers, as it resolves the current 2PA design flaw that sometimes results in the highest bidder losing an auction. Certain SSPs and DSPs have also claimed the move to 1PA should be cost-neutral for buyers once they adapt their bidding strategies. <\/span><\/p>\n Auction theory predicts that this cost-neutrality claim is inaccurate and our real-world testing bears this prediction out. Based on our data, for a given set of inventory, advertisers are better off buying through an SSP conducting a 2PA. This is true even when a DSP auto bidder is used to buy the 1PA inventory; in this case, the 2PA-priced inventory still has a lower price. Advertisers can and should adjust their programmatic supply path accordingly, both by shifting investment toward 2PA-priced inventory when possible and by negotiating fixed-price private marketplace (PMP) deals with their biggest publishers. However, these near-term mitigation strategies are not the end of the story.<\/span><\/p>\n The programmatic auction system should be restructured as a classic single-stage 2PA. This would retain the benefits of the present 1PA model (highest bidder always wins) while adding the benefits of a 2PA system (bidding truthfully is optimal). This could be achieved through the following reforms: <\/span><\/p>\n This combination of changes would make the marketplace much more advertiser-friendly by making bid strategy straightforward and clearing the air around programmatic auctions while maintaining or improving yields for publishers. Hearts will be working with partners across the programmatic supply chain to bring this vision to life.<\/span><\/p>\n Figure 1: Private value, clearing price, and buyer surplus<\/p><\/div>\n Auctions have been used for millennia to allocate difficult-to-price items to buyers. While a tremendous variety of auction forms exist, this paper focuses on two: sealed-bid first-price auctions (1PA) and sealed-bid second-price auctions (2PA), the two forms that dominate the digital media landscape. Buyers must pursue different bid strategies as the auction form dictates (e.g., 1PA vs. 2PA) but will typically<\/span>\u00a0seek to maximize the difference between the value they or their clients place on an item (their \u201cprivate value\u201d) and what they ultimately pay for it (the \u201cclearing price\u201d).1\u00a0<\/sup><\/span>This difference is called a \u201cbuyer surplus.”<\/span><\/p>\n In a 1PA, the highest bidder wins and pays what they bid, regardless of the behavior of other bidders. By contrast, in a classic 2PA, the highest bidder wins the auction but only pays the second-highest bid amount. This is true even if the highest bidder values an item dramatically more than the second-highest bidder; in every case, the winner will pay the second-highest bid amount. By extension, the same bid amount will yield a different clearing price depending on whether the auction is 1PA or 2PA.<\/span><\/p>\n This has a direct bearing on the optimal bid strategy for each of these auction forms. The dominant strategy in 1PAs is to guess the lowest bid amount between the bidder\u2019s private value and the second-highest bidder\u2019s bid that will result in an auction win. Contrary to what some commentators have written, it is suboptimal to bid your private value in a 1PA; doing so eliminates the possibility of a buyer\u00a0<\/span>surplus. On the other hand, 2PAs are \u201cdominant-strategy incentive compatible\u201d (DSIC), meaning that the best or \u201cdominant\u201d strategy for bidders is to bid their private value. 2<\/sup><\/span><\/p>\n The simple dominant bid strategy afforded by the 2PA form benefits both buyers and sellers. Bidders do not have to concern themselves with determining an optimal bid distinct from their private value. Sellers, assuming a constant set of bidders, should receive an average selling price identical to what they would have received in a first-price auction (more on this later).<\/span><\/p>\n Thinking about the marketplace more dynamically, the elegance of the 2PA should give additional bidders the confidence to buy ads at auction\u2014driving up publisher revenues! \u00a0<\/span><\/p>\n Because of these benefits, second-price auctions have been widely adopted in digital media, with Google and Facebook most prominent among the firms using second-price auction variants to sell billions of dollars of digital ad inventory each year. Since all demand for Google search ads and Facebook social ads flows to those firms directly through their proprietary systems, all bids are aggregated into each relevant ad auction. As such, these auctions follow the classic 2PA auction design closely. The programmatic 2PA, on the other hand, diverges from the classic design dramatically. <\/span><\/p>\n Figure 2: Header bidding’s position atop the programmatic “waterfall”<\/p><\/div>\n Demand for programmatic ads is diffused across many DSPs, so digital publishers rely on SSPs to coalesce demand for their inventory. Not every DSP is integrated with each SSP, so publishers employ multiple SSPs to capture as much demand as possible. SSPs were originally given the opportunity to bid in sequence, with each publisher working its way down its list of partner SSPs until one transmitted a bid above the price floor. This sequential \u201cwaterfall\u201d approach was simple enough but resulted in forgone revenue for publishers since the SSPs never had to compete directly in a unified auction. \u201cHeader bidding\u201d was developed to solve this problem, forcing all SSPs to bid simultaneously into an instantaneous, yield-maximizing publisher auction.3<\/sup><\/span><\/p>\n The implementation of header bidding, however, broke the logic of classic 2PA. Now, the SSPs would independently receive bids from DSPs, conduct a 2PA, then pass the clearing price to the publisher. When presented with multiple clearing prices from SSP 2PAs for a given impression, the publisher conducts a 1PA, selecting the highest clearing price from the set of SSP auctions. The two-stage nature of this auction\u20142PA at the SSP, then 1PA at the publisher\u2014destroys the DSIC quality of a classic 2PA because the highest initial bidder has no guarantee of victory! <\/span><\/p>\n Figure 3: The problem: in 2PA, sometimes the highest bidder loses<\/p><\/div>\n The use of 2PA at the SSP in header bidding created bid-strategy problems for advertisers, who could not be assured that a high bid would result in an auction win. And for SSPs, it created a business problem: SSPs have a business model based on earning a commission, a \u201ctake rate,\u201d on winning bids, so their inability to submit the highest received bid unaltered to the final auction hurt their win rates and, therefore, their revenues<\/span>.4<\/sup> Even when an SSP won an auction, their take rate was a function of the second-highest bid only, further depressing their earnings. As a countermeasure, SSPs often non-transparently adjusted the bid transmitted to the publisher such that it was no longer a true second price.5<\/sup><\/span>\u00a0This boosted SSP revenues but damaged advertiser confidence in the integrity of the auction system. <\/span><\/p>\n Finally, while the move to header bidding did force SSPs to compete with each other head-to-head, it did not fully aggregate advertiser demand. There is still tremendous bid attenuation between advertisers and publishers. As the simple model in Figure 3 above illustrates, four advertisers may bid into two SSPs, but the publisher will only see two bidders. This means that the publisher can\u2019t conduct its own 2PA without suppressing revenues.<\/span><\/p>\n First-price auctions in the SSP were envisioned as a way to resolve all of these issues in one stroke. After all, if the two-stage auction is a 1PA in both stages, a bidder can be confident that she will win if she has the highest bid among all bidders, across all SSPs. Concordantly, SSP win rates should be identical to the percentage of the time a given SSP has the highest bid among its competitors. The business implications are positive: For a fixed set of bids, SSP revenues should increase. Advertiser confidence in the integrity of the auction system should be improved, too, as winning bidders\u2019 clearing prices should be identical to the bids submitted by their DSP.6<\/sup><\/span><\/p>\n Figure 4: First-price auctions ensure the highest bidder always wins but drive up cost<\/p><\/div>\n Unfortunately, the move to 1PA is not a win for advertisers for two key reasons. First, costs will increase. Second, it is impossible to construct an optimal bid strategy. Both predictions are grounded in existing auction theory literature and supported by our testing. The remainder of this paper examines the theoretical basis for our claims, the results of our field experiments and recommendations for advertisers.<\/span><\/p>\n Why does moving to 1PA directly increase advertiser costs?\u00a0William Vickrey is the father of modern auction theory, so much so that the sealed-bid second-price auction (introduced above) is commonly referred to as the \u201cVickrey auction\u201d in his honor. Vickrey\u2019s 1961 paper <\/span>Counterspeculation, Auctions, and Competitive Sealed Tenders<\/span><\/a> not only created the field of auction theory but also established a core concept called \u201crevenue equivalence.\u201d Vickrey used tools from game theory to demonstrate that if certain conditions were met, four standard auction types would generate equivalent revenues for sellers (and prices paid by auction winners):<\/span><\/p>\n The Vickrey auction and Blind auction are commonly analogized to programmatic advertising\u2019s 2PA and 1PA auctions, respectively. This is a mistake, because programmatic 2PAs deviate from the classic design in a way that violates the revenue equivalence requirements. Vickrey\u2019s revenue equivalence theorem requires the following conditions to be met to apply to a given auction design:<\/span><\/p>\n Programmatic second-price auctions, as shown in Figure 1 above, do not satisfy the first criterion: The impression does not always go to the highest bidder. Sometimes, a lower-bidding advertiser can win. <\/span>Therefore, programmatic 2PAs and 1PAs cannot be revenue equivalent.<\/b>7<\/sup><\/p>\n We can take this a step further than merely saying the two forms are not revenue equivalent. Since the programmatic 2PA sometimes allocates impressions to non-high bidders, <\/span>it is a mathematical certainty that average clearing prices will be higher in a 1PA auction form.<\/b>8<\/sup><\/p>\n Why is it impossible to construct a dominant bid strategy in programmatic 1PA?\u00a0Constructing a dominant strategy in a second-price auction is easy: A classic 2PA is fully incentive compatible. All a bidder must do is be aware of his private value and then bid it. The auction form takes care of the rest, ensuring that the high bidder captures the maximum possible buyer surplus.9<\/sup><\/span><\/p>\n Sealed-bid first-price auctions call for a more complex strategy that is rooted equally in probability and the bidder\u2019s private valuation of an item. After all, if a bidder values a media vehicle at a $10 CPM, bids $10 and wins at $10, no surplus is captured. Bidders, therefore, need to construct a strategy that includes \u201cbid shading\u201d\u2014the reduction of their private value to a bid that captures a surplus while still beating competing bidders with lower private values. Put simply, the more competition you face, the smaller the gap between your private value and your optimal bid. The underlying math is presented in depth in a blog post at <\/span>Mathalicious<\/span><\/a>, but the takeaway is straightforward: Against \u201cn\u201d opponents, the dominant strategy is to bid n\/(n+1) of one\u2019s private value. The visual below illustrates how this plays out in a case where a bidder\u2019s private value is $5:<\/span><\/p>\n Figure 5: Optimal Bid Strategy in 1PAs: \u201cbid shading\u201d<\/p><\/div>\n Unfortunately, this ideal approach to bid shading is technically impossible to execute in the world of programmatic advertising due to information deficiencies on the advertiser side. While advertisers do have access to certain metrics related to the auction like submitted bid values, clearing prices and win rates, one thing they never see is the number of competing bidders. In fact, no single party in the programmatic supply chain sees all bids:<\/span><\/p>\n Constructing a dominant 1PA bid strategy requires information that no one in programmatic advertising has: the number of competitors a given advertiser has in each auction. Without this knowledge, constructing a dominant bid strategy for 1PA is impossible.10<\/sup><\/span>\u00a0Bidders will necessarily realize suboptimal outcomes compared to what they could achieve with a classic 2PA, which does not require knowledge of competitor bidder volume.<\/span><\/p>\n Auction theory predicts that the move from the programmatic 2PA to 1PA will be negative for advertisers. Irrespective of any changes advertisers make to their bid strategy, it is a mathematical certainty that clearing prices will increase. Separately, the information-poor auction environment that bidders encounter makes it impossible to implement a dominant 1PA bid strategy.<\/span><\/p>\n We set out to see if the predictions of auction theory would be supported by field tests in the current programmatic auction environment. <\/span><\/p>\n Acknowledging that, in a two-stage auction, using a second-price auction in the first stage can lead to undesirable losses due to an uncompetitive second price being passed to the final first-price auction.<\/span><\/p>\n We hypothesize that these losses are nevertheless outweighed by the superior buyer surplus produced by the second-price auction format.<\/span> Our test was designed to be as favorable to 1PAs as possible. Of our three cells, only one was a 2PA. 1PAs were granted the benefit of a cell with an auto bidder. We assumed that if our predictions about the effects of changing auction types were inaccurate, 1PAs using a DSP\u2019s bidding algorithm would easily outperform 2PAs using a simple flat bid strategy.<\/span><\/p>\n Figure 6: Our three-cell test<\/p><\/div>\n To construct the test cells, we relied on SSPs\u2019 own reports of their auction type, our DSP\u2019s guidance and confirming coverage in the advertising trade press. The 1PA test cells only utilized SSPs that had publicly switched to this auction format. For our 2PA test cell, all media ran through a single SSP that was widely known to only conduct 2PAs. <\/span><\/p>\n The test cells ran concurrently through a single DSP, all with a simple demographic target serving across the United States. All media was bought exclusively from authorized sellers as declared by publisher ads.txt files, so we know that none of the inventory was \u201cspoofed.\u201d Each test cell also had the same daily rate of spend in order to control for volume of demand as a confounding factor. <\/span><\/p>\n Conscious of the complexity here, we took steps to control for certain factors when conducting our analysis of performance:<\/span><\/p>\n We found that our 2PA test cell with flat bids outperformed both 1PA cells. While the use of an auto bidder reduced clearing prices relative to the 1PA cell with flat bids, these prices were still higher than those returned by the 2PA auction cell. <\/span><\/p>\n Figure 7: Aggregate Results: CPM and vCPM (second-price flat bid CPM indexed to 100)<\/p><\/div>\n There was no significant variation in either viewability rates or in-demo delivery accuracy across our test cells, which tells us that we were serving inventory of a consistent quality and reaching the same types of users. That is important because it contradicts any claims that 1PAs provide access to \u201cmore premium\u201d inventory.<\/span> Figure 8: Aggregate Results: Viewability and In-demo Accuracy<\/p><\/div>\n We further saw that these patterns were repeated publisher by publisher, even when we controlled for variations in ad unit composition:<\/span><\/p>\n\n
Auction game basics<\/b><\/h2>\n
![\"\"](\"https:\/\/www.hearts-science.com\/wp-content\/uploads\/2018\/11\/sold-for-image-01.jpg\")
<\/a>Programmatic second-price auctions stand apart<\/b><\/h2>\n
<\/a>The diagram below illustrates the problem:<\/strong><\/h3>\n
<\/a>The shift to first-price auctions<\/b><\/h2>\n
<\/a>Auction theory <\/b><\/h2>\n
\n
\n
<\/a>\n
Empirical results<\/b><\/h2>\n
Hypothesis<\/b><\/h3>\n
\n<\/span><\/p>\nExperimental design<\/b><\/h3>\n
<\/a>Analytical controls<\/b><\/h3>\n
\n
Findings <\/b><\/h3>\n
<\/a>
\n<\/span><\/p>\n<\/a>