“Data-Driven Thinking" is written by members of the media community and contains fresh ideas on the digital revolution in media.
Today’s column is written by Eric Berry, CEO at TripleLift.
Every major ad exchange uses second-price auctions, yet almost no major ad exchange really uses second-price auctions. Sell-side platforms have created yield-enhancing technologies, such as cascading floors, hard and soft floors and other techniques – all ultimately with the single objective of clearing auctions at a price higher than the second bid price, as often as possible. This underscores the questions of why the industry leverages second-price auctions in the first place, whether they continue to be appropriate and how to move forward.
Some of these questions have been raised before, but the exchange space has matured and consolidated since then – suggesting it's a good time to revisit them.
Real-time bidding achieved meaningful scale largely because Google’s AdX enabled the protocol. In the early days of RTB, startups designed new bidding algorithms, each processing a previously unfathomable number of auctions every day. The possibilities for optimization were endless – but as with any complex nonlinear optimization, the potential for unknowns meant buyers needed an abundance of caution. A runaway algorithm could bankrupt a company.
Google had successfully leveraged second-price auctions – also known as Vickrey auctions – as the foundation for AdWords and later continued their use as the mechanic for real-time bidding. This allayed many of the above concerns. Buyers could bid and they would be assured that any unreasonable price would be reduced to someone else’s bid. This use of second-price auctions increased the appeal and reduced the friction. In a world with dozens of billions of daily auctions, this significantly increased liquidity and buyer confidence, with a side effect of keeping prices low enough to make RTB attractive for continued investment.
It must be noted that the efficient properties of second-price auctions do not hold where each bidder’s valuation methodology may have errors based on unknown, faulty logic or unpredictable optimization strategies. Thus, while the second-price auction is risk-mitigating, its precise function here neither ensures its efficiency for pricing nor broadly justifies its use. The ultimate purpose of Vickrey auctions is to maximize expected utility for each participant. This contrasts from expected value in that expected utility considers the riskiness of individual bets. In the real-time bidding context, this is not as salient of a concern because millions of impressions are won. In a context with such high numbers of trials, the concerns that would apply for an isolated risky bet are significantly less important.
As RTB matured as an ecosystem, DSPs began to consolidate. Only a few DSPs currently represent the vast majority of the demand, and the algorithms have been refined and optimized. Vickrey auctions work when every market participant submits their true values – values that they believe the underlying asset is worth.
This holds in AdWords, where every buyer is evaluated concurrently in Google’s own system, and each with the objective of paying a certain amount for a click. This does not work in the display advertising context of sparse bid density, when each of the six or so major DSPs evaluates its share demand, and submits its top bid (or, sometimes, top few). The bid price is thus not reduced to the actual second-highest price for that auction, but the second-highest price that was submitted for that auction – necessarily one from a different DSP. Thus a basic tenet of second-price auctions is not held, specifically that not every bid is actually represented.
Further, unlike AdWords, where there is a single, accurate means of determining the value to the buyer, a fundamental issue with the application of real-time bidding to display advertising is the fact that RTB is still almost entirely used for direct-response marketing. The only data that can truly feed back into any optimization system is clicks, conversations and the like. In a world where search is known to be overrepresented in conversion attribution, and the branding impact of display is not (and possibly cannot be) fully appreciated in an RTB context, the bids submitted will almost never be the “true” values that could be used for display ads.
Contrary to RTB, publishers and their direct sales teams often highlight the branding impact of the ads they sell. Among numerous other reasons, this has led to a higher clearing price. When reselling inventory in ad exchanges, publishers tend to be somewhat disappointed with the clearing prices. Exchanges and supply-side platforms are expected to deliver service and yield to their partners, but are unable to control the amount that DSPs bid. Thus their relative yield advantage must come in the form of yield-management techniques. The DSPs tend to be aware – cascading floors, after all, are trivial to detect. Thus they either redirect their demand or include negative modifiers. This further violates a basic principle of second-price auctions: that participants submit their true value.
Second-price auctions have succumbed to the divergent incentives between supply- and demand-side platforms. DSPs seek to purchase the “best” impressions for the lowest price possible. SSPs prefer these same impressions transact at the highest possible price. Because SSPs and exchanges actually control the auction mechanic, do not reveal actual bids for any given auction and are paid more for higher-priced impressions, they are actually acting completely rationally for their position by elevating the price paid beyond the true second price.
In reality, second-price auctions are used by DSPs for price discovery, buyer confidence and value creation, but they are neither accurate nor currently serving their true purpose. Instead, second-price auctions undervalue banner inventory and lead to the unfortunate practices discussed above.
The solution is twofold. First, move to a first-price auction. Buyers should bid what impressions are truly worth to them, explicitly subtracting their margin. Over the millions of transacted impressions, deviations from statistical expectations will decline to nearly zero. Unexpected profits deriving from low bid-density impressions will be lost, but in its place DSPs will bid in an honest and predictable environment that would promote liquidity and actually incentivize publishers to make higher-quality inventory available. Indeed, under many circumstances, first-price auctions are shown to promote the same or greater honesty in bidding than second-price auctions.
To the question of price discovery, we recommend a system of full price transparency, whereby each bidder receives the clearing price for every impression (though not the identity of the winner). This allows accurate forecasting and incentivizes a more traditional English (ascending) auction for high-quality impressions. We understand that these are significant deviations from the norm in online advertising, but over the long run, they stand to significantly improve the value that all parties receive from real-time bidding.