Bidding Models for Exchanges

RTB auction dynamics is a topic we've discussed a few times in the past, but it's worth revisiting from time to time. Today we discuss what the theoretical ideal auction dynamic should be in today's RTB ecosystem as well as the practical ideals, and why they're different. This article is almost entirely about display (banner) ads at the moment, but it's possible that it will eventually apply to native or other formats.


Second-price auctions became popular in advertising as part of Google's AdWords platform (search keyword bidding). In this context, each separate advertiser put its own bids into AdWords at the maximum value that will still drive them profit, with the additional comfort that the bid will be reduced at least somewhat, resulting in additional profit. The advertisers don't know each other's bids but are incentivized to reveal their maximum true value to win the most clicks and still derive profit from the bid reduction.

Google continued the use of second price auctions in AdX, their real-time bidding ad exchange. In the early days of this RTB exchange, it created value and comfort for buyers for ad impressions that would largely be unfilled otherwise. Further, there was a single auction and all the buyers participated in a single unified auction. In the years since, nearly no part of the RTB ecosystem resembles this early exchange.

In AdWords, every single buyer bids into Google's consolidated keyword auction system. The most expensive keyword in 2017 was "insurance." Geico, Allstate, Progressive - as well as Quote Wizard,, etc all bid separately for this word. There are likely hundreds of bids, creating a dense auction with a number of ad slots. In RTB, however, there is no precise analog -- perhaps individuals that are "in-market auto buyers" based on audience data, who are then browsing on a car reviews website.  Even assuming there were one consolidated auction (which there isn't), market share has consolidated substantially into two (maybe three or four depending on who you ask) demand-side platforms. Each DSP performs its own auction, applies some internal bid reduction logic and returns a single bid. What may have been a very dense auction with hundreds of potential buyers could appear to the exchange like a sparse set of bids - perhaps only a couple bids, none of which represented the true value of any potential buyer. Thus the auction no longer has the full set of bids nor does it have the bids that accurately reflect the true amounts. The theoretical framework of a second-price auction does not apply to this scenario. It's further distorted in a header bidding context, where there are a number of parallel auctions for the same impression, each of which may be some form of a second-price auction.

The ideal, of course, would be a real second price auction where each DSP submits all of their bids (or at least their top two prices) into a single consolidated auction. This would also require that the DSP not reduce their bids based on some estimate of market price and that the single consolidated exchange hold a pure second-price auction - and that the publisher understand that the monetization on their property is simply a function of market dynamics. No part of this idealized case is likely to become reality any time soon, if ever. This begs the question of what the idealized outcome should be given the dynamics that do exist.

DSPs bid some value into the various exchanges in which they participate. Given their knowledge of the ecosystem, including that they are bidding against themselves in the various auctions for the same impression, it would be imprudent to bid the actual true value for a particular bid. Instead they will be reducing the bid by some amount - with the expectation that other DSPs will be doing the same, that there aren't too many DSPs so demand is somewhat sparse and power is substantially with the buyers, that there will likely be enough inventory to fill their campaign, and with the understanding that exchanges will likely not be reducing their impression by much, if at all.

Exchanges conduct their own auction based on the values received from the DSPs. They are receiving what are likely reduced bids and are competing against other exchanges who also likely received reduced bids to win the auction. The exchange's first priority is to win the most impressions for a publisher so they get paid the most and are also the most important to that publisher. But exchanges also know that DSPs will be conducting supply path optimization.

The relationship between the current state of affairs discussed above and actual second-price auction mechanics is, at best, attenuated. DSPs are not bidding true prices, the full set of bids are not disclosed, downstream auctions are often not true second price, especially since the final auction is in a subsequent DFP auction.  This leaves two outcomes. The first is that exchanges effect a pure first-price model. This means that the DSP's bid, minus the exchange's margin, is passed right into the header. At this point, the SSP is adding little value beyond ad quality - and the DSP might as well integrate into the publisher's header to avoid the duplicate impressions and the waste of the commoditized format SSP's margin (this is effectively what HeaderDirect expects and is building technology to facilitate as close to this relationship as possible).

The second outcome is that the exchange as a business moves away from purely acting as an agent for the publisher and instead tries to get the DSP additional value for its inventory. In this case, the exchange uses its understanding of the ad impression itself, what it is likely to clear for, and reduces the DSP's bid to some amount that ostensibly gets the DSP a similar likelihood of winning but at a lower amount. This, of course, changes the dynamics in the relationship between the exchange and the publisher.

It is very possible that both outcomes will exist at once in the ecosystem - both soon and for some time. In the first outcome, DSPs that have adapted to the not-second-price model will algorithmically price in what they believe the best price for the inventory would be, based on their own understanding of the inventory, including minimum likely clear price (why would the exchange do a better job pricing than the DSP, a reasonable DSP might ask). DSPs that have yet build this form of pricing or focus on other parts of their algorithm may see a sensible partnership opportunity in the second outcome. And thus it is reasonable to expect both models to proliferate and the ecosystem move even further from anything that could sensibly be called a second price auction.