Header Bidding And The Fate Of RTB

ericberry-sellsiderThe Sell Sider” is a column written for the sell side of the digital media community.

Today’s column is written by Eric Berry, CEO at TripleLift.

Header bidding has added significant value to the publisher ecosystem, maximizing revenue and creating opportunities to democratize access to the ad server.

A series of fascinating trends in the header bidding context have emerged that, if they continue, will significantly change the dynamic of real-time bidding (RTB).

Real-time bidding is based on the proposition that the highest bid wins at the second-highest price. While the justification for this principle may or may not be sound, agents can profit if they can correctly value an impression with low demand – or at least with a relatively high differential between the first and second auction prices, reducing the bid price significantly. This dynamic only works if there is a single auction for the impression. That means all bids are compared at the same time, quickly determining the highest and second-highest bids.

Header bidding relies on a number of parallel auctions generally submitting bids to compete against demand in AdX, Google’s ad exchange, and first-party line items in Google’s DoubleClick for Publishers (DFP). DFP does not conduct a second-price auction. Header bids pay the number they submit.

This means that a number of header bidding results compete against the prices from DFP line items and the price submitted through AdX, possibly the result of a second-price auction. If an exchange that submits a bid through a header bidding system includes a buyer willing to pay the highest overall price – high enough that it would win against any other buyer in any other header bid and DFP, but the second-highest price on that same exchange is sufficiently low that it would not win – the choice of what to submit in a header bid presents a challenge.

It would be highly inefficient if the exchange responded with a value in the header at the second price. The publisher loses because its revenue isn’t maximized. The “winning” advertiser loses because it doesn’t actually win the impression it wanted and submitted a high bid for. And the theoretically winning exchange loses because it doesn’t win an impression it could have won and doesn’t collect its margin.

All constituents here would be better served by an auction that clears between the first and second price. But what if that same exchange had submitted a number higher than its second price, such that the advertiser that submitted the highest bid had been able to win at a price still lower than its own bid? The publisher would then earn higher revenue per thousand impressions. The advertiser wins the impression it wants while still retaining some margin below its bid. The exchange representing the highest single demand source also wins.

As a result, exchanges will be tempted to increase the price they’re willing to submit. This would put them somewhere in the range between the first and second prices so they could clear more impressions. This will likely create a race to the bottom where bids submitted by exchanges in the header bidding context approach the first price, minus some margin.

This first-price effect is exacerbated by the fact that the same impression is sent to multiple exchanges in the form of multiple header bid requests. This results in each exchange in turn sending the bids to generally the same set of demand-side platforms (DSPs).

Theoretically, a DSP should bid the same amount for the same exact impression, regardless of exchange. That, however, is not the case in reality. DSPs bid against themselves across exchanges – not just for impressions lower in the waterfall, but at the same exact position in the waterfall.

This means a DSP may see an impression on two different exchanges and bid accordingly based on its biases for each exchange. Because both exchanges will eventually move closer to a first-price auction, the DSP will simply end up paying the highest of its bids for any impression, across any exchange. This duplicates bid processing efforts.

DSPs could be tempted to only choose to buy inventory for a certain publisher from a certain exchange, reducing the overhead from header bidding. Publishers, however, could easily circumvent this by randomly excluding certain header tags on certain impressions, meaning no exchange could see every impression.

These problems are significantly less prevalent for native advertising because few vendors support header bidding and there is no “replication” of bids due to each vendor providing a completely unique experience.

For banner ads, however, the advertising product is more fungible between exchanges. It’s essentially the same underlying asset that’s being shown to the user, without any special rendering by the exchange. In this context, one natural result may be larger DSPs developing their own header bidding solutions to preempt declining bid reduction and the cost of highly duplicated bid requests.

Another result might simply be the natural evolution of RTB toward first-price auctions. This has been discussed before, but header bidding may accelerate what to many industry observers appears to be the natural conclusion of RTB.

Follow Eric Berry (@ezberry), TripleLift (@triplelifthq) and AdExchanger (@adexchanger) on Twitter.

1 Comment

  1. Eric: great stuff. I’ve been wondering when someone might point out the problem with 2nd price in the header bid. For the most part, I was thinking about it from the standpoint of the interaction with the line items in the ad server, which will be fixed price (I’m thinking of the process described by Ben Kneen at http://www.adopsinsider.com/header-bidding/header-bidding-holistic-ad-serving/). Once a winner is found in the header, it gets passed to a fixed price line item in the ad server. Price reduction can’t really be honored in this scenario. I believe this is true even in the case where there is only one header tag.
    Your point seems to be about the intelligence required of the header bidders in determining what price to submit, and I agree that there’s a “race to the bottom” (though I’m not sure I would call 1st price “the bottom”). The other problem you’ve identified is the fact that if two exchanges are in the header, and the same DSP is integrated with both, then that DSP is bidding against itself. Based on my selling experience, buyers HATE that. The main impact is that the 2nd price bid is effectively converted to 1st price, which exposes the buyers True Private Value, which is against the rules in a Vickrey auction.
    I think there’s pressure on second price auctions from programmatic in general, since Vickrey’s theory never contemplated the multiple cascading auctions that occur in the handoff between the DSPs and the Exchanges. This means that the auction we stole from search may not be serving us properly in programmatic. Esco Strong pointed out other problems with second price back in prehistoric times (2012): http://adexchanger.com/data-driven-thinking/second-guessing-the-second-price-auction-model/. Thanks for raising this very interesting issue.


Add a comment

XHTML: You can use these tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>