Internet Advertising and the Generalized Second-Price Auction

Benjamin Edelman (Harvard), Michael Ostrovsky (Stanford GSB) and Michael Schwarz (Yahoo! Research), American Economic Review 97(1):242–259, March 2007. Read in full from the authors’ copy (25 pp, dated 1 August 2006).

The paid half of this spoke has run on b2b-ppc-metrics and groas — what an advertiser measures after the fact — with no page on the mechanism that sets the price. This is that mechanism, described by the economists who named it.

Why it exists at all

“Google’s total revenue in 2005 was $6.14 billion. Over 98 percent of its revenue came from GSP auctions.” Yahoo!‘s was $5.26 billion with over half believed to come from the same mechanism, and the two companies’ combined market capitalization passed $150 billion by May 2006. The auction was not studied in the mechanism-design literature before this paper; it was invented in the market and analysed afterwards.

The two auctions before this one

Manual sales. Ads were sold “manually, slowly, in large batches, and on a cost-per-impression basis,” a few thousand dollars a month per contract.

Generalized first-price (Overture, then GoTo, 1997). Per-click bidding on keywords, ads ordered by bid, each click billed at the bidder’s own most recent bid. Cheap to enter and transparent — and unstable, because bids could change at any moment. The paper’s example: two slots (200 and 100 clicks an hour), three advertisers valuing a click at $10, $4 and $2. Bids climb by a cent at a time, $2.01 → $2.02 → $2.03, with no pure-strategy equilibrium in the one-shot game. Everyone’s best move is to re-bid as often as possible, so advertisers buy bidding robots, and “the costs that buyers incur while trying to ‘game’ an auction mechanism are fully passed through to the seller.” Worse, if only one bidder has a fast robot, the engine’s revenue collapses to 2.02 cents per click in that example no matter how high the true values are.

Generalized second-price (Google AdWords Select, February 2002). An advertiser in position i pays the bid of position i+1 plus a minimum increment, typically $0.01. Yahoo!/Overture followed. Same three advertisers, same slots: payments become $4 and $2 per click, so $800 and $200 in total. The gaming incentive is gone because a bidder cannot be pushed up by a cent-at-a-time war.

The result the paper is known for

GSP looks like the Vickrey-Clarke-Groves auction — both charge on other bidders’ bids, not your own — and Google’s own marketing leaned on the resemblance, saying its “unique auction model uses Nobel Prize-winning economic theory to eliminate … that feeling that you’ve paid too much.”

It is not VCG. With one slot they coincide; with several they diverge, because VCG charges each advertiser the externality imposed on everyone below, while GSP charges the next bid down. In the running example the top advertiser pays $800 under GSP and $600 under VCG. The consequences:

GSP “generally does not have an equilibrium in dominant strategies,” and truth-telling is not an equilibrium. Bidding your value is the wrong move by construction.
Restricting attention to what the authors call locally envy-free equilibria, the set contains one with VCG payoffs — and that one is the worst such equilibrium for the search engine.
If everyone did bid true values under both, revenue would always be higher under GSP.
The corresponding generalized English auction has a unique ex-post equilibrium with VCG payoffs.

Quality score, in the model

Google’s real mechanism ranks by bid × quality score rather than bid alone. The paper handles this without special pleading: a set of bids is an equilibrium of GSP with position factors, advertiser quality scores γ and values s exactly when {γ·b} is an equilibrium of the basic model with values {γ·s}. Quality score rescales the problem, it does not change its shape.

The historical note matters more for this spoke: “Initially, Google simply used click-through rates to determine quality scores,” and only later added other factors, becoming less transparent as it went. The authors also record that Google’s estimated CTR is computed conditional on the ad attaining the first position, which is why position and quality cannot be read off each other. What Google says about this now is in google-ad-rank.

What it settles here

It gives search-ad-auction its foundations, and it supplies the domain’s clearest documented case of the pattern this spoke keeps meeting elsewhere: a platform changing a mechanism to remove a gaming incentive, and the optimization industry re-forming around whatever the new mechanism rewards. The 1997→2002 shift was about bid timing; the equivalent move on the organic side is publishing-volume-hurts-seo and scaled-ai-content-crawl-economics, and on the generative side c-seo-bench reports that the interventions do not work as advertised.

It says nothing about ad rank as operated today — the auction has twenty more years of changes on it, none of them in this paper.

search-ad-auction · google-ad-rank · b2b-ppc-metrics · groas · seo-operating-model-shift · c-seo-bench · synthesis