Randomizing in Search Ad Auctions to Extract Revenue from Runaway Winners

This paper studies a randomizing threshold auction of the kind online platforms use to sell search-advertising slots. The mechanism awards the slot outright only when the highest bid sufficiently exceeds the runner-up's, and otherwise assigns it by lottery among the close bidders. Such relative-bid randomization can raise revenue above the standard second-price auction when advertisers' values are bimodal: usually low but occasionally very high, reflecting whether an ad matches the user's query. It does so by threatening runaway winners with the risk of losing, and charging them a certainty premium to restore a sure win. Welfare effects cut both ways. A second-price auction having only its reserve with which to tax a runaway winner can, once the best-matched advertisers are far enough ahead, only refuse to sell to the weaker ones outright. The randomizing auction has a second instrument and keeps serving weaker bidders. Over the range, randomization raises both revenue and total welfare. However, the best-matched advertisers are nonetheless taxed and ad relevance for users is lowered. A learning exercise further shows that randomization prevents advertisers from separately inferring rival values and the pricing rule from observed outcomes. These results are then interpreted in the light of the U.S. antitrust case against Google, which ability to impose the randomized generalized second-price auction (rGSP) was held to be direct evidence of monopoly power and the resulting remedies have begun to require auction transparency.