First-Price Ad Auctions Raise Prices Before Bid Shading Catches Up

Internet display advertising moved from second-price auctions to first-price auctions because exchanges and publishers wanted a simpler, more transparent mechanism for real-time bidding. The economic question is whether bidders immediately respond as theory predicts: in a first-price auction, a bidder should shade below value rather than bid truthfully as in a second-price setting. This paper studies that adjustment using publisher-level field data from display ad auctions, where different publishers adopted first-price auctions at different dates.

The central finding is empirical rather than theoretical. Prices rose sharply when publishers switched formats, which is what one would expect if bidders did not immediately shade enough. In several later switches, those price gains then weakened over time, consistent with bidders learning to adapt rather than jumping straight to the new Bayesian Nash equilibrium.

Core Contribution

The contribution is a field measurement of behavior around an auction format change that had mostly been studied through theory, simulations, or controlled settings. The authors observe real advertising auctions and use staggered adoption across publishers to compare treated publishers against control publishers that remained under second-price auctions during the same period.

That design matters because ad markets have strong common shocks: demand changes by week, campaigns come and go, and publisher quality differs. Figure 2 lays out the treatment timing and the control groups used for the difference-in-differences comparisons, separating early and later switches across the global company and European media companies.

Technical Approach

The paper estimates the effect of switching from second-price auctions to first-price auctions on revenue per sold impression, measured as price. The baseline specification is a difference-in-differences regression with publisher and time controls. The authors then use a synthetic difference-in-differences estimator to address a central concern in this setting: treated and control publishers can have different pre-treatment trends even when they are exposed to the same broader ad market.

The empirical design uses several treatment-control comparisons rather than a single before-after change. The figures and tables distinguish publisher groups by company, geography, and adoption date. Figure 3 shows the raw weekly average price series around the switches. The visual pattern is useful: several treated groups show a price jump near the format change, but the pre-periods are not identical across groups, which is why the synthetic difference-in-differences analysis is doing real work rather than serving as a cosmetic sensitivity check.

The interpretation is tied to bidding incentives. Under a second-price auction, bidding close to value is weakly dominant in the standard private-values model. Under a first-price auction, the winner pays its own bid, so bidders should shade. If bidders are slow to change bid logic, budgeting rules, or platform-side bidding algorithms, first-price adoption can temporarily raise publisher revenue.

Results and Analysis

The headline result is a large short-run price increase. Across the studied format changes, the paper reports increases in revenue per sold impression ranging from 25% to 70% of the treated group’s pre-treatment price level. That range is too large to read as a minor implementation artifact. It indicates that the auction rule change transferred surplus toward publishers, at least initially.

Figure 4 summarizes event-study estimates for average price around the format change. The estimates show a post-treatment increase rather than an immediate flat adjustment, while the confidence bands make clear that the effect size and persistence vary by comparison.

The time path is as important as the initial jump. For later switches, the paper finds that the price increase under first-price auctions tends to dissipate over time in several specifications. The authors read this as evidence of gradual bid shading: bidders appear to learn or update systems after facing the new payment rule. This is reminiscent of revenue equivalence, but the paper is careful about the wording. The data do not show instantaneous convergence to the textbook first-price equilibrium; they show a transition period in a production market.

The evidence is strongest for the direction of the initial price response. It is more qualified on the long-run endpoint, because the degree of reversion depends on the estimator, comparison group, and adoption episode. That distinction is the main practical takeaway. Auction designers should not assume that equilibrium predictions describe day-one revenue after a format change. They should also not assume the initial revenue gain is permanent.

Caveats in Practice

This is a quasi-experimental field study, not a randomized switch. The authors address pre-trends with synthetic difference-in-differences, but treated publishers may still differ from controls in ways that are hard to observe. The dataset is also drawn from particular advertising companies and publisher groups, so the magnitude may not transfer one-for-one to other exchanges, bidder populations, or eras of automated bidding.

The paper’s value is that it measures a real transition. For ad exchanges, publishers, DSPs, and auction designers, the result narrows the gap between auction theory and operational deployment: the format change matters not only through the equilibrium it implies, but through how quickly bidders can recognize it and change their bidding systems.