Nonparametric Estimation of Sponsored Search Auctions and Impact of Ad Quality on Search Revenue

dongwoo-kim (Simon Fraser University) and pallavi-pal (Stevens Institute of Technology), Management Science 71(12):10047–10066, 2025; earlier as CESifo Working Paper 10312 (2023) and cemmap CWP16/24. Read in full from the cemmap working paper (63 pp), extracted locally.

Ingested via the hub research pass against the edge the ad auction as operated, measured by anyone but the platform. It is the first source in this spoke to put a number on what the quality weight does — and the number belongs to the platform's side of the trade, not the advertiser's.

What it is

An empirical model of the quality-weighted generalized second-price auction (GSPw) — the mechanism search-ad-auction and google-ad-rank describe, where rank is bid × quality rather than bid alone. Advertisers' private values are recovered nonparametrically from observed bids under incomplete information, without density estimation or tuning parameters, and the estimated value distributions then drive counterfactual auctions under alternative scoring rules.

The point of the machinery: you cannot ask an advertiser what a click is worth to them, and the platform will not tell you what its scoring rule does. Recovering values from bids lets someone outside the platform simulate the rule being changed.

The data

Yahoo! Webscope A3 (Search Marketing bid–impression–click data on competing keywords), covering all non-brand search queries January–April 2008 in five categories: cruise travel, car insurance, laptops, cable TV and collectible coins. 77,850,272 raw observations, cut to 1,610,333 — February only, first two result pages, the top 10% most popular keywords, and days with all ranks 1–14 present. Advertiser and keyword identities are masked but trackable over time.

Note what that means for this spoke: it is Yahoo!, in 2008, not Google today. The mechanism class is the same one still in use; the market, the ranking inputs and the ad formats are not.

Findings

Bidding. Advertisers with higher weighted values shade their bids more, and shading rises with less competition. Across all five categories bid shading is close to zero at the median and climbs rapidly through the upper percentiles — the median advertiser is bidding near value, the top advertisers are not.

Score squashing — the platform's dial. Squashing raises the quality score to a power θ ∈ [0,1], shrinking quality's weight relative to bid as θ falls. The counterfactual (2,000 simulated auctions per category × θ, θ stepped 0 to 1) finds squashing increases auctioneer revenue by 8.5%–24.3%, "at the expense of advertiser profit and consumer welfare," and that the revenue-maximizing θ differs by market — so the scoring rule is a local decision, not a global constant.

The paper quotes Yahoo!'s then-chief economist Preston McAfee on what the dial is for:

"When someone has a really high ad click probability, they're very hard to beat, so it's not a really competitive auction. So that they don't just win [every auction], we do squashing. This makes the auction more competitive. It's like handicapping. We handicap the people with the high click probability."

Quality-score precision is also a dial. Computing scores from user attributes ("user targeting") raises click rates but reduces competition between advertisers, so the platform has a reason to tune how accurate its quality scores are rather than to maximize accuracy. Optimally adjusting that precision raises revenue 0.4%–3.0%.

What this settles for the spoke, and what it does not

Settles: the quality weight is not only a relevance mechanism, it is a revenue instrument, and its magnitude is now estimated by someone outside the platform rather than asserted. Two levers — the squashing exponent and score precision — are shown to trade advertiser profit and consumer welfare for auctioneer revenue, with the sizes attached.

Does not settle: the edge asked how much of a CPC difference an advertiser's Quality Score explains, or where google-ad-rank's thresholds sit. This paper answers the question from the auctioneer's side (what does moving the rule do to revenue), on Yahoo! 2008 data, in five categories. The advertiser-side number the industry repeats — a low Quality Score costing multiples of a high one — is still sourced only to vendor blog posts with no published method.

Related

search-ad-auction · google-ad-rank · gsp-auction-paper · b2b-ppc-metrics · dongwoo-kim · pallavi-pal · synthesis

The advertiser's side. return-on-quality-improvements-sem (2012, T1) measures what a quality gain does to the advertiser's own bill: in 37% of cases it lowered profit, because the gain bought rank and rank costs more.