What the IBPA Numbers Actually Show
The empirical question is not whether a proposed auction sounds more sophisticated than GSP. It is what the reported IBPA numbers compare, which parts come from observed retail-media data, and which parts come from estimates and simulation. Keeping those layers separate lets a reader judge the size of the result without mistaking a counterfactual mechanism for a production cost report.
What was observed, and what was estimated
The auction-design paper studies a large retail-media platform over the period May 17 to July 14, 2025, using all position auctions in one product category during that interval. The observed sample contains 9.6 million auctions, 1,207 advertisers, 72 segments, and 44.4 million submitted bids. Most auctions contain either eight or twenty slots. In the platform environment represented by the data, the auction used generalized second-price (GSP) with full disclosure: advertisers saw the realized segment before bidding, were ranked by bid multiplied by quality score, and paid the minimum bid needed to retain their assigned position under the platform's rule.
Those are observations about the data-generating environment, not observations of IBPA. No publisher is reported as running the proposed mechanism in production. The paper uses the existing auction to estimate the ingredients needed for a counterfactual: the distribution of segments, advertiser quality effects, slot effects on click-through rates, and advertiser valuation distributions.
The segment structure combines audience and device information. The platform's selected dimensions include three consumer-behavior categories, eight demographic categories, and three device types, creating 72 possible combinations. These categories are not just descriptive labels in the paper's model. They define the states over which an advertiser's value can vary and therefore the dimensions that a richer auction might use.
The click-through-rate ingredients come from a separate dataset covering the same period. It contains daily observations of impressions and clicks by advertiser and slot position. The authors estimate multiplicative slot and advertiser effects, imposing a monotonic relationship in which higher positions have at least as large a slot effect as lower positions. Segment effects are set to one because the click data do not break clicks out by segment; that normalization is applied uniformly across the counterfactual mechanisms.
Valuations require a more indirect step because GSP is not truth-telling. Observed bids cannot simply be treated as true willingness to pay. The authors use inequality bounds implied by envy-free GSP conditions to create intervals containing the underlying valuations, then estimate segment-level valuation distributions with a non-parametric Turnbull estimator. They assume advertisers are ex ante symmetric within the one product category so that interval observations can be pooled across auctions and advertisers. The resulting distributions are estimated primitives, not directly observed valuations.
That distinction is central. The data show that particular auctions, bids, quality scores, segments, positions, impressions, and clicks existed in a particular sample. They do not directly show what those advertisers would have bid under a mechanism they never used. The counterfactual depends on the estimated distributions and on the model that maps them into alternative outcomes.
How the counterfactual was constructed
The paper compares the proposed mechanism with two kinds of benchmark. One is the optimal scalar-bid mechanism under different disclosure regimes. The other is GSP under full or null disclosure, because GSP remains a common format in search and retail media.