Price Floors Rebuilt on Bid Landscape Data, With Holdouts

What a floor is: A floor is a bet that the bid you refuse would have been beaten by the next one. We size that bet from your own bid landscape, meaning every bid response the wrapper received rather than the winners. On the accounts we have rebuilt, the revenue-maximising floor sits between the 55th and 70th percentile of the bid distribution for a given slot, device and geography, and the spread between segments is wide enough that one site-wide floor loses money in both directions at once: refusing fill in the thin segments, leaving margin on the table in the rich ones. The arithmetic: Per segment, expected revenue is the floor multiplied by the observed probability that a bid clears it, minus what the unfilled impression actually returns. That last term is where most floor work is sloppy. A passback or a house creative is not worth zero and it is not worth the same on every page, and until you have priced it you cannot tell a good floor from a greedy one. Since the move to first price the auction mechanism no longer protects you, and buy-side shading means the bids you are reading already encode a guess about your floor. Set one badly and you are training that guess. Segments we cut on: Slot and size, device, geography, session depth, refresh index, and consent state. Consent state is where most setups go wrong: non-consented traffic clears lower, and the correct response is a lower floor there, not a higher one to recover the difference. We add recency only where the volume supports a decision. Implementation: A price floors rule set with an explicit minimum, published as a JSON file we refresh weekly, with enforcement behaviour written down rather than left at defaults. On the Google side, unified pricing rules aligned so that the constraint that binds is the one you chose; a wrapper floor sitting under a higher unified pricing rule is decoration. You get the generation query, so your analyst can rerun the whole thing next quarter without us. How we protect you from us: Ten percent of traffic stays at your existing floors for the length of the engagement. Every weekly change is read against that holdout rather than against last week, because last week had different demand. We do not ship a floor change in the same week as a layout change or a demand partner change, and there is a one-step rollback. When the holdout wins, you hear it from us first. What we refuse: We do not take a floor recommendation from a demand partner as an input; their optimum is not yours and they are not shy about it. We do not resell a black-box floor product whose output we cannot audit. And we will not touch floors before your ads.txt and supply chain declarations are resolved, because until then the landscape you are measuring includes paths that are not yours and every percentile in it is fiction. Not for you if: Below roughly two million monthly ad impressions the segments will not reach a decision inside a quarter and you would be buying noise; the same money spent on layout or on direct demand will do more. Also not for you if unsold inventory is politically unacceptable in your business. Floors work by refusing bids, so some impressions go unsold. That is the mechanism, not a defect in it.