Price Floor Tuning Judged on ARPDAU, Not eCPM

Raising a price floor raises eCPM. It does that reliably, which makes floor work the easiest thing in app monetization to report as a success and one of the easiest to get wrong. The pattern across the floor tests I have run: a floor moved far enough to lift eCPM by ten percent takes eight to fourteen percent of impressions with it, and revenue per daily active lands flat or slightly down. This engagement is the search for the cases where that trade goes the other way. There are fewer of them than any network's account manager will tell you. Where floors actually pay: By country, and not by the tier list the networks publish. A floor that is right for the United States leaves money behind in Norway and chokes fill in Brazil, so the split is drawn from your own bid density, the share of auctions with more than two bidders clearing the current floor, rather than from a ranking of GDP. By format, because rewarded tolerates a high floor and a banner does not. By segment where the stack supports it: a cohort that will never subscribe can carry a higher floor than one still inside a trial window, and in a resale app the seller cohort and the browsing cohort are two businesses sharing a bundle id. The protocol: Every change runs as a split in MAX or through your own remote config, one variable at a time, with a fourteen-day read and no decision before day ten. Ramadan, back-to-school and the four weeks after American Thanksgiving are excluded windows, because demand curves move enough inside them to produce whatever conclusion you were hoping for. The report carries a per-market table with floor, eCPM, impressions per daily active, fill, and revenue per daily active with a confidence interval on the last one. Where that interval crosses zero I say so, rather than picking the flattering side of it. Out of scope: I do not negotiate rate cards or chase a network's promotional rate for the quarter. I do not tune floors on a stack whose impression-level data I cannot read, and where that data does not exist, building it is the first fortnight. I do not touch the paywall here, even when the floor problem turns out to be a paywall problem. I will tell you that it is, and it becomes a separate piece of work with its own scope. Who should not buy this: Anyone who needs a number for a deck by Friday. The fourteen-day read is not negotiable and no version of this produces a trustworthy answer in four days. Anyone running mostly direct-sold inventory, where the floor is a contract term rather than an auction parameter. And anyone whose stated goal is beating a competitor's published eCPM, which is a statistic about their traffic mix rather than about anybody's skill. The ongoing version: Floors decay. Demand shifts each quarter, a bidder changes its shading, a network loses a large advertiser and stops competing in your third-largest market. The retainer reads bid density monthly, runs one test at a time, and reports what moved and what I deliberately left alone, including the months where nothing needed changing. In a well-set stack that is most months, and a report saying so is worth more than a change made to look busy. You keep the test log: every floor tried, the date, the market, the result and the reasoning. It is the document that stops your successor re-running the same test in eighteen months and calling it an insight.