Waterfall to In-App Bidding Migration, With a Holdout

Migrations get reported at plus fifteen to thirty percent because the team compares the fortnight after with the fortnight before, and the fortnight after happens to contain an acquisition push, a seasonal peak, or a release with a new placement in it. The number I hand you comes from a randomised holdout held for twenty-one days. It is usually smaller and it survives the first question a CFO asks. The read: Users split at install by a stable hash, never by country or device tier. The control arm keeps the waterfall you run today. Every decision after that is judged on revenue per daily active, with impressions per active and eCPM reported separately, because those two move in opposite directions and a stack that lifted eCPM while cutting impressions per active by a fifth has cost you money while looking like a win. Latency sits in the same report: bidder timeouts, request to fill at the median and the ninety-fifth percentile, and the share of interstitial requests ending in no fill on a session-two return. Order of operations: Bidding goes on one format first, usually rewarded, because it carries the clearest user-side signal and the least exposure if fill drops. Interstitial follows. Banner is last, and sometimes never, in apps where the banner is a fixed-price direct deal worth more than an auction will pay for it. A small number of waterfall lines stay pinned above the auction throughout: the direct sales, the seasonal apparel deals booked months in advance, the network that reliably overpays in Canada. There is no prize for a stack that is purely programmatic. What I refuse: I will not run a migration through a fourth-quarter retail peak or across an acquisition burst, because neither the holdout nor your finance team will trust what comes out of it. I will not switch every format in one release; the rollback path matters more than the calendar does. I will not report a lift that has not survived the holdout, even when the pre-post figure is friendlier and you would rather see it. And I will not drop a network for bidding low on average, because average is the wrong statistic for a bidder that wins your best users in one market and abstains everywhere else. Not for you if: You already run bidding on every format and want the last few points squeezed out. That is floor work, a different engagement, and buying this one would be paying migration prices for it. Also not for you if your mediation SDK is more than four versions behind, in which case the upgrade is the project and the migration is one step inside it. What you get: A branch with the changes and a remote config flag per format. A twenty-one day holdout report covering revenue per active, impressions per active, eCPM, fill and latency for both arms, split by your three largest markets and by platform. A written decision on which waterfall lines survive and at what pinned price. And a note on what the migration does to your acquisition bids, because ad revenue entering the predicted LTV model changes what you can afford to pay for an install, and that is frequently worth more than the lift itself. Six weeks: one to instrument, one to build, three to hold, one to write up and hand over. If the holdout says the migration is worth less than a point, I will tell you to keep the waterfall and the engagement ends there, with the instrumentation still yours.