Bid-floor optimization for publishers in mobile ad auctions is difficult: bids are heavy-tailed, user-level signals are sparse, and publishers observe only partial auction feedback. Mediation infrastructure also supports auction retry paths, where an auction that fails to fill at one floor can be attempted again at a lower floor. These retry paths complicate floor setting, but also create an opportunity: publishers can target high-value demand while preserving a low floor safety path.
We study this problem through FloorGym, a publisher-side simulator calibrated on production data. We show that single-level bid floors often perform poorly under heavy-tailed first-price bids: floors high enough to extract price also lose enough fill to cancel out the gains. We show using simulations that a cascaded bid-floor policy can mitigate this trade-off by attempting an aggressive learned primary floor first and, if it does not fill, reissuing the auction opportunity at a near-zero safety floor. We evaluate this cascaded bid-floor design in production through a seven-day A/B test on 44 mobile games from the portfolio of Game District, a large mobile-game publisher, covering approximately 5 million daily active users. Treatment increased average revenue per daily active user (ARPDAU) by +13.2% relative to incumbent floor strategies. The effect was primarily price-led: revenue per impression increased by +9.3%, while impressions per daily active user (ImpDAU) increased by +3.7%.