Better AI predictions don't always improve ad auction outcomes

Researchers at ICML 2026 have introduced the concept of "model monotonicity" in autobidding auctions, addressing a critical gap between ML model quality and auction outcomes. The paper, led by Ashwinkumar Badanidiyuru, defines model improvement as a refinement relation based on probability filtrations — essentially, when a model's predictions become more granular and accurate. The study systematically evaluates how these improvements affect platform-level metrics like revenue, welfare, and liquid welfare across different auction formats.

The key result: First-price auctions with uniform bidding guarantee revenue monotonicity for tCPA (target cost-per-acquisition) bidders without budget constraints, thanks to Jensen's inequality. However, second-price auctions and the introduction of budget constraints can break monotonicity — meaning better predictions can paradoxically reduce revenue. The findings have immediate practical implications for ad platforms like Google, Meta, and Amazon, which must align their ML model upgrades with auction mechanisms to avoid unintended negative business outcomes.