A model can pass offline evaluation and still fail in production. Offline evaluation measures performance on historical data. It does not measure how users, markets, reviewers, or downstream systems react when the model changes the live experience.

Online evaluation fills that gap. It tests models on real traffic, measures real outcomes, and catches failures that static datasets cannot reveal.