GrowthBook

AI Evals tell you if the model works. A/B tests tell you if users care. These answer different questions. Evals measure competence—can it do the job? A/B tests measure value—do users actually behave differently? DoorDash saw a model that looked great in testing drop 4.3% in accuracy with real users. The test data was too clean. The teams shipping GenAI reliably treat these as stages: evals catch regressions early, shadow mode tests real traffic, A/B tests prove business impact. You can't A/B test a broken model. You can't eval your way to product-market fit. Wrote up the full pipeline: https://lnkd.in/gc_yVfYw