Experiments
Auto-optimization
Use multi-armed bandit allocation to shift traffic toward stronger variants.
Auto-optimization uses Thompson Sampling to update allocation as conversions arrive. It favors stronger variants while continuing to learn. This differs from a fixed A/B test: its purpose is reducing opportunity cost during the run, not preserving a constant split.
Choose the confidence requirement and optional automatic stop before launch. Monitor update notifications and primary-goal quality. Sparse, delayed, or incorrectly tracked conversions can make any adaptive decision unreliable.