Multi-armed bandit

An assignment strategy that shifts traffic toward better-performing variants as a test runs, instead of a fixed split.

A bandit strategy allocates traffic dynamically instead of holding a fixed split for the whole test: as evidence accumulates about which arm is performing better, more traffic shifts toward it. Common implementations are ε-greedy (mostly exploit the current best arm, occasionally explore the others) and Thompson Sampling (sample from each arm's estimated performance distribution and let probability do the shifting).

The tradeoff against a fixed-horizon factorial test is real: a bandit optimizes continuously rather than giving you one clean, fixed-sample statistical read at the end, which makes it a better fit for a high-traffic page you want to keep improving than for a one-shot decision you need to defend with a p-value. TraqLyte names this strategy as planned but not yet implemented.