Multi-armed bandit (MAB)

A multi-armed bandit is an algorithm that allocates traffic across two or more variants dynamically, sending a growing share to whichever is currently performing best, rather than splitting visitors evenly for the whole test as in classic A/B testing. The name comes from a gambler choosing among slot machines ("one-armed bandits") with unknown payout rates while trying to maximize total winnings.

Common algorithms include epsilon-greedy, which mostly plays the current best arm but occasionally explores at random, and Thompson sampling, which samples from a probability distribution over each arm's true performance and tends to be more traffic-efficient. This trades some statistical rigor for lower opportunity cost: a bandit minimizes the traffic sent to underperforming variants ("regret"), while a fixed-split A/B test prioritizes a clean, unbiased estimate of the effect size.

Bandits suit short-lived decisions, such as picking a headline or an image, where the goal is the best outcome now rather than a precise, generalizable effect estimate; A/B testing remains the better choice when the goal is measuring a durable causal effect. A frequent pitfall is using a bandit for a question that needs a defensible effect size, since its adaptive allocation biases naive estimates of the difference between arms.