Multi-Armed Bandit

An adaptive alternative to a fixed A/B test that automatically sends more traffic to whichever variant is winning as the data comes in, balancing “exploit what looks best” against “explore to be sure”. The name comes from imagining a row of slot machines (one-armed bandits) and trying to win the most while learning which pays out best.

Bandits shine when you mainly want to maximise outcomes during the test itself — for example picking the best of many headlines for a one-day news story — because they cut the cost of showing losing variants. The trade-off is that they are worse for clean cause-and-effect measurement than a classic A/B test, so for trustworthy long-term decisions a fixed-split experiment is usually still preferred.