Skip to content

Stockertim Dictionary · Artificial intelligence · Business

Multi-armed bandit

Also called bandit testing, bandit algorithm, multi-armed bandit testing

In short

An adaptive testing method that shifts more traffic towards the better-performing versions as results come in, while still giving weaker versions a small share.

  • Advanced
  • Process
  • Checked

In detail

A multi-armed bandit is a way of testing several versions of something, such as ads or landing pages, that adapts while the test runs. The name comes from a gambler facing a row of slot machines, each with an unknown payout, who must balance trying different machines against playing the one that seems to pay most.

A classic A/B test splits traffic evenly until the end and then picks a winner. A bandit instead updates the split as results arrive: versions that convert well receive more visitors, and weaker ones receive fewer. It keeps a small share of traffic exploring the weaker versions in case early results were luck. This trade-off is called exploration against exploitation.

The benefit is less traffic spent on poor versions during the test, which suits short campaigns where there is little time to learn and then act. The cost is weaker evidence: because the groups end up very different in size and are exposed at different times, it is harder to say with confidence how much better one version is.

Many advertising and testing tools use bandit-style methods inside their automated optimisation. Teams that need a clear, reportable answer about which version works better still run a controlled test with a fixed split.

Worked example

Lagoon Logistics tests three landing pages with 9,000 visitors a day. On day one each page gets 3,000 visitors.

  1. Day one conversion rates: 60 / 3,000 = 2%, 90 / 3,000 = 3%, 45 / 3,000 = 1.5%
  2. The bandit moves day two traffic to 20%, 70% and 10%.
  3. Day two visitors to the second page: 9,000 x 70% = 6,300
  4. Day two visitors to the first and third pages: 9,000 x 20% = 1,800 and 9,000 x 10% = 900

The page converting at 3% gets most visitors on day two, while the other two still receive 2,700 visitors between them in case the early results were noise.

In practice

Bandits are used for headline tests on news sites, short promotional campaigns, recommendation slots and ad creative rotation. Teams set a minimum share for every version, watch for results that change by time of day and keep a fixed-split test for decisions that need firm evidence.

Watch out

  • A bandit is built to earn while testing, not to measure precisely; its final split is not proof of a winner.
  • Early luck can starve a good version of traffic if the minimum exploration share is too small.

Learn it properly

Official sources

Facts last checked on 3 November 2026.

Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.

English