LaunchRanked

GlossaryGrowth & analytics

What is an A/B test?

Also called: split test, A/B testing, split testing, multivariate test

Definition

An A/B test randomly splits visitors between two versions of a page or feature, A and B, and compares a chosen metric, such as sign-up rate, to see which version performs better beyond chance.

A/B test, explained

The mechanics are simple: decide what you're testing and which metric decides the winner, split traffic randomly, run until you've reached the sample size you planned, then compare. The hard parts are statistical and practical.

The biggest statistical trap is peeking. If you check results repeatedly and stop the moment one version looks significantly better, you'll declare many false winners. Evan Miller's well-known essay "How Not To Run an A/B Test" shows how badly repeated peeking distorts significance levels. Decide the sample size in advance, or use a testing method designed for continuous monitoring.

The practical trap is running tests without enough traffic. Detecting a modest improvement in a low conversion rate can require tens of thousands of visitors per version. Most early startups don't have that. At low traffic, test big, bold changes (a different headline and offer, not a button color), or skip testing and learn from user interviews and session recordings.

Tests on public pages also need to respect search engines. Google's guidance on website testing is specific: don't cloak, meaning don't show Googlebot a different version than users see; if variants live on different URLs, use rel="canonical" pointing to the original; use 302 (temporary) redirects rather than 301s for test redirects; and run experiments only as long as necessary, then remove the test setup.

Record every test: the hypothesis, the variants, the result and what you decided. The log becomes a guide to what your audience responds to.

Why it matters for founders

A/B tests replace opinions with evidence, but only when run correctly. Bad tests produce false winners that you then build on.

Example

A startup with 40,000 monthly visitors tests a new pricing page layout, planned for two weeks at a fixed sample size, with the variant on a separate URL canonicalized to the original. Sign-ups rise, and the result holds after launch.

Common mistakes

  • Stopping a test the moment it looks significant.
  • Testing tiny changes on too little traffic.
  • Showing Googlebot a different version, which is cloaking.

Sources

Checked

← All 274 glossary terms