Free
For anyone who wants to check what has been tested before.
- Every test, learning and idea
- Rigor and evidence scores
- Growth Toolbelt and Playbooks
- Save items to your workspace (free account)
Real A/B tests from Duolingo, Booking.com, Netflix and 93 more, each scored for rigor and pooled into learnings with a confidence interval. The losers stay in.
Free to browse. No account needed.
Tests that ran at
A screenshot and a "+40%" is a story, not evidence. You can't tell a well-powered test from a lucky week, and you never see the ten that failed. That is how teams ship ideas that already lost somewhere else.
Filter by funnel stage, surface, team and the behavioural mechanism the change relies on. Paywall, onboarding, checkout, cancel flow: see what was tried, by whom, and what happened.
Every test gets a rigor score from the stats it publishes. A score is capped by its evidence tier, so a "big lift" post with no numbers can never outrank a test with a sample size and an interval.
Learnings combine every linked test, weighted by rigor, freshness and independence. You get one effect with an interval, how much the tests disagree, and the next test that would settle it.
Every number on this page is computed from the library itself, and every result links to where it was published.
results link to the original write-up: engineering blogs from Netflix, Microsoft ExP, Booking.com and Duolingo, plus GoodUI, abtest.design, Built for Mars and more.
Tier A is first-party data. Tier D is a claim with no figures, capped at a rigor of 25.
For anyone who wants to check what has been tested before.
For growth and product people who decide what to test next.
For experimentation teams building their own evidence base.
Paid plans are opening gradually. Requesting one reserves your place and charges nothing; we email you before anything is billed.
From write-ups the companies or their agencies published: engineering blogs such as Netflix, Microsoft ExP, Booking.com and Duolingo, and curated collections such as GoodUI, abtest.design and Built for Mars. Every record links to its source, and screenshots from abtest.design are credited to it.
From what the test published: sample size per variant, p-value or confidence interval, how long it ran, a sample ratio mismatch (SRM) check, and whether it was pre-registered. The score is capped by evidence tier: 100 for first-party data, 85 when full stats are published, 50 for an effect with no stats, and 25 for a claim with no figures. Records store only the inputs, so when the formula changes, everything is re-scored.
A test is one experiment at one company. A learning is a reusable claim, such as "show value before asking for sign-up", that several tests support or contradict. Its evidence score grows with each linked test, weighted by rigor, freshness and how independent the tests are, and falls when they disagree.
No. Browsing every test and learning is free and needs no account. A free account adds a workspace: saved tests and learnings, what you viewed recently, and a short getting-started path. Sign-in is by email link, so there is no password to manage.
No. Paid plans are opening gradually. Requesting one records your interest on your account; nothing is billed, and we email you before it is.
Yes. Pro includes the library as structured JSON and agent skills that review, classify and group tests, so Claude Code or Cursor can check prior evidence before building a paywall or an onboarding step.
Not necessarily, and the library is built to show that. Learnings report how much tests disagree, older results lose weight as they age, and lab findings carry replication discounts. Use it to choose better tests and set realistic expectations, not to skip testing.
Look it up first. 175 tests and 61 learnings, free to browse.