Experimentation

Most A/B testing programs produce confident conclusions from insufficient data. The discipline is almost entirely in what you do before launch.

Before running

• Hypothesis with a mechanism. "Moving the pricing table above the fold will raise trial starts, because visitors currently leave before seeing pricing." Not "let's try a green button."
• One primary metric, chosen in advance. Secondary metrics are context, never the verdict.
• Sample size calculated in advance, from your baseline rate and the smallest lift that would change a decision. If the required sample is unreachable, do not run the test — decide by judgment and say so.
• Duration set in advance, covering at least one full weekly cycle, and two if the buying cycle is long.
• Guardrail metrics that would make you reject a win: refunds, support volume, downstream retention.

While running

Do not look at results and act on them mid-flight. Peeking and stopping at significance is the single most common way to generate false positives, and it is very effective at it.

Check only that the test is running correctly — even split, no broken variant, tracking firing.

Reading

• At the pre-set duration, not before, and not extended because it is nearly significant. Extending until significance manufactures it.
• Significance is not size. A statistically significant 0.3% lift may not be worth shipping.
• Inconclusive is a real result and the most common one. It means the change did not matter enough to detect, which is useful.
• Check the guardrails before declaring a win.
• Segment afterward for hypotheses only, never for verdicts. Slice enough ways and something is always significant.

Program level

Test where the traffic and the leverage are. Most sites can only run a handful of adequately powered tests a year — spend them on structural questions, not button colors.

Keep a log of every test: hypothesis, result, decision. Without it, teams re-run the same tests every eighteen months and re-learn the same things.

Sources

references/sources.md in this skill lists the outside authorities that settle the questions here — what each one is authoritative for, and what you may do with it. Check them before answering on anything they cover, and cite what you used. Most are free to read and not free to reproduce; the use note on each is binding.

Tooling

Client-side and web testing: Optimizely, VWO, AB Tasty, and similar. Warehouse- or product-native: GrowthBook, Statsig, Eppo, PostHog, and similar — these compute against your own event data, which is what you want once the metric definitions matter.

Feature flags are the server-side path to the same thing: LaunchDarkly, Unleash, Split, and similar. Running an experiment behind a flag you already use for release control is cheaper than adding a second system, and technology:release-and-deployment covers the release side of it.

No tool fixes an underpowered test. The platform reports a result either way, which is exactly the risk.

Never

• Stop a test because it reached significance early. Peeking until it looks conclusive manufactures the result.
• Run a test that cannot reach adequate sample size in a reasonable window. Ship the change on judgment instead and say so.
• Change more than one variable and attribute the outcome to the one you liked.
• Count a flat result as a failure. A well-run test that rules out a plausible idea has bought information.