Blog

Non-significant doesn't mean 'no effect.' It means 'not enough evidence yet.' Here's the five-step diagnostic that tells you whether to keep running, replan, ship anyway, or call it null — and which one your test actually is.

CUPED uses each user's pre-experiment behavior to subtract predictable noise from your metric. Done right, it shrinks required sample size by 30–50% with no extra traffic, no bias, and no extra weeks on the calendar. Here's the intuition, the math, the numbers, and the four ways teams get it wrong.

Most A/B tests fail before they start because the sample size was wrong. Here's how to plan one that has a real chance of detecting the lift you care about — with the math, the trade-offs, and a calculator.

P-value, confidence interval, lift, power — what each one means, what each one doesn't mean, and the order to read them in so you don't fool yourself.

MDE is the smallest lift your test can reliably detect. Pick it too small and your test runs forever; pick it too big and you'll miss real wins. Here's how to choose it like a grown-up.