Free tool
A/B Test Sample Size Calculator
Enter your current conversion rate and the difference you want to detect to see how many people each variant needs. Everything runs in your browser and nothing is sent anywhere.
Conversion of the current version (A). Example: 5 if 5 of 100 visitors sign up
How much of a lift you want to detect. 20% on 5% means a 6% target
Add it to see how many days it takes to collect the sample
- Sample per variant
- 8,158
- Both variants
- 16,316
- Baseline → variant
- 5% → 6%
- Days needed
- —
Which formula does it use?
The normal-approximation formula for comparing two proportions (Fleiss, Levin & Paik, ch. 4) without continuity correction: n per variant = (z(1-α/2)·√(2p̄(1-p̄)) + z(1-β)·√(p1(1-p1)+p2(1-p2)))² ÷ (p2-p1)², where p̄ is the mean of the two rates. The result is rounded up.
Smaller effects need far bigger samples
The sample grows with the inverse square of the difference. Detecting 2 points on a 10% baseline takes 3,841 people per variant; detecting 1 point takes almost four times as many. With little traffic, test bigger changes first.
Don't stop before you reach the sample
Checking every day and stopping on the first significant-looking day inflates false positives. Fix the sample size first, collect it, then decide once. The result is only as good as the conversion counts behind it.
Frequently asked questions
How do I choose significance and power?
A significance level of 0.05 and 80% power are the usual choice. Significance is the chance of seeing a difference that isn't there; power is the chance of detecting one that is. Raising power to 90% needs about 1.3 times the sample.
Should I pick two-sided or one-sided?
Choose two-sided if you also want to know when the variant is worse. One-sided needs fewer people but only looks in one direction. When in doubt, two-sided is the safer choice.
Why do other calculators give slightly different numbers?
Continuity correction and the way variance is estimated change the result by a few percent. This calculator uses the Fleiss formula without continuity correction; corrected formulas give slightly larger numbers.
What if I have three or more variants?
Multiply the per-variant sample by the number of variants, and remember that comparing many variants at once raises the chance of a false positive. Dividing the significance level by the number of comparisons (Bonferroni) is a conservative fix.
To trust an experiment, trust your conversion counts first
funnellog counts signups and payments from server events only. It isn't for sale; it's an operator-only tool.