Sample size calculator. How many visitors your A/B test needs before the result means anything, and how long that takes at your traffic.
Assumes a two-sided test comparing two independent proportions, a 50/50 split between control and variant, one variant against one control, and a single analysis at the end rather than peeking as results come in. It uses the standard normal approximation, stated in full on this page. Everything is computed in your browser.
In short
How many visitors does an A/B test need?
It depends on your baseline conversion rate and the smallest improvement worth detecting. At a 3% baseline, detecting a 20% relative lift with 95% significance and 80% power takes about 13,900 visitors per variant, so roughly 27,800 in total. Smaller effects cost far more traffic: halving the effect roughly quadruples the sample.
Which test does this assume?
A two-sided z-test comparing two independent proportions, using the standard normal approximation. Per variant:
n = ( z1-α/2× √(2 × p̄ × (1 - p̄)) + z1-β × √(p1(1 - p1) + p2(1 - p2)) )2 ÷ (p2 - p1)2
p1 is the baseline rate, p2is the rate you want to be able to detect, and p̄ is the average of the two. The first term uses the pooled variance under the null hypothesis, the second the unpooled variance under the alternative. That pairing is what published sample size calculators use, and it is slightly more conservative than the familiar shortcut of 16 × p × (1 - p) ÷ delta².
The z values are exact standard-normal quantiles for the significance and power levels offered, so nothing is rounded there. The approximation lives in the normal model itself. It holds well at ordinary web conversion rates and stops being trustworthy when a variant would produce only a handful of conversions in total, which is why the tool warns you below about ten expected conversions per arm.
Four things it assumes and you should check: the test is two-sided, so you are asking whether the variant differs rather than whether it is better; traffic is split evenly between the two arms; there is one control and one variant; and you read the result once, at the end, rather than watching it daily and stopping when it looks good.
Compared
What each input actually costs you. Every knob on the calculator trades traffic against certainty. This is what each one buys, holding everything else at a 3% baseline, 95% significance, 80% power.
| Change | Visitors per variant | What it buys you |
|---|---|---|
| 20% relative effect (the default) | 13,914 | The reference point |
| 10% relative effect | 53,211 | Catches smaller wins, at nearly four times the traffic |
| 30% relative effect | 6,455 | Cheap, but only detects large changes |
| 99% significance instead of 95% | 20,704 | Fewer false winners, at about 1.5x the traffic |
| 90% power instead of 80% | 18,626 | Misses a real effect 1 time in 10 rather than 1 in 5 |
| 6% baseline instead of 3% | 6,719 | Higher baselines need less traffic for the same relative lift |
Figures computed with the formula above, rounded. Reproduce any row by entering it into the calculator.
You still need the baseline.
Every number this calculator produces depends on the rate your control converts at today. Guess that wrong and the sample size is wrong with it, usually in the direction that makes the test end too early.
Take the baseline from the same funnel step, the same traffic mix, and a period long enough to cover a full weekly cycle. If your rate swings between 2% and 5% depending on the week, size the test on the low end and expect it to run longer than the estimate.
Where tests go wrong.
Peeking and stopping early
Stopping at the first significant reading pushes your real false positive rate past 20%.
Running less than a week
Tuesday traffic is not Sunday traffic. Run whole weeks so both arms see the full cycle.
Sizing for a lift you will not get
Most copy changes move conversion a few percent, not twenty. Size for the realistic ceiling.
Counting bots in both arms
Crawlers land in control and variant and convert in neither. On one live mrkr site they ran at 4.1x human pageviews in September 2026.
Testing many variants at once
Four variants is four comparisons, and the odds one wins by luck climb fast.
Treating a null result as failure
A powered test that finds nothing has told you the change is not worth shipping.
Live demo
Where your baseline comes from. Sizing a test starts with knowing the rate you have now, per step and per source. The mrkr demo shows funnels, goals and events on a fully populated dashboard, with no signup.
Open the live demoWhat if you do not have the traffic? If the calculator says 14,000 per variant and you get 3,000 visitors a month, an A/B test on that page is not a tool you can use. Three honest alternatives.
Test earlier in the funnel
A step with more traffic and a higher base rate reaches significance far sooner.
Test bigger changes
Sample needed collapses as the effect grows. Two offers, not two headlines.
Use qualitative evidence
Twenty session replays of an abandoned form beat a test you cannot power.
FAQ
How many visitors does an A/B test need?
It depends almost entirely on your baseline conversion rate and how small an effect you want to detect. A 3% baseline and a 20% relative improvement, at 95% significance and 80% power, needs roughly 13,900 visitors per variant. Halve the effect you want to detect and the sample needed roughly quadruples.
What formula does this calculator use?
The standard normal approximation for two independent proportions: n = (z(1-alpha/2) * sqrt(2 * pBar * (1 - pBar)) + z(1-beta) * sqrt(p1(1-p1) + p2(1-p2)))^2 / (p2 - p1)^2, per variant. It assumes a two-sided test, an equal split between control and variant, and one analysis at the end.
What is the difference between significance and power?
Significance is your tolerance for calling a winner that is not real, so 95% significance means a 5% false positive rate. Power is your chance of detecting an effect that is real, so 80% power means you would miss a genuine improvement of the size you specified one time in five.
Should minimum detectable effect be relative or absolute?
Relative is how most teams talk, as in a 20% lift. Absolute is in percentage points, as in 3% to 4%. They are very different asks at low baselines: on a 3% baseline, a 1 point absolute gain is a 33% relative gain. This calculator takes either and shows you the resulting variant rate.
Can I stop the test as soon as it looks significant?
No. Checking repeatedly and stopping at the first significant reading inflates the false positive rate well beyond the 5% you asked for, often past 20%. Fix the sample size before you start, and read the result at the end. Sequential testing methods exist for peeking, but they are not what this calculator computes.
Does this work for more than two variants?
Not directly. The formula compares one control against one variant. Testing three variants against a control means three comparisons, so you need to tighten significance to keep the overall false positive rate at 5%, which increases the sample per arm. Running fewer variants for longer is usually the better trade.
Start with one line of HTML.Your first visitors show up within a minute.