Do I have enough traffic to A/B test my landing page?
Short answer
Probably not, unless the change you test is large.
To catch a 20% lift on a page that converts at 5%, a standard A/B test needs 8,158 visits per variation, about 16,300 in all, or 16 months at 1,000 visits a month. At a few hundred to a few thousand visits a month, only a large change on an action many visitors take can be read in weeks.
Last checked 25 October 2026 · Agshin Rajabov, founder of Rawuser
Months to finish
16.3 months at 1,000 visits a month
5.4 months at 3,000 visits a month
Last checked 25 October 2026
6 sources, each with its date
What is the minimum sample size for an A/B test?
It depends on how often the page converts today and on the smallest change you want to catch: 80,682 visits per variation at a 2% rate and a 10% lift, 686 at a 10% rate and a 50% lift.
Every cell below uses the same settings: two variations, a 50/50 split, 95% confidence (two-sided) and 80% power. A lift is relative, so a 20% lift on 5% means 5% against 6%. Worked out by us on 25 October 2026 with the formula in the next section.
Double a cell for the whole test. Halve the lift and the visits roughly quadruple.
| Conversion rate today | To catch a 10% lift | A 20% lift | A 50% lift |
|---|---|---|---|
| 2% | 80,682 | 21,109 | 3,826 |
| 5% | 31,234 | 8,158 | 1,471 |
| 10% | 14,751 | 3,841 | 686 |
What is the A/B test sample size formula?
Add 1.96 times the spread you would see with no difference to 0.84 times the spread with the difference, square the sum, and divide by the gap between the two rates squared; the result is visits per variation.
In words: take the two rates you want to tell apart, say 5% and 6%. Multiply 1.96 (95% confidence) by the square root of twice their average times one minus that average. Add 0.84 (80% power) times the square root of each rate times one minus itself, summed. Square the total, then divide by the gap between the rates squared.
We checked it against Evan Miller's sample size calculator on 25 October 2026 by running the formula his page publishes. For 2% and a 20% lift his gives 19,784 per variation against our 21,109; for 5% and a 50% lift, 1,273 against 1,471.
The gap comes from one choice. His uses today's rate in the no-difference term, ours the average of the two rates. Our code with his choice returns his numbers exactly. The choice moves this table by 3% to 16%, which is small next to the gap between a 2% and a 10% page.
n per variation = (1.96 x sqrt(2 x p x (1 - p)) + 0.84 x sqrt(p1 x (1 - p1) + p2 x (1 - p2)))^2 / (p2 - p1)^2
p1 = today's rate, p2 = the rate you want to catch, p = (p1 + p2) / 2How long do you need to run an A/B test at a few hundred visits a month?
At 1,000 visits a month, a 20% lift takes 7.7 months on a page that converts at 10% and 42.2 months on one that converts at 2%.
Months to finish, if every visit to the page enters the test and the action happens on that page. Real tests lose some visits to bots and to people who never reach the button, so treat these as the shortest case.
| Rate today, lift to catch | 500 visits a month | 1,000 a month | 3,000 a month |
|---|---|---|---|
| 2%, 20% lift | 84.4 | 42.2 | 14.1 |
| 5%, 20% lift | 32.6 | 16.3 | 5.4 |
| 10%, 20% lift | 15.4 | 7.7 | 2.6 |
| 2%, 50% lift | 15.3 | 7.7 | 2.6 |
| 5%, 50% lift | 5.9 | 2.9 | 1.0 |
| 10%, 50% lift | 2.7 | 1.4 | 0.5 |
Why not stop the test when it looks significant?
Because checking often and stopping at the first good look turns a 5% false-win rate into a much higher one.
In Evan Miller's example (2010), a test with no real effect, checked as it ran, showed a significant result 26.1% of the time instead of 5%. Fix the number of visits before you start, or decide in advance when you will look.
His simple sequential test (2015) does the second: pick a number of conversions N first; stop when one side leads by 2.25 times the square root of N, or call it even when both sides together reach N. With a real lift it can finish sooner; with no effect it usually takes longer than a fixed test.
Does showing the original to only 1 in 5 visits need fewer visits?
No. For the same change, a split where 1 in 5 visits see the original needs about 1.6 times the visits of a 50/50 split.
In return, 4 in 5 visits get the new page while the numbers build up, which suits a change you already believe in. The small group also collects conversions slowly; at a 2% rate it reaches 10 only after about 2,500 visits in all.
| Rate today, 20% lift | 50/50, visits in all | 1 in 5 on the original, visits in all |
|---|---|---|
| 2% | 42,218 | 67,326 |
| 5% | 16,316 | 25,999 |
| 10% | 7,682 | 12,223 |
What you can do this month, with no tool
Fix what is broken first.
Open the page on a phone and on a slow connection, click every button, and read the headline as a stranger. A button that does nothing or a headline that never says what the product does needs no test.
Count a closer action.
A click on your main button happens more often than a paid signup. At a 10% click rate a 20% lift needs 3,841 visits per variation; at a 2% signup rate, 21,109.
Test a big change.
Change the headline and the offer together; at low traffic only a large lift can be read. At 5% and a 50% lift a test needs 1,471 visits per variation, about 3 months at 1,000 visits a month.
Watch five people use the page.
Jakob Nielsen's model (2000) finds about 85% of usability problems with five similar users. It tells you what confuses people, not which version converts more.
Set the end before the start.
Double the cell for your rate and lift, divide by your monthly visits, and write the end date down. Do not stop early on a good week.
Tag the links you post.
Add
utm_sourceandutm_mediumto every link you share, so you can see which place sends the visits that click, even while a test would take a year.
Sources
- Evan Miller, sample size calculator read 25 October 2026
- Evan Miller, the calculator's published formula (sample-size-fixed.js) read 25 October 2026
- Evan Miller, How not to run an A/B test (18 April 2010) read 25 October 2026
- Evan Miller, Simple sequential A/B testing (13 October 2015) read 25 October 2026
- RevenueCat, effective testing strategies for low-traffic apps (Daphne Tideman, 10 December 2024) read 25 October 2026
- Nielsen Norman Group, why you only need to test with 5 users (Jakob Nielsen, 18 March 2000) read 25 October 2026
Related questions