How long should an A/B test run?
How long an A/B test should run, planned before it starts, in whole weeks, with the smallest change your traffic can detect worked out for 2, 4 and 8 weeks.
Read the guideEnter how many people would see the test each week and how often they convert. The table shows the smallest change a test of each length can find.
An A/B test splits people 50/50 between the original (A) and one change (B). The calculator asks how big a difference between them a test can tell apart from chance, with a two-sided test at 95 % confidence and 80 % power.
It uses the arcsine method, the same maths as the planner in MIRA FIVE: B must reach a rate of sin²(asin√p + 1.4 × √(2 / n)), where p is today’s rate and n the people in each variant. The smallest change is that rate against p.
At 4 % and 1,000 people a week, a 2-week test finds only a change of about +70 %; 4 weeks about +48 %; 8 weeks about +33 %. Finding +10 % would take about 39,400 people per variant. Plan the length before you start, run whole weeks, and decide once at the end.
The calculator assumes every person is new each week. MIRA FIVE’s planner projects from your own last 28 days, so returning people count once.
In MIRA FIVE you pick 2, 3, 4, 6 or 8 weeks, see what the test can find before it starts, and get one verdict at the planned end plus its window, on the goal you chose.
How long an A/B test should run, planned before it starts, in whole weeks, with the smallest change your traffic can detect worked out for 2, 4 and 8 weeks.
Read the guide