A product manager wants to run an A/B test next week. How do you work out how long it needs to run?

Work backward from the minimum detectable effect. You need the baseline rate, the smallest lift worth shipping, the significance level, and the desired power (usually 80 percent). Those give a required sample per arm; divide by daily eligible traffic to get days. Then round up to whole weeks so you cover weekday cycles, and fix the duration before you start.

I ask four things: what is the current conversion rate, what is the smallest lift that would actually make us ship this, what alpha are we using, and what power do we want. The second one is the real conversation, and it is a business question rather than a statistical one. If they say any lift at all, I show them the math: at a 3 percent baseline, detecting a relative half percent takes millions of users per arm, and we do not have that, so we would be running a test that cannot succeed. Usually they land on something like a 5 percent relative lift, which at our traffic might be eleven days. Then I round up to fourteen so we cover two full weekly cycles, because weekend users behave differently and a ten day test overweights whichever days it happens to include. And I write the stop date down before launch. The habit that kills teams is checking daily and calling it the moment p dips under 0.05, which badly inflates the false positive rate. If they genuinely need early reads, I set up a sequential test properly instead of pretending peeking is free.