Why do A/B tests usually run for whole weeks rather than an arbitrary number of days?

## The weekly cycle Traffic to almost any consumer product rises and falls on a seven-day cycle. The volume differs by day, and so does the *composition* of the audience: weekday daytime visitors are often at work, browsing on desktop, in short sessions; weekend visitors skew towards longer, leisure-driven sessions and different device mixes. Conversion rate, average order value, session length and retention all typically move with the day of week. A business-to-business product usually shows the mirror image, with weekends nearly empty. That cycle is the reason experiment runtimes are quoted in weeks rather than days. ## What a partial week actually costs It is worth separating two claims that are often mashed together, because interviewers probe exactly here. **It does not bias the treatment-versus-control contrast.** Units are randomised into arms concurrently, so Tuesday's visitors are split between arms by the same mechanism as Saturday's. Both arms therefore see the same day mix, whatever that mix is. Someone who says "the control got the weekend and the variant did not" has misunderstood how simultaneous randomisation works; that failure belongs to designs that compare this week against last week, not to a concurrent A/B test. **It does damage generalisation and precision.** Two real costs remain: 1. *Heterogeneous effects.* If the effect differs by day — a delivery-promise banner that matters far more when people shop for the weekend, a checkout change that matters more on mobile-heavy evenings — then a Tuesday-to-Thursday window estimates the mid-week effect and calls it the effect. Ship on that basis and the realised impact over a full week can be materially different, in either direction. 2. *Population mismatch.* A short window over-represents your most frequent visitors, because heavy users show up first and light users trickle in. That is a different audience from the one the rollout will serve. A third, smaller consideration: day-to-day variance is part of the noise in the estimate, and a lopsided window can make the observed variance unrepresentative of the variance the sizing calculation assumed. ## Whole multiples, not round numbers "Whole week" means a multiple of seven consecutive days, not a window aligned to Monday. A test launched Thursday and stopped after the following Wednesday has covered each weekday exactly once and is perfectly balanced. What breaks balance is a duration that is not a multiple of seven: a 10-day test starting Monday covers Monday, Tuesday and Wednesday twice each and the other four days once, so early-week behaviour carries roughly double weight in the pooled estimate. Ten days sounds more conservative than seven and is in fact less balanced. Fourteen days is the common default because it covers two full cycles, smooths week-to-week wobble, and is long enough to be robust to one odd day. ## Committing before launch The runtime belongs in the test plan alongside the sample target, written down before the first user is exposed. Two things follow from that. First, the plan should state the calendar end date, not just a sample count, so that the sample and the calendar are both preconditions for reading the result. Second, the plan should note any known calendar hazard in the window — a promotion, a holiday, a marketing push — because those decisions are much cheaper before launch than after. ## When the rule bends The rule is a heuristic about metric behaviour, not a law. If you can show from historical data that the metric has no meaningful weekly pattern — some internal tools, some low-frequency flows — a non-multiple window is defensible, though you should still commit to it in advance. Conversely, if the metric has a monthly cycle, such as payday-driven spend, seven days is not enough and the runtime argument scales up accordingly: you want the window to cover whole cycles of whatever periodicity your metric actually has. ## What a strong answer sounds like Name the weekly cycle, explain that whole multiples give each weekday equal weight, and then show the extra layer: concurrent randomisation already balances the day mix across arms, so the real prize is an effect estimate that generalises to a normal week, plus a stable variance. Finish by saying the duration was fixed in the plan before launch.