How test duration is decided
Duration is not a preference, it is arithmetic with a calendar wrapped around it. The statistical part asks how much traffic the test needs to detect the change you care about; divide that by your weekly visitors and you have a floor. The calendar part then rounds that floor up to something the business actually experiences.
Rounding up matters because behaviour is not evenly spread. Weekday visitors behave differently from weekend visitors, paydays cluster, and in Nepal festival weeks distort almost every category. A run that stops mid-week hands one variant more of the good days than the other, and that imbalance is indistinguishable from a real result. Always run complete weeks.
On paid platforms there is a third constraint. Meta and Google both re-optimise delivery when something changes, and results during that learning phase reflect the algorithm settling rather than the creative or audience being tested. A test that ends before delivery stabilises measures the platform, not your idea.
Why test duration matters
Short tests produce confident nonsense. Early in a run the numbers swing widely, and whichever variant happens to be ahead looks decisive. Stopping there is how teams end up rolling out changes that quietly do nothing, then wondering why the year’s conversion rate never moved despite a long list of wins.
Long tests have their own cost. Traffic sent to a losing variant is revenue you chose to spend on learning, seasonality creeps in, and cookie churn means returning visitors may not stay in the group they started in. The right length is the shortest one that answers the question properly, which is why it is worth calculating rather than guessing.
Common mistakes with test duration
The first is deciding the end date by watching the dashboard. Checking daily and stopping on the first good day is peeking, and it turns a disciplined experiment into a search for a flattering moment.
The second is ignoring the buying cycle. A form for a plumbing callout is decided in minutes; a decision about an education consultancy, a property or an overseas remittance provider may take weeks. If the test is shorter than the time customers take to decide, it counts the visits and misses the conversions.
The third is launching a test alongside a campaign change, a price change or a public holiday, then reading the result as though the page was the only thing that moved.
How to act on it
Fix three things before launch: the effect you need to detect, the traffic that implies, and the end date that follows once you round up to complete weeks and clear the learning phase. Put the end date in the calendar and treat it as a commitment rather than a suggestion.
Then leave it alone. Do not change budgets, audiences, creative or the page itself while the test runs, and do not add a second test to the same journey. When the end date arrives, read the result once, including the range around it, and accept an inconclusive answer as a real answer. Booking the next test straight away is far more productive than extending this one until it says something you like.