Calculate how long to run an A/B test based on traffic, baseline conversion rate, and minimum detectable effect. Avoid stopping tests too early.
This calculator turns the three numbers you have before a test starts - daily traffic, current conversion rate, and the smallest improvement worth detecting - into a required sample size and a run duration in days and weeks. It applies the standard sample-size math for comparing two conversion proportions at a chosen confidence level and power, then divides by traffic to reach a calendar date. The number people get wrong first: sample size does not grow linearly with what you want to detect. Detecting half the effect needs about four times the visitors, which is why so many under-powered tests end inconclusively after a month.
A concrete input and expected output from the current implementation.
Input
Baseline conversion rate 10%, minimum detectable effect 2 percentage points, 100 visitors per variant per day, 95% confidence, 80% power.
Expected output
Required sample: 3,841 visitors per variant, 7,682 total. At 100 visitors per variant per day, the test should run about 39 days - roughly 5.5 weeks.
The two-proportion sample-size formula returns 3,841 per group for detecting a shift from 10% to 12% at 95% confidence with 80% power, and 3,841 divided by 100 visitors per day is 38.4, rounded up to a 39-day run.
The tool first computes the visitors needed per variant from four inputs: baseline conversion rate, minimum detectable effect, confidence level, and power. Because it compares two proportions, the requirement depends on the spread of both the baseline rate and the assumed improved rate. Only afterwards does it divide by daily traffic to express the same sample as calendar duration. Traffic changes the duration; it never changes the sample size.
The required sample scales with the square of one over the effect: halving the minimum detectable effect multiplies the needed visitors by about four, because a smaller difference sits closer to the noise floor. At a 5% baseline, detecting a 1-point lift takes roughly 8,200 visitors per variant; at a 10% baseline, the same 1-point lift takes about 14,750. Low baselines make the same absolute lift easier to detect.
Not and keep the 5% error rate. Watching the result day by day and stopping at the first significant p-value is peeking; it pushes the real false-positive rate well above the nominal 5%, which is exactly why the planned duration exists. If you want early stopping, pick a sequential testing design before the test starts, and treat its result as the only stopping rule.
The formula scales with the inverse square of the effect. Halving the effect halves the difference you must resolve against the same background noise, and statistical precision only grows with the square root of the sample, so the visitors must roughly quadruple. At a 10% baseline, a 2-point effect needs about 3,840 visitors per variant; 1 point needs about 14,750.
The example and behavioral notes were checked against the browser implementation. Standards and primary references below define the relevant format, formula, or platform behavior.
Shorten URLs with local redirect tracking via localStorage. See click counts and generate a QR code for each link.
Build UTM-tagged URLs for campaign tracking with source, medium, campaign, term, and content parameters.
Enter 2+ headlines and see simulated previews in Google results, email subject lines, and social posts side-by-side.
Count characters for Twitter/X, LinkedIn, Facebook, and Instagram with platform-specific limits and warnings.
Calculate click-through rates from impressions and clicks with industry benchmarks and trend analysis.
Check email subject line length against best practices for desktop and mobile with preview text optimization.