Check A/B test sample size adequacy and conversion significance

It also checks whether both variants contain enough conversions and non-conversions for the normal approximation to be reasonably supported. Use it for a quick, transparent signal before making a landing-page decision, while remembering that statistical notation does not replace experiment design, traffic-quality checks, or a preselected stopping rule.

Enter clean totals from the same experiment

Provide the number of visitors and conversions for variant A and variant B. A visitor count represents the people or eligible sessions assigned to that version, while a conversion count represents those that completed the same predefined goal. Keep the attribution window, audience rules, bot filtering, and conversion definition identical across both variants. The calculator rejects a variant with zero visitors because no conversion rate can be observed, and it rejects conversions greater than visitors because that would not describe a one-conversion-per-visitor binomial outcome. If your metric allows repeat purchases or several events per visitor, aggregate it into a visitor-level converted-or-not outcome before using this test. Do not mix totals from different dates, campaigns, devices, or allocation rules unless those differences were part of the planned experiment. Random assignment and simultaneous exposure matter: a mathematically precise z-score cannot repair weekday effects, a broken split, duplicated visitors, or tracking that fires differently between versions. Confirm those basics before treating the result as evidence about the landing page itself.

Read the rates, difference, and z-score together

The observed rate for each variant is conversions divided by visitors. The absolute difference is variant B's rate minus variant A's rate, so a positive result favors B and a negative result favors A. The percentage-point difference expresses the same gap on a familiar 0-to-100 scale; it is not relative uplift. The calculator then pools conversions across both groups to estimate the standard error expected under the null hypothesis that the underlying rates are equal. Dividing the observed difference by that standard error produces the z-score. A result is flagged as statistically notable when its absolute z-score is at least 1.96, the familiar two-sided 5% threshold. This is a simple screening rule, not a promise that the winning page will reproduce the same uplift. It does not return a Bayesian probability, correct for repeated checking, adjust for multiple variants, or incorporate revenue value. Read direction, practical size, and statistical flag together instead of promoting whichever single number looks most favorable.

Respect the sample adequacy caveat before deciding

The two-proportion z procedure relies on a normal approximation. This checker marks the sample as adequate only when each variant has at least five conversions and at least five non-conversions. That rule is a practical minimum, not a guarantee of strong power. When it fails, the z-score is still displayed for transparency, but its threshold flag should be interpreted cautiously; an exact method or more observations may be more appropriate. Even when the check passes, decide whether the experiment could detect the smallest effect that matters commercially. A tiny but statistically notable gain may not pay for implementation, while a promising large observed gap may remain uncertain when traffic is limited. Avoid stopping the moment the score crosses 1.96, because repeated peeking increases false positives unless the experiment uses a sequential design. Set the target metric, minimum detectable effect, duration, exclusions, and stopping rule before launch. Finally, inspect guardrail metrics and segment balance. This calculator summarizes one binary outcome; it cannot reveal slower pages, refund changes, sample-ratio mismatch, or an uplift isolated to suspicious traffic.