Conversion Rate Calculator
Plug in visitors and conversions. Get the conversion rate with a proper 95% confidence interval (Wilson score), optional revenue per visitor / per conversion — or flip to A/B mode for uplift and a real significance test between two variants.
Variant A
—Variant B
—Absolute difference: — · Relative uplift (B vs A): —
—
How many visitors do I need for ±X precision?
Why the CI matters
Most online conversion-rate calculators show only the point estimate: "5%". This is a partial answer. With 100 visitors and 5 conversions, the actual rate could plausibly be anywhere from 2.2% to 11.1% — a range so wide that "5%" is more a starting hypothesis than a fact. Decisions made on the point estimate alone (kill this campaign, double down on this variant) often reverse when the next 100 visitors arrive.
The 95% CI is the honest version: "the true rate is probably somewhere in this range, given how many people we've actually seen." If the range is too wide for your decision, get more data.
The "rule of 30" doesn't apply here
A common myth: "n=30 is enough for the CLT to apply, so we can use normal-approximation formulas." For binomial proportions — which is what conversion rate is — this rule undercounts the sample needed when rates are extreme (very high or very low). A 1% conversion rate genuinely needs hundreds of conversions (tens of thousands of visitors) before the CI is tight enough to do anything with.
Wilson handles this correctly without you having to remember when which formula applies. The tool above tells you when your n is too small to draw a confident conclusion.
A/B tests: uplift is not evidence
The A/B mode reports three things: the absolute difference in rates, the relative uplift of B over A, and a pooled two-proportion z-test with a two-tailed p-value. The first two answer "how big does the effect look?"; only the third answers "is there an effect at all, or is this noise?". A headline like "+30% uplift" from 50 vs 65 conversions per 1000 visitors sounds decisive but tests at p ≈ 0.15 — a result you'd see one run in seven with no true difference. Ship decisions on p-values (or a pre-registered Bayesian analysis), not on uplift screenshots.
Related
FAQ
What's a 'Wilson score' confidence interval?
It's a more accurate confidence interval for proportions than the textbook formula. The textbook (Wald) interval is p ± 1.96·sqrt(p(1-p)/n), which breaks down at small n and at extreme rates (e.g. 0/100 gives '0% ± 0%' — meaningless). Wilson handles small samples and extreme proportions correctly. It's been the recommended default in modern statistics textbooks since around 2000 and is what statsmodels, R, and most production analytics tools use.
Why does the CI matter?
Because conversion rates are estimates with uncertainty, not facts. '5% conversion rate from 100 visitors' could plausibly be anywhere from 2% to 11% in reality. Treating the point estimate as the truth is how teams ship A/B tests that revert two weeks later when they sample more users.
How big does my sample need to be?
Depends on the precision you need and the rate itself. For ±2 percentage points at 95% confidence with worst-case rate (50%), you need ~2400 trials. For ±5 pp, ~385. For ±0.5 pp, ~38,400. The 'How many visitors do I need' section computes this directly.
How does the A/B test mode decide significance?
It runs a pooled two-proportion z-test: pool the conversion rate across both variants (p̂ = (convA + convB) / (nA + nB)), compute z = (pB − pA) / √(p̂(1−p̂)(1/nA + 1/nB)), and convert |z| to a two-tailed p-value. If p < 0.05, the badge reads 'statistically significant' — meaning a difference at least this large would appear less than 5% of the time if the variants truly converted identically. Example: A 50/1000 vs B 65/1000 is a +30% relative uplift, but z ≈ 1.44 and p ≈ 0.15 — not significant yet.
My uplift is +30% — why isn't it significant?
Because the uncertainty around each rate is bigger than the gap between them. Relative uplift is a point estimate; significance asks whether the gap could plausibly be noise given your sample sizes. Small samples and low base rates need surprisingly large n: detecting a real 10–20% relative lift on a 5% base rate typically takes thousands of visitors per variant. Keep the test running — and decide the sample size before peeking, or you'll inflate your false-positive rate.
What do revenue per visitor and revenue per conversion tell me?
Enter the revenue those visitors generated (single mode) and you get revenue / visitor (RPV — the value of traffic, useful for valuing ad clicks) and revenue / conversion (average order value). RPV = conversion rate × revenue per conversion, so it moves when either lever moves — which is why a variant can win on conversion rate and still lose on revenue.