Ad A/B Test Significance Calculator – Real Win or Noise?
Creative A pulled 200 clicks from 5,000 impressions. Creative B pulled 250 from the same 5,000. B is ahead by a quarter, the dashboard is green, and someone wants to kill A this afternoon. The only question worth asking first is whether that gap would survive being run again, and that is what an A/B test significance calculator exists to answer. Most free ones print a bare “Significant!” and stop. This one shows you the uncertainty, because the uncertainty is the finding.
The two-proportion z-test, in four lines
Two rates, two sample sizes, one test. The null hypothesis is that both creatives share a single true rate, so the two arms get pooled into one estimate before the standard error is computed.
p1 = c1 / n1 4.00%
p2 = c2 / n2 5.00%
p_pool = (c1 + c2) / (n1 + n2) 0.045
SE_pool = √(p_pool(1 − p_pool)(1/n1 + 1/n2)) 0.0041461
z = (p2 − p1) / SE_pool 2.4119
p = 2(1 − Φ(|z|)) 0.0159A p-value of 0.0159 means a gap this large would appear by chance in about 1.6% of tests where the two creatives were truly identical. Flip it round and you get the figure marketers quote: 98.41% confidence the difference is not chance. That wording is deliberate and worth guarding. It is emphatically not “a 98% probability that B is better” — that is a Bayesian claim about your specific test, and a frequentist p-value cannot make it.
Why the interval matters more than the verdict
“Significant” is one bit of information. The confidence interval is the whole distribution of plausible answers, and it is what you actually plan against. For the example above the absolute difference is +1.00pp with a 95% interval of +0.19pp to +1.81pp, and the relative lift is +25.0% with an interval of +4.2% to +49.9%. B won — but the honest forecast is “somewhere between a rounding error and a transformation”, not “25%”.
Note that the interval uses the unpooled standard error while the test uses the pooled one. That is not an inconsistency: pooling assumes the null hypothesis is true, which is right for a test statistic and wrong for an interval that has to cover a difference which may be real. The lift interval, meanwhile, comes from the delta method on the log risk ratio rather than from dividing the absolute interval by p1 — the shortcut produces a symmetric range where the truth is skewed, and it understates the upside every time.
The peeking problem
The single most common way marketers fool themselves is checking the test every morning and stopping the moment it turns green. A p-value is computed on the assumption of exactly one look at the data. Armitage, McPherson and Rowe showed in 1969 that repeating a fixed-level test on accumulating data pushes the false-positive rate above the nominal level, and it climbs with every extra look. Fix the sample size before the test starts, run to it, read the result once.
Statistical significance is not practical significance
With enough traffic, almost any difference becomes statistically detectable. At ten million impressions per arm, moving a 0.50% rate to 0.51% — a gap of one hundredth of a percentage point — clears the 95% bar comfortably at z = 3.15, and can still be worth nothing, because rebuilding the creative costs more than the extra conversions return. Before running the test, write down the smallest lift that would actually change your decision. If the confidence interval sits entirely below that number, the correct conclusion is “B wins and it does not matter”.
Small samples, and the warning you should not skip
The z-test approximates a binomial with a normal curve, and that approximation needs enough successes and failures in both arms — conventionally five at the absolute minimum and ten to be comfortable. Three conversions out of forty thousand impressions is an enormous sample and nowhere near normal. This is why the calculator checks all four cells and says so loudly when one is thin: a confident p-value computed off three conversions is arithmetic dressed up as evidence.
Reading a null result properly
An inconclusive test is not a tie. It is a statement that your data cannot separate the options at the precision you asked for. A gap of +8% with an interval running from −8.7% to +27.7% leaves a real win entirely on the table — keep going. An interval of −1% to +1% is a genuine answer: the creatives are interchangeable, so decide on production cost instead and spend the traffic on a bolder test.