Experiments & Data

Experiment Readout Calculator

The other half of the test, and the half that goes wrong. Two conversion rates always differ; the question is whether they differ by more than chance would produce anyway. This gives you the lift, the range the true lift plausibly sits in, and a p-value — and then tells you plainly when that range includes zero, which means you learned nothing, or when the sample is too small for any of it to mean much.

Inputs

  • Control visitors
  • Control conversions
  • Variant visitors
  • Variant conversions

How to use this calculator

  1. Enter the visitors and conversions for each arm exactly as your analytics reports them: visitors who were exposed to the variant, and the conversions among those visitors. Counting all site traffic against test-only conversions is the most common way this comes out wrong.
  2. Read the confidence interval before the lift. The headline lift is one sample's answer; the interval is the range of true effects consistent with what you saw, and it is usually far wider than people expect.
  3. If the interval contains zero, the test did not find anything. That is not 'trending positive' and not 'directionally good'. It means a loss is as consistent with your data as a gain. Do not ship on it.
  4. If the interval clears zero, look at how wide it still is. An interval running from +1% to +30% tells you the change works and tells you almost nothing about how much, so plan on the low end.
  5. Check the smallest cell in the table. Under about ten conversions the maths behind every number here stops being trustworthy, and the tool will refuse to print a p-value rather than dress up four events as a finding.

What each term means

Conversion rate
Conversions divided by visitors in that arm. Two rates always differ; the question is whether they differ by more than chance produces.
Absolute difference (pp)
Variant rate minus control rate, in percentage points. Going from 4% to 4.65% is 0.65 percentage points, and also a 16% relative lift. Both are true, and the relative one always sounds bigger.
Confidence interval
The range of true differences consistent with this data at 95%. Run the same test twenty times and roughly nineteen of these intervals would contain the real effect.
p-value
The chance of seeing a difference at least this large if the two versions were actually identical. It is not the probability that the variant is better, and a great many decisions have been made on that misreading.
Statistically significant
Only means the result is unlikely to be pure chance. It says nothing about whether the effect is large enough to be worth the engineering, and a big enough sample makes trivial differences significant.
Normal approximation
The shortcut that lets binomial counts be treated as a bell curve. It needs roughly ten conversions and ten non-conversions per arm before it can be relied on.

Educational disclaimer: Outputs are simplified educational estimates built from the numbers you enter — they are not financial, legal, tax, or investment advice, and real decisions deserve verified figures and qualified professionals.

Quantitative Methods

More Experiments & Data tools