Glossary · Éprouver
A/B test
An A/B test compares two versions of the same thing — a page, a destination, a hook — by splitting traffic between them at random. It answers one precise question: which of the two produces the better result, all else being equal.
Tester deux versions, et ne désigner un gagnant qu’au-dessus du seuil. The forty terms of the trade
What it is for
It replaces an opinion with a measurement. Two people can argue for an hour about the best landing page; a test settles it in a week. It also keeps you from breaking what works: the version in place keeps serving half the traffic for the whole test.
How it is set up
One single difference between the two versions, traffic split at random, and a duration decided in advance. Changing two things at once makes the result unusable: you will know one version wins, never why.
The classic mistake
Stopping the test as soon as the favoured version pulls ahead. On small volumes, the lead changes sides several times a day — you are not watching a winner, you are watching randomness.
A test that means something
- One differenceJust one. Two changes, and the result explains nothing.
- A random drawEach click falls one side or the other, at the chosen weight.
- The thresholdThirty clicks per version, and at least one sale. Before that, silence.
- The verdict95% confidence, or “no difference”. Never a winner out of politeness.
Three conditions before the fourth beat. Skipping the third is what moves budgets onto noise.
What Subtraq does with it
Subtraq splits clicks by weight between versions and records the variant on every click. Stopping a test never changes the link's destination, and no verdict is given before thirty clicks per version and at least one sale.
Read next
Un terme seul ne sert à rien. Voici ceux qui le rendent utilisable.
Variante
Read the definition →Signification statistique
Read the definition →Taux de conversion
Read the definition →Les fonctionnalités du produit
Open the page →« A/B test », une fois posé sur de vraies ventes
A test that picks a winner on twelve clicks moves a budget onto noise. The product's thresholds are public: thirty clicks per variant at least, and 95% confidence before announcing anything.
See the features