Learning cards

A learning card is the record of one finished test. Once learnings are on for a website, Pertento reads each test an hour after it ends, when its final statistics are in, and writes a card: what each variant changed on the page, step by step, what the test taught in one sentence, whether the result supported the hypothesis, a possible reason, and one test to try next.

The labels are per variant, because a test with three variants often tries three different things. Each variant gets its change types and page type from fixed lists, the element it changed, and a theme, which is how tests chasing the same idea under different names end up side by side. Those labels are what the patterns are built from.

The AI reads the test itself: the changes, the variant code, the hypothesis, the targeting, screenshots of the variants if you allow it, and the team’s debrief. It runs its analysis on a result that code has already computed, and every number on the card comes from the statistics. A sentence holding a figure of the AI’s own is thrown away rather than shown.

A card says it was written by AI until somebody marks it reviewed, and it sits on the test’s row in the Library next to the team’s own read from the debrief. It is written once, and again only when what it describes changes, so a recomputed interval never rewrites it or spends credits on it.