Assignment without correct exposure logging mixes people who never saw the change into the treatment arm. Results flatten, and teams conclude “nothing works.”
Check that exposure waits for the variant UI, that bots are filtered, and that primary metrics can physically move from the change under test.
Secondary metrics still matter for guardrails — support tickets, crash rate, payment failures — but they should not silently become the decision metric mid-test.
A short readiness check before launch costs less than a month of arguing about a contaminated result.