A green test suite proves that its assertions passed. It does not prove that an admin can still read the billing page in dark mode, that German labels fit, or that the empty state survived an AI-generated refactor.
The gap is structural. Unit tests observe logic. Component tests observe isolated render states. End-to-end tests usually observe a small number of happy paths. Users observe the whole screen under a particular identity, theme, locale, viewport, and dataset.
Treat visible output as release evidence
Define the journeys whose visible result would block a release. Capture each journey in the contexts that materially change it. When a pull request changes the output, attach the before-and-after evidence to the commit and keep the check pending until someone approves it.
This is not a demand to screenshot every pixel. Start with checkout, onboarding, permissions, billing, and the states support sees most often. Visual verification works when it protects decisions, not when it maximizes image count.
The useful question is no longer “did the test run?” It is “did a person see what this change actually ships?”