A score is not a fix list

A tool run on a fast connection from one location tells you what the page could do. Field data tells you what it does for the people loading it on a mid-range phone on mobile data, which is usually a different and less flattering story.

Both get looked at, and where they disagree the field numbers win. Chasing a lab score to ninety while real visitors wait four seconds is optimising the measurement.

Automated checkers find contrast ratios, missing alt attributes and broken landmarks, and they are worth running first. They cannot tell you that the focus order jumps across the page, that a modal traps a keyboard user, or that a label describes the wrong field.

So the pass is manual as well: through the site with a keyboard only, then with a screen reader, on the flows that matter most — sign-up, checkout, contact, whatever yours are.

Every finding names the page, the problem, who it affects and the specific change — this image is 2.4 MB and should be served at 600px wide in AVIF; this control is a div and should be a button.

That is the difference between a report that gets forwarded to a developer and one that gets acted on. A list of failures with no remedies attached is work moved, not work done.