A clean scan report is good news. It is not a certificate. An image with alt="photo" passes the check "does this image have alternative text?" and still tells a blind visitor nothing. We build a scanner, so we say this plainly: no automated tool can tell you a page is accessible.
The W3C says the same thing in its guide to choosing tools: "Web accessibility evaluation tools can not determine accessibility, they can only assist in doing so." That sentence is worth keeping in mind every time a dashboard shows a big green number.
Three kinds of checks
It helps to sort accessibility checks by who, or what, can actually decide them.
Checks a machine can settle. An <img> with no alt attribute. Body text at 3.1:1 contrast on a solid background. A form field with no label. An aria-labelledby pointing at an id that doesn't exist. A page with no lang attribute. These are facts about the code and the rendered page, and a scanner that gets them wrong has a bug.
Checks that need a browser behaving like a person. Does focus move into a dialog when it opens, and back to the button when it closes? Can you reach the menu with Tab? Does the page still work at 400% zoom? Where does focus go after you submit a form with errors? Answering these means driving a real browser: pressing keys, opening things, resizing. Our scanner does this with a real Chromium session, and it is also where scanners (ours included) miss things, because a site can hide a problem behind an interaction the tool never tries.
Checks that need judgement. Is that alt text a good description, or just a description? Is the heading "Overview" meaningful where it sits? Are the captions accurate? Is the error message understandable? A tool can flag suspicious patterns, such as alt text that is a file name like "IMG_2041.jpg", but the final call belongs to a person who understands the content.
Reading confidence levels
Not every finding deserves the same weight. When a check rests on a heuristic, for example deciding that a large bold paragraph is really a heading, the finding is less certain than "this button has no accessible name". Our reports label the least certain findings "Needs review". Treat them as "look at this" rather than "this is broken", and check them before you file a ticket.
How we measure our own accuracy
The W3C's Accessibility Conformance Testing (ACT) task force publishes test cases for its rules: small pages labelled "passed", "failed" or "inapplicable" by people who wrote the rules. We run our scanner against 547 of those examples on every major change. On 1 October 2026 it caught 94.9% of the failing examples (168 of 177) and wrongly flagged 1.6% of the passing or inapplicable ones (6 of 366).
Those numbers mean something narrow. They cover the ACT rules that have test cases, on small test pages, not the whole of WCAG on a real site. The misses cluster in a few rules, mostly ones that need judging what an image actually shows, which takes a person or image understanding. We publish the number because it is independent and repeatable, not because it proves more than it does.
What a report can't list
Some criteria can only be partly tested. A scanner can find that a video has a caption track, not that the captions are accurate. It can find a heading, not decide whether the heading describes the section. A few criteria, mostly at Level AAA, such as sign language for prerecorded video, have no automated test in our scanner at all. A report that stays silent about something is not saying it passes. Ours includes a manual checklist for the parts a person has to check, so the gap is visible instead of hidden.
A sensible way to use a scanner
- Run it on every deploy, so regressions get caught while the change is still fresh in someone's head.
- Fix the high-confidence findings first; they are usually quick and certain.
- Do a keyboard-only pass of your key journeys (sign up, search, checkout, contact) at least once per release.
- Try a screen reader on the same journeys: NVDA on Windows and VoiceOver on macOS and iOS are free.
- Before you claim conformance, get a manual audit, and test with disabled people if you possibly can.
A scanner is a very good smoke detector. It is not the fire inspection.
Further reading: Selecting Web Accessibility Evaluation Tools (W3C) and the ACT Rules.