Skip to main content
Back to Digital Inclusion Agent Blog

Web Accessibility

How Accessibility Checkers Work

A plain English tour of what happens when you run an accessibility checker: the visit, the rule engine (the part that runs the tests), the report, and the judgment calls software leaves to people.

· 6 min read

Glowing blue gears and a scanning beam inspecting a bright website page panel, against a near-black background.

The one sentence version

An accessibility checker is software that loads your web page, rebuilds it the way a browser does, and runs a long list of yes or no tests against the result. Every failed test becomes a finding in a report.

It can feel like magic when a tool hands you a verdict in seconds. It is not magic. A checker is three parts working in sequence: a visit, a rule engine, and a report. Once you see the three parts, you will know what a checker can and cannot promise.

Step one: the checker visits your page

A checker starts by fetching your page and actually running it. Modern pages are assembled by scripts, so the checker uses a browser engine to load the page, execute the code, and wait until everything settles into its final state. Testing the raw code alone would miss most of what a visitor gets.

Then it builds a map of the result, called the document object model (the browser's map of the page): the tree of headings, links, buttons, images, and form fields your page describes. Many checkers also build the accessibility tree (what screen readers use), the version of the page that screen readers work from, where each element carries a role, a name, and a state. The checker tests these trees. It cannot test your intentions, only what the page actually became.

The same engine can scan one component, a full checkout flow, or a whole site through a crawl, which is why you find the same technology in quick online tests and in thorough professional audits.

Step two: a rule engine asks the questions

Against that map, the checker runs its rules. Each rule is a narrow question with a definite answer. Does this image have alternative text? Does this button have a name a screen reader can announce? Is this text at least 4.5 to 1 against its background? Does this form field have a label connected to it?

The best known rule engine is axe-core, an open source library from Deque. Deque describes axe as a suite of accessibility testing tools built on that engine, and axe-core powers a large share of the checkers on the market, including tools built into the developer panels of popular browsers. Its rules are tied to specific WCAG (web accessibility standards) criteria, so a failure points you at the standard you missed rather than a vague warning.

Axe is one engine among many. The W3C Web Accessibility Initiative maintains a searchable list of evaluation tools that runs to hundreds of entries: browser extensions, full platforms, and command line tools developers wire into their build pipelines.

Step three: the report sorts the answers

Results come back sorted into buckets. Violations are definite failures, where the rule fired and the answer was wrong. A second bucket holds the questions the software could not settle on its own, flagged for a person to review. Passes are recorded as well, so you can see what was actually tested and found clean.

Some checkers show findings in a report file. Others show them on the page itself. WAVE, the long running evaluation tool from WebAIM, offers a suite of tools that provide visual feedback about the accessibility of web content, marking errors and alerts right beside the elements they belong to. For an owner looking at their own site, that picture is often worth more than a spreadsheet.

Good reports also rank what they find. A missing form label that blocks a purchase outranks a decorative image flagged for review, and a useful checker makes that order clear, so a small team fixes the barriers that lock people out before it polishes the rest.

What checkers are great at

Checkers earn their keep on mechanical failures, the problems that have exactly one right answer:

  • Images with no alternative text attribute at all.
  • Text that falls short of the minimum contrast ratio against its background.
  • Form fields with no connected label, and buttons with no accessible name.
  • Invalid ARIA, duplicate element IDs, and a page with no declared language.
  • The same template failure repeated on forty pages, found once and fixed everywhere.
  • They run in seconds, they never get tired, and they give the same answer on the hundredth page as on the first. That consistency is exactly what you want from a machine.

What a checker cannot decide

Every rule needs a definite pass or fail, so anything that requires judgment lands outside the engine. A checker can confirm an image has alternative text. It cannot confirm the text describes the image well, or that a decorative image was the one left blank on purpose. It can confirm your headings nest in order. It cannot tell you whether they make sense read aloud, in sequence, to someone who cannot see the page.

This is why honest checkers return a needs review bucket instead of pretending to settle those questions. It is also why a clean report means something real and limited at the same time: your code passed every test that can be automated. The judgment calls are not rare edge cases. They are the difference between a page that passes and a page that works.

Putting the checker to work

The routine that works: run a checker early and often. Run it in the browser while pages are being built, in the pipeline before anything ships, and on a schedule after launch, because new content quietly reintroduces old failures. Then put human eyes on the needs review bucket, and human hands on a keyboard walkthrough of your key pages, before anyone calls the site accessible.

That pairing is how we run accessibility work at Thanos Labs. The checker finds mechanical failures at machine speed. Our accessibility AI agent drafts the fixes. A human reviews and approves every change before it goes live. Software asks the questions. People still make the calls.

Sources

Want Thanos working on your site?

Discuss this topic with us

Built for agencies

Run the four AI scanners and client ready reports under your own brand.

See how we work with agencies