What it is

Rules can tell that a link has a text. They cannot tell whether “Read more” says where the link goes. The AI review asks Claude, a language model made by Anthropic, a few narrow questions of that kind about each page. Its findings are suggestions for a person to check, never verdicts.

What it looks at

Four questions, each about one WCAG success criterion:

  • Page title (2.4.2). Does the title describe the page’s topic or purpose?
  • Link purpose (2.4.4). Can a reader tell where each link goes from its text and its surroundings?
  • Headings and labels (2.4.6). Does each heading describe the content after it, and each form field’s label say what to enter?
  • Image descriptions (1.1.1). Does each image’s text alternative serve the same purpose as the image?

For every item the model answers pass or fail with a one-sentence reason. It is told to fail an item only when it clearly fails the criterion, and to pass anything that could merely be better. It does not judge spelling, style or tone.

What is sent to Anthropic

To ask these questions, Supervisor sends parts of the page to Anthropic’s API. Only things a visitor can perceive on the page are gathered, and only from the pages a crawl reaches, which are pages open to anyone without signing in.

  • The page’s address, title and language, and its first headings.
  • The first 600 characters of the page’s text, for the title question.
  • Up to 40 headings, each with a little of the text that follows it.
  • Up to 30 form fields: the label, the field type, the placeholder and the text around it. Not what anyone has typed.
  • Up to 40 links: the link’s name, its address, the text around it and the heading before it.
  • Up to 20 images: the alt text, the file name, the caption or nearby text, where a linked image goes, and a small crop of the image itself, at most 400 px on its longer side.

The page’s full HTML is not sent. The text of the page is treated as data. Instructions written on a page are not followed.

Where the findings go

  • Every finding of the AI review waits as “needs review”. It never goes into your accessibility statement by itself.
  • The model’s reason is saved with the issue, so the person reviewing sees why it was raised.
  • You, or your agent with your say, confirm it or mark it as a false positive with update_accessibility_issue. That decision sticks.
  • When a later review of the same page no longer raises a finding, the issue is marked fixed.

In the issue list the rule ids are ai:page-title, ai:link-purpose, ai:headings-labels and ai:alt-text.

When it runs

  • On the desktop view of pages in a crawl. It does not run when a check follows a journey.
  • How many pages of a check are reviewed depends on the site’s plan. Pricing has the numbers. One check sends at most 50 pages.
  • The result of a check says how many pages were reviewed, under ai_judges in get_accessibility_check.

What it cannot do

  • It can be wrong in both directions. That is why a person decides.
  • It sees the gathered texts and image crops, not the whole page as a visitor does.
  • It does not test anything by using the page. Keyboard use, zoom and reflow are covered by Supervisor’s own browser checks.
  • It does not show that a page meets a WCAG criterion. A pass means only that nothing clearly failed in what it was shown.

Supervisor is an independent service. It is a customer of Anthropic’s API and is not otherwise affiliated with Anthropic.