Accessibility and compliance

Web accessibility audit with a concrete list of fixes

Automated checking of all 55 WCAG 2.2 level A and AA criteria, plus a vision model's assessment of whether the alternative texts actually say anything.

Client
e-commerce and corporate website operators
Status
Production · 2026
Outcome
  • The whole site is crawled, not a sample of pages, including content behind a load-more button.
  • The output is a single self-contained HTML report that works with no internet connection.
  • The report includes the legal context, so it is usable by management, not only by a developer.
Stack
Python and FastAPIPlaywright and Chromiumaxe-coreLLaVA via OllamaJinja2

The problem

The European Accessibility Act and its national transposition mean commercial websites must meet a specific technical standard. Most operators find out they have a problem when a complaint or an inspection arrives.

Ordinary free tools test one page and return a list of technical errors. That is insufficient for two reasons. The errors tend to be on the pages nobody tests, such as the basket or a form. And automated checking fundamentally cannot judge the thing that matters most — whether an image's alternative text says anything, or is merely present so that something is present.

What was built

The tool crawls the whole site with its own browser, including pages that only load after interaction. On each page it runs a standard checking engine and evaluates the page against all 55 level A and AA criteria.

Above that sits a second layer. Every image, together with its alternative text, is presented to a vision model which judges whether the text actually describes the image. This exposes the most common formally correct failure: alternative text that exists, passes the check, and helps nobody.

The output is a single HTML file. Nothing is fetched from the internet, so it can be emailed, opened a year later, and archived as evidence of the state on a given date. Besides the technical findings it explains the legal framework, because the person who decides on remediation is usually not the person who reads code.

Automated checking covers a large share of the criteria but not all of them. Text comprehensibility, the logic of keyboard focus order and usability with a screen reader must be judged by a person. The report marks these as requiring manual review rather than pretending the audit is complete.

Result

The tool is working and usable as a standalone service. For a customer it is the fastest way to find out the size of the problem before somebody else finds out.