I ran an accessibility audit on a dashboard that an AI coding assistant had built in an afternoon. The UI looked great. Genuinely — good spacing, sensible hierarchy, nice color choices.
Then I tried to use it with a keyboard.
The modal didn't trap focus, so tabbing from inside it went straight through to the page behind it. Three icon buttons had no accessible name, just <button><svg/></button>. The form's error state was conveyed by turning the input border red, which is invisible to a screen reader. None of this was exotic. It was the standard checklist, failing in the standard ways.
The model writes plausible markup. Visually plausible, structurally plausible. But the things a11y audits check for — focus management, accessible names, ARIA usage, contrast — are invisible in a screenshot, and invisible in a code review that happens at reading speed. The person who prompted the UI often can't tell anything is wrong, because for them, nothing is.
Meanwhile the volume is going up. When generating a screen costs a sentence instead of a sprint, the audit queue grows at the same rate. Manual audits don't scale to that, and "we'll fix it before launch" keeps not happening.
The boring WCAG checklist exists because these failures repeat. Alt text missing. Contrast under 4.5:1. Form inputs without labels. Click handlers on divs. Focus that never lands where the user expects. Every one of them is mechanically checkable, which means every one of them can be caught before the screen ships.
That's the idea behind A11yGuard: paste the AI-generated component markup or point it at a URL, and it runs a WCAG/ADA-oriented checklist — alt text, contrast, ARIA, keyboard navigation, labels — with specific fix guidance, not just "consider improving accessibility." The rules lean toward the failure modes that AI-generated frontends actually produce, which are a bit different from hand-written ones.
It won't certify compliance with anything. WCAG conformance involves real users, assistive technology testing, and judgement calls that no automated tool can make. Automated checks reliably catch roughly the easy half of the problem. The other half — is this focus order actually sensible, is this alt text meaningful — still needs a person, ideally someone who uses assistive tech daily.
But the easy half is exactly the half that "we'll audit it later" lets ship, and it's half the problem gone in minutes.
Tab through it without touching the mouse. Hit enter on the thing you focused. If focus vanishes, or lands somewhere surprising, you found the first real bug before the checklist even runs.
a11yguard.lxsaihub.com — paste the markup, see what falls out. And if the fix guidance for a rule is wrong or vague, tell me which one; vague guidance is a bug in the tool, not in you.