The 2026 WebAIM Million measured something that should worry anyone patching accessibility with ARIA attributes instead of fixing markup: home pages that use ARIA average 59.1 detected errors. Home pages with no ARIA at all average 42. That's 17 additional errors, not fewer, on the pages that reached for the "accessible" toolkit.
82.7% of the top million home pages now use ARIA, up from 79.4% last year. ARIA usage overall is up 27% in a single year, and over 6 times what it was in 2019.
Why more ARIA means more errors, not less
An empty or incorrect ARIA attribute doesn't leave the accessibility tree blank. It fills the tree with confident, wrong information. A role="button" on an element with no keyboard handler tells assistive tech and AI agents alike "this is interactive," when it isn't. An honest gap is recoverable: a screen reader user or an agent can at least tell something is missing. A false attribute is worse, because there's no signal that it's wrong.
The W3C's own First Rule of ARIA has said this for over a decade: if a native HTML element already has the semantics and behavior you need, use it instead of re-purposing a generic element and bolting on ARIA. A <button> is a button in the accessibility tree with zero extra work. A <div onclick> with role="button" and tabindex="0" is a reconstruction project that's easy to get subtly wrong, and WebAIM's data suggests a lot of teams are getting it wrong at scale.
What this means for AI agents specifically
Agents reading the accessibility tree (Playwright MCP, and per OpenAI's own developer documentation, ChatGPT Atlas) trust the role and name a node reports. They don't have a way to verify it against the visual page. A button that claims to be enabled and isn't, or a region marked as a dialog that doesn't trap focus the way a real dialog should, doesn't just confuse the agent. It gives it wrong information with full confidence, which is a worse failure mode than a node with no information at all.
The fix
Native HTML first. ARIA only for the genuine gaps native HTML can't express, and even then, tested against a real screen reader, not just validated against a linter. "Passing an automated scan" and "reporting the right thing to the accessibility tree" are not the same claim.
Fuente de los datos: WebAIM Million 2026 (webaim.org/projects/million), sección ARIA.
Top comments (0)