AI agents fail on inaccessible websites: 96% vs 31%

An AI agent sent to buy something on your website does not see your website. It reads the accessibility tree. AudioEye published research on September 24 measuring exactly what that costs: on its least accessible test site, agents completed 96% of tasks against an accessible version and 31% against the inaccessible one. Same site, same tasks, a 65-point gap.

What actually happened

AudioEye ran 1,560 independent agents across six commercial models, on 13 real-world tasks, against six websites, repeating each test ten times. Results were independently scored with WebJudge, an evaluation system out of Ohio State, with 95% agreement against AudioEye's own scoring.

Completion rate was the headline, but token spend is the number an operator should read twice. Agents working on inaccessible versions burned a median 43% more tokens. On the worst site in the set, every model spent at least twice as many, and one spent six times as many.

CEO Kelly Georgevich framed the mechanism plainly: agents use a website's accessibility tree, not the DOM. Unlabeled buttons, div-based controls, images without alt text, forms whose inputs are not tied to their labels — a human squints and figures it out from visual context. An agent gets a tree with holes in it and starts guessing, which is what those extra tokens are.

Why website accessibility matters for your business

This reprices accessibility work. Accessibility has historically been sold to small businesses as legal exposure — an ADA demand letter you would rather not get. That framing made it a compliance line item, which is where budgets go to die. The new framing is operational: the same markup that fails a screen reader fails a shopping agent, and now roughly half of US consumers have an AI tool somewhere in their purchase path.

Token cost is a conversion signal. An agent that needs 6x the tokens to check out on your site is an agent that is slower, more expensive to run, and more likely to bail or pick a competitor's flow. You do not control the agent's budget, and you will never see the abandonment in your analytics. It looks like traffic that did not convert.

The fix list is the one you already have. Semantic HTML. Real <button> elements instead of clickable divs. Labels bound to inputs with for/id. Alt text. ARIA roles where native semantics do not exist. Focus order that matches visual order. None of this is new work invented for agents; it is the WCAG checklist your developer already knows and has been deferring.

Overlays do not fix the tree. A widget that repaints contrast and adds a toolbar does not repair the underlying markup an agent parses. Fix the markup.

Run one test this week: open your checkout in a browser's accessibility inspector and read the tree. If you cannot tell what the buttons do from the labels alone, neither can the agent.

Sources: AudioEye study announcement, NIQ Agentic Commerce Tracker.