AI agents finished far fewer tasks on websites with accessibility problems

An AudioEye study sent 1,560 AI agents to complete everyday tasks. On one site they finished 96% of tasks when it was accessible and 31% when it was not.

AudioEye, a web accessibility company, published a study on 24 September looking at how AI agents cope with websites. These are the assistants that browse and click through sites for a person, for example to book something or find a product. It found that agents completed far fewer tasks on sites with accessibility problems.

AudioEye ran 1,560 agents from six commercial AI models through 13 everyday tasks on six websites, repeating each test ten times. Five were retail, travel and commerce sites, tested with and without accessibility fixes. The sixth was a W3C demonstration site that exists in an accessible and an inaccessible version. On that site, agents completed 96% of tasks on the accessible version and 31% on the inaccessible one, and accessibility was the only difference. Across the tests, completion fell by up to 68%, and agents used 43% more tokens on average when the fixes were missing. One task needed a number that only appeared inside an image with no alt text, and all 60 attempts failed.

The reason is how agents read a page. Many rely on the accessibility tree, the same structure screen readers use. If a button has no label or an image has no description, the agent cannot see what it needs. If customers start sending assistants to buy, book or compare for them, a site the agent cannot read will lose those sales.

AudioEye sells accessibility tools, so read the study with that in mind. The fixes it points to are still basic ones. Add alt text to images that carry information, label your buttons and form fields, and do not put prices or key details only inside images.

Work that makes your site easier to use with a screen reader also makes it easier for an AI agent to use.