AudioEye Study Finds AI Agents Fail Up to 68% More Often on Inaccessible Websites
AudioEye, the Tucson-based digital accessibility company (Nasdaq: AEYE), published research on Sept. 24, 2026, finding that website accessibility directly determines how well AI agents can carry out tasks on a site. The study sent 1,560 independent AI agents, built from six commercial models, through 13 real-world tasks on six websites, running each test once with accessibility fixes in place and once without.
On the site with the most accessibility issues, agents completed just 31% of assigned tasks, compared with 96% on an accessible version of the same site, a task-completion drop of up to 68%. Accessibility was the only difference between the two versions. According to the findings, completion rates fell by about two-thirds for every model tested on the inaccessible site, including the most advanced models available. The sixth site in the study was the W3C's demonstration site, which is published in both accessible and inaccessible versions.
The cost of inaccessibility extended to compute as well. Across all 1,560 tests, the median run consumed 43% more tokens when accessibility fixes were absent, rising from 90,000 to 128,000 tokens, a gap AudioEye describes as a compute tax. On the site with the most accessibility issues, every model used at least twice as many tokens without the fixes, and the highest-cost model used six times as many. The company attributes the gap to how agents read a page: they navigate through a site's accessibility tree, the same structure screen readers have relied on for about 25 years, rather than the browser's DOM. When an element carries no accessible name or role, the agent cannot tell what it is or what it does, and workarounds follow.
Missing accessibility labels stopped every model tested. In one task, the required numbers existed only inside an image that had no text alternative; in another, survey answer buttons had no code linking each answer to its control. Both were clear to a sighted user, but six models each attempted the tasks ten times and all 60 attempts failed. AudioEye had the results independently rechecked using WebJudge, an open-source evaluation system from Ohio State University's NLP group, which agreed with the company's scoring 95% of the time.
"Agents use a website's accessibility tree, not the DOM, which powers web browsers," said Kelly Georgevich, CEO of AudioEye. "A core part of AI readiness in the future will be having an accessibility solution, not just for compliance, but for the agents themselves." The study arrives as agentic tools spread: AudioEye cites a Gartner survey of 3,566 customers in early 2026, in which 58% of generative AI users said they had already delegated a task to AI, rising to 74% in B2B, and points to Meta's September 2026 launch of the Muse personal agent, which acts across Meta's apps. AudioEye's separate 2026 Digital Accessibility Index found an average of 62 accessibility issues per page, and the company argues that each one is a gap in what an agent can read, making accessibility an increasingly commercial as well as a compliance concern.