AudioEye has published research suggesting that web accessibility can materially affect how effectively AI agents complete tasks on websites. In tests involving 1,560 agents across six commercial AI models, the company reported that task completion fell from 96% on an accessible version of a website to 31% on an inaccessible version, while median token consumption increased by 43%.
As AI agents move from answering questions to performing tasks, the structure of the websites they interact with is becoming part of the AI infrastructure problem.
AudioEye’s latest research puts that issue into measurable terms. The digital accessibility company tested 1,560 AI agents across six websites and 13 real-world tasks, comparing website experiences with and without accessibility fixes.
According to AudioEye, agents completed 96% of assigned tasks on an accessible version of one website but only 31% on the inaccessible version. The company says accessibility was the only difference between the two versions.
The findings point to an emerging connection between web accessibility, AI agent reliability and AI compute efficiency.
Accessibility becomes an AI interface
AI agents do not interact with websites in exactly the same way as human users.
AudioEye’s research focuses on the accessibility tree, a structured representation of webpage elements that exposes information such as roles, names, states and relationships to assistive technologies. W3C’s WAI-ARIA specification describes the accessibility tree as a parallel structure to the DOM that exposes relevant web content to accessibility APIs.
That structure can also provide machine-readable information about interactive website elements.
For an AI agent attempting to complete a transaction, missing semantics can therefore become an operational problem. A button without an accessible name, a form control without an identifiable label or an image containing critical information without a text alternative can make a task difficult or impossible for software to interpret.
AudioEye’s testing illustrates the potential impact.
The company reported that one task required agents to identify numbers appearing inside an image without corresponding text information in the code. Six models attempted the task 10 times each, resulting in 60 failures.
The finding is significant for agentic AI because an agent that cannot identify a control or understand its state may not simply provide a less accurate answer. It can fail to perform the requested action altogether.
Inaccessible sites can increase AI compute
AudioEye also reported a cost dimension to the problem.
Across its 1,560 tests, the company said median token consumption was 90,000 when accessibility fixes were present compared with 128,000 without them, representing a 43% increase.
On the website with the greatest number of accessibility issues, AudioEye said every tested model used at least twice as many tokens without the fixes, while the highest-cost model used six times as many.
AudioEye calls this additional consumption a “compute tax.”
The underlying implication is that poorly structured web content could force an AI agent to make more attempts, interpret more information or pursue longer reasoning paths before completing a task.
However, these figures should be treated as results from AudioEye’s controlled research rather than a universal measurement of the compute cost of inaccessible websites. The test covered six websites, 13 tasks and six commercial models.
Agentic AI raises the stakes
The timing of the research reflects a broader change in how consumers are beginning to use AI.
Generative AI systems are increasingly being positioned not only as interfaces for retrieving information but also as agents capable of navigating websites and carrying out actions on behalf of users.
That changes the requirements for digital infrastructure.
A website traditionally needs to be understandable to people and compatible with browsers, search engines and assistive technologies. Agentic interfaces introduce another software participant that may need to identify navigation elements, understand forms, select options and initiate transactions.
W3C’s current accessibility work already defines how semantic information is exposed through accessibility APIs. Its September 2026 draft on web user agents also describes user agents broadly as software that retrieves, presents and acts on web content on behalf of users.
The emerging agentic web therefore adds another reason for organizations to think about the machine-readable structure of their digital experiences.
Accessibility and AI readiness converge
AudioEye’s argument is not that accessibility should replace conventional AI-readiness practices. Rather, the research suggests that accessibility can become one component of an organization’s AI infrastructure strategy.
The same labels, roles and states that help assistive technologies understand an interface can also make interactive content more explicit to software attempting to operate it.
That creates an overlap between accessibility engineering and agentic AI engineering.
For e-commerce, travel, financial services and other transaction-heavy websites, the implications could be particularly important. An AI agent that can discover a product but cannot operate the site’s filters, understand a selector or complete checkout cannot deliver the full value of delegated shopping.
AudioEye’s own 2026 Digital Accessibility Index separately reported that two-thirds of AI search sessions were bypassing homepages and reaching interior pages, highlighting how AI-mediated discovery can change which parts of a website receive attention.
From SEO to agent readiness
The development also extends a broader evolution in digital optimization.
Organizations have spent years optimizing websites for human usability, search engines, mobile devices and accessibility technologies. AI agents add another layer: whether software can reliably interpret and operate the interface.
That does not make accessibility an AI-specific technology. Accessibility standards and practices were developed primarily to improve digital experiences for people with disabilities, and those objectives remain independent of AI adoption.
But the convergence is increasingly relevant to enterprise AI strategies.
AudioEye’s study was independently rechecked using WebJudge, an open-source evaluation system from Ohio State University’s NLP group. AudioEye said the independent scoring agreed with its own scoring 95% of the time.
The research remains a company-sponsored study rather than a broad industry benchmark. Even so, its central question is likely to become more important as agentic systems gain the ability to transact directly with websites.
For enterprises, AI readiness may increasingly depend not only on which models they deploy, but also on whether the digital environments those models must navigate are structured for reliable machine interaction.
Market Landscape
The AI infrastructure market is expanding from models and inference toward the systems that allow agents to interact with real-world software.
Model providers are developing increasingly capable agents, while enterprise technology teams are building orchestration, browser automation, retrieval, identity and governance layers around them. In parallel, web standards continue to define how semantic information is exposed to browsers and assistive technologies.
This creates a new intersection between web accessibility, AI agents and digital experience infrastructure.
AudioEye’s research adds a performance dimension to that discussion by linking inaccessible web structures with lower agent task completion and higher reported token consumption. The study is limited to its test environment, so additional independent research across more sites, tasks and models will be needed to determine how broadly the findings apply.
Top Insights
- AudioEye reported a 96% versus 31% task-completion gap between accessible and inaccessible versions of one tested website.
- The company reported 43% higher median token consumption when accessibility fixes were absent across its 1,560 tests.
- Missing labels and semantic information can prevent AI agents from identifying or operating important website controls.
- Accessibility trees expose structured web semantics through accessibility APIs, creating an important interface between websites and software agents.
- The research suggests AI readiness may increasingly include the machine-readability and operability of customer-facing digital experiences.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
