Can AI Agents Use Your Website? What a 1,560-Agent Test Found

What did the AudioEye test actually measure?

On 24 September 2026 AudioEye, a digital accessibility company, published a study that asked a simple question: can AI agents finish ordinary jobs on a website, and does the way the site is built change the answer? It sent 1,560 AI agents, built on six commercial AI models, to complete 13 real-world tasks on six websites. Five were retail, travel or e-commerce sites; the sixth was a W3C demonstration site. Every test ran ten times.

The only thing that changed between runs was whether accessibility fixes were in place. On the accessible version of a site, agents completed 96% of tasks. On the inaccessible version of the same site, they completed 31%. AudioEye describes that as a drop of up to 68%, and says it showed up across all the models it tested. The results were graded with WebJudge, an evaluation tool from Ohio State University’s NLP group, which agreed with the study’s own grading 95% of the time.

Failing was also more expensive. The median run used about 90,000 tokens on an accessible site and 128,000 without the fixes, 43% more. On the hardest site, the most expensive model used up to six times as many tokens. For anyone paying for agents to work on the web, that is a cost line; for a business, it means the agent sent by a customer is more likely to give up.

One caveat belongs right here: AudioEye sells accessibility software, so it has a stake in the answer. The design is still useful because the comparison is the same site with and without fixes, and the failure points it describes are specific enough to check on your own pages.

Why agents fail where people do not

A person looks at a page. An agent mostly reads it. AudioEye’s explanation is that agents rely on the accessibility tree, the structured version of a page that browsers build for screen readers: what each element is, what it is called and what state it is in. When an element has no name in that tree, the agent has to guess, and often it cannot.

The study gives two concrete examples. In one task, the numbers the agent needed existed only inside an image with no text alternative: across six models and ten tries each, 0 of 60 attempts succeeded. In another, survey answer buttons had no code tying each button to its option. AudioEye’s summary is blunt: missing accessibility labels stopped every model it tested.

OpenAI says the same thing about its own agent. Its publisher and developer FAQ states that ChatGPT Atlas uses ARIA tags, “the same labels and roles that support screen readers,” to interpret page structure and interactive elements, and it tells site owners to follow the W3C’s WAI-ARIA practices by giving buttons, menus and forms descriptive roles, labels and states. That is as close to an official agent-readiness checklist as any AI company has published.

Half of U.S. shoppers already use AI tools

This would be a technical footnote if few people used AI to shop. They do. The same day, NIQ published results from its Agentic Commerce Tracker: 51% of U.S. consumers said they had used at least one AI-powered tool to support shopping in the past month, the first time the figure has passed half. AI product recommendations were the most used, at 20%, followed by AI personal shopping assistants at 16%. The research is NIQ’s monthly Quick Question survey, about 500 U.S. consumers a month in 2026.

Liz Buchanan, President of North America at NIQ, put it this way: “The brands and retailers that succeed in this next era of commerce will be those that optimize not only for consumers, but also for the AI systems helping consumers navigate their choices.”

Most of that use today is comparing and narrowing, not an agent buying on its own. But the direction is set. When a buyer asks an assistant to compare three agents, book a tour or fill in a contact form, the business whose site the assistant can read and operate gets the inquiry. The one it cannot use drops out of the comparison without ever knowing it was in it.

What we found on 63 Miami real estate websites

To see how common these barriers are outside a lab, on 1 October 2026 we re-read the homepages of the same 63 Miami-area residential real estate websites from our Miami real estate AI readiness study last week. We checked three of the failure points the studies describe, in the HTML each site serves. We publish the totals, not the sites.

The forms stand out. On a real estate site the contact or showing-request form is where the money is, and on 27 of the 53 homepages with a form, at least one field gives an agent nothing to go on. A person sees an empty box next to a phone icon and understands it. An agent reading the accessibility tree sees an unnamed text field.

Images matter less on most of these homepages, because few of them hold data the visitor needs. They matter a great deal when they do: a price, a floor plan’s square footage or a rate chart that exists only as pixels is exactly the case that failed 60 out of 60 times in the AudioEye test.

How to make your website usable by AI agents

None of this needs a redesign. It is the order we would do it in, starting with the pages that make money:

1. List the five jobs a customer sends an agent to do. Request a showing, book a call, find a price range, filter listings, send a message. Those pages come first.
2. Label every form field. Each input gets a visible label tied to it, or at least an aria-label. Placeholder text disappears as soon as someone types and is not a dependable name. If your forms come from a builder such as GoHighLevel, check the rendered field, not the settings screen.
3. Name every button and icon link. A magnifying glass, a heart, a WhatsApp icon or an arrow needs a text name an agent can read: “Search listings,” “Save this home,” “Message us on WhatsApp.”
4. Put facts in text, not only in images. Prices, hours, addresses, rates and square footage should exist as text on the page. If an image carries them, its alt text should say them.
5. Use real controls. Buttons should be buttons and links should be links, with the right ARIA role and state on menus, tabs and filters, following the W3C’s ARIA Authoring Practices.
6. Test it the way an agent would. Give an AI browser agent one of the jobs from step 1 and watch where it stalls, then fix that element and run it again.

This is the part of agentic AI marketing that is easy to overlook: before you build agents that work for you, make sure the agents working for your customers can use your site. The same discipline runs through our wider AI marketing work.

Is this an accessibility project or a marketing project?

Both, and that is the useful part. The fixes that let an agent finish a task are the same fixes that let a person using a screen reader finish it. AudioEye’s own index puts the average at 62 accessibility issues per page, so most sites have work to do on both counts.

What changes is who pays attention. Accessibility work has often been framed as risk and compliance. Seen through agents, it becomes conversion work: every unlabeled field on a lead form is a lead an assistant could not send. For businesses that already run their follow-up through GoHighLevel systems, it is the front door to that pipeline; a form an agent can fill is a contact that lands in the CRM with the right tags.

It also connects to search. A page that names its elements clearly and states its facts in text is easier for search engines and AI assistants to read and quote. That is the same principle behind our generative engine optimization: say it plainly, in a form machines can read.

What these studies do not prove

It is worth being precise. AudioEye tested six sites and 13 tasks; that is a controlled comparison, not a measure of how agents perform across the web. NIQ’s 51% counts people who used any AI shopping tool in a month, including product recommendations, which is a much wider group than people who let an agent act for them. Neither study measures sales lost to inaccessible sites.

Our own check is narrower still. It reads 63 homepages once, from one location, without running JavaScript, so fields or buttons added by scripts are not counted, and a field we count as unlabeled may be named in a way our parser does not see. It measures barriers, not outcomes: we did not send agents to these sites. Treat the table as a sign of how common the problems are, not as a score for any one business.