LLM view: the page, as an AI model reads it

In December 2024, OpenAI’s crawler made 569 million requests a month¹ across Vercel’s network, and Anthropic’s made 370 million. Google’s search crawler made 4.5 billion.

Crawlers gather pages for training and for search. Agents are different: they open a page because a person asked them to, to compare prices, fill in a form or summarise a long document. Many of those people are disabled. In WebAIM’s 2026 survey² of screen reader users, six in ten use AI to describe images, and four in ten use it for guidance through a task online. Since 2023, Be My Eyes³ has let blind people ask an AI model what their camera sees.

So when a model reads your page badly, it is often a person who goes without.

The major AI crawlers don’t run JavaScript¹. They fetch the script files and never execute them, so anything a page builds in the browser never reaches them.

Agents see more, in one of two ways. Some look at screenshots: Claude’s computer use⁴ works from images of the screen, and OpenAI’s Computer-Using Agent⁵ perceives the browser as pixels. Others read the accessibility tree, the same list of roles and names a screen reader uses. Playwright MCP⁶, which many agents use to drive a browser, hands the model lines like textbox “What needs to be done?” and needs no vision at all.

Google’s own Lighthouse guidance⁷ puts it plainly: agents rely on the accessibility tree as their primary data model, and semantic HTML is the machine-eye view of your page. A button that a screen reader can’t name, an agent can’t name either.

On WebArena⁸, a benchmark of 812 everyday web tasks, people succeeded 78% of the time and GPT-4 14%, in 2023. By 2025, OpenAI’s agent reached 58%.

Agents are improving quickly, and they still fail in familiar places. In A11y-CUA⁹, a computer-use agent completed 78% of everyday tasks as usual, 42% when it had to work by keyboard alone, and 28% through a screen magnifier: the same conditions many disabled people work in every day. In a three-week diary study¹⁰ with blind participants, the best agent succeeded about half the time.

That is the honest picture. An agent is not yet a fix for an inaccessible page. An accessible page is still the fix, for people and for the agents working for them.

OWASP ranks prompt injection¹¹ the first risk for applications built on language models, and web pages are one of the main ways in.

A line written in white on white, or pushed off the screen, is invisible to people and plain to a model. Some pages use that gap to give agents orders. In Anthropic’s own tests of an agent in the browser, such attacks succeeded 23.6% of the time¹² before safeguards and 11.2% after. Researchers have shown the same trick working through the accessibility tree¹³.

The design rule is old and simple: say the same thing to everyone. Words kept for screen readers are right to keep. Words that tell a machine something different from what people see are not.

This tool reads the page’s structure, so it needs real text, headings and buttons. A Figma prototype is drawn as a picture, with nothing underneath to read, and the tool will say so rather than report nonsense.

It does work on prototypes built as real pages: Axure, UXPin and Justinmind exports, Framer and Webflow previews, Storybook stories, and apps generated by tools like Figma Make or v0. If the content sits in a frame from another site, the tool offers to open the frame on its own.