LLM view
AI assistants read your page for people, without seeing its design.
Try it here
Press “Open the LLM view”. A panel opens with this page as an AI model is given it: the text, the buttons it can name, and any hidden words.
Why should I care?
AI helpers read your page as text: whatever lives only in a picture or a script, they can’t pass on.
- 6 in 10screen reader users use AI to describe images.
- 4 in 10use it to help them through a task online.
- About halfof the time, the best AI helper succeeded, in a study with blind people.
What to do on your website
- Make sure your words arrive with the page.
- Give every button a name in words.
- Say the same thing to people and machines.
More, if you want it
Check your own website with this tool
This is the button you’ll drag.
- Make sure you can see your bookmarks bar, the row just under the address bar. If it isn’t there, press Ctrl + Shift + B, or ⌘ + Shift + B on a Mac.
- Drag the LLM view button up into that bar and let go.
- Open your own website and click LLM view in the bar. Click it on any page and a panel opens with the page as a model receives it: as Markdown with its token count, as first sent before JavaScript, as a map of the controls an agent can name, its facts for machines, and every hidden word. Click again to close it.
This works on a computer, in Chrome, Safari, Firefox or Edge. Phones and tablets have no bookmarks bar, so there are three other ways below.
On a phone or tablet
iPhone or iPad, from the Share button
- Copy the tool:
- Open the Shortcuts app, make a new shortcut, and add the action Run JavaScript on Web Page. Paste the tool in place of its sample script.
- In the shortcut’s details, turn on Show in Share Sheet, and name it LLM view.
- Once: in Settings, open Apps › Shortcuts › Advanced and turn on Allow Running Scripts.
- On your website in Safari, tap Share, then LLM view.
Android, from Chrome’s address bar
- Copy the tool:
- Bookmark any page with the star, then edit that bookmark: name it gc llm view and paste the tool in as its address.
- On your website, type gc llm view in the address bar and tap the starred bookmark it suggests. Choosing it from the bookmarks list won’t work.
Use it later, on a computer
Send yourself the link, or a reminder for tomorrow morning. Both are made on your phone and sent nowhere else.
Get all 10 tools in one go
One Good Citizens folder with every tool in it. Download the file, then bring it into your browser:
Download the bookmarks file- Chrome: Bookmarks › Import bookmarks and settings › Bookmarks HTML file
- Safari: File › Import From › Bookmarks HTML File
- Firefox: Bookmarks › Manage bookmarks › Import and Backup › Import Bookmarks from HTML
More free tools: Fovea, Color vision, Aging eyes, or all 10.
Why this matters
A customer asks an AI assistant to compare prices for them, and it opens your page on their behalf. It doesn’t see your layout; it is given text, a list of named buttons and fields, or a screenshot, like a friend sent on an errand with only a shopping list.
An LLM, a large language model, is the kind of AI that does this reading. This tool doesn’t show what a model thinks. It shows what a model is given, so you can see whether your page still makes sense when that is all there is.
A new reader, often running an errand
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, programs that visit page after page, gather them 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 summarize 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, the errand fails, and it is often a person who goes without.
What an AI reader is actually given
The major AI crawlers don’t run JavaScript1. JavaScript is the code that builds parts of a page inside the browser; they fetch those files and never run them, so they get only the page as it was first sent, the first HTML.
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.
Agents stumble where people are made to
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 a page disabled people can’t use. An accessible page is still the fix, for people and for the agents working for them.
Text people can’t see is read in full
OWASP ranks prompt injection11 the first risk for applications built on language models, and web pages are one of the main ways in. Prompt injection means slipping a model instructions its person never gave.
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 slip a note into the errand-runner’s list. 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.
All the advice, in full
Put the words in the first HTML
Render content on the server, or at build time, so the page arrives with its words in it. Crawlers that skip scripts, and people on slow connections, both get the page.
Use real controls, with names
A button that is a <button>, a link that is an <a>, each with words or a label. Agents find them by role and name, as screen readers do (WCAG 4.1.2).
Prefer HTML to ARIA
Native elements carry their roles, names and states for free. A clickable div is invisible to anything reading roles.
Say the same thing to people and machines
Keep structured data true to the visible page: the same price, the same date, the same author. Google asks for exactly this14.
Hide nothing you wouldn’t show
No text in the background color, no instructions meant only for machines. If it matters, show it; if it is for screen readers, make sure it says what the screen says.
Choose your robots.txt rules by purpose
AI companies name their crawlers by what they do: training, search15, or fetching for a person16. Block training if you like, and still let people’s assistants in.
On prototypes: once it is built
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.
What people get wrong
- You need an llms.txt
- It is a proposal, not a standard. Google’s John Mueller said in 2025 that no AI system uses it, and Google says its AI features need no special files. Lighthouse does check for one, so it does no harm.
- AI crawlers see what browsers see
- The major ones don’t run JavaScript. If the words arrive by script, they don’t arrive.
- Blocking GPTBot keeps you out of ChatGPT
- Training, search and fetching for a person are separate crawlers, and fetches a person asks for may not follow robots.txt at all.
- Agents will make accessibility unnecessary
- They fail most where pages are least accessible. The fix for both is the same page, built well.
Where these numbers come from
Every figure on this page comes from one of these 18 sources. The small numbers in the text point here.
Vendor data
The rise of the AI crawler·
Vercel, with MERJ
Survey
Screen Reader User Survey #11 results·
Article
Introducing Be My AI (formerly Virtual Volunteer) for People who are Blind or Have Low Vision, Powered by OpenAI’s GPT-4·
Guidance
Computer use toolClaude Platform Docs
Vendor data
Computer-Using Agent·
Reference
Playwright MCPGitHub
Guidance
Lighthouse agentic browsing scoring·
Chrome for Developers
Peer-reviewed study
WebArena: A Realistic Web Environment for Building Autonomous Agents·
ICLR 2024
Article
A11y-CUA Dataset: Characterizing the Accessibility Gap in Computer Use Agents·
arXiv
Peer-reviewed study
Are We There Yet? Assessing Computer-Use Agents for Blind Users’ Accessible Interaction with Desktop Applications·
EMNLP 2026
Standard
LLM01:2025 Prompt injection·
OWASP Top 10 for LLM Applications
Vendor data
Piloting Claude in Chrome·
Peer-reviewed study
Manipulating LLM Web Agents with Indirect Prompt Injection Attack via HTML Accessibility Tree·
EMNLP 2025 Demos
Guidance
AI features and your website·
Google Search Central
Guidance
Overview of OpenAI CrawlersOpenAI Platform
Guidance
Does Anthropic crawl data from the web, and how can site owners block the crawler?Claude Help Center
Reference
The /llms.txt file·
llmstxt.org
Article
No AI system currently uses llms.txt·
Bluesky