Can AI read your website? How structure decides whether you get cited

Two very different machines now read your website, and neither sees it the way a person does. One is the crawler that feeds AI answers (ChatGPT, Claude, Perplexity, Google's AI surfaces). The other is the AI agent that browses and shops on a user's behalf. Both reach your page through its structure, not its design. If that structure is clean, they can extract what you say and repeat it. If it isn't, you are invisible to them even when the page looks perfect to a human.

This guide is about the layer underneath GEO: whether AI can mechanically read your page at all, and what to do about it.

Do AI crawlers even see your page?

Mostly they see less than you think. Vercel's crawler study found that the major AI crawlers do not execute JavaScript: GPTBot and ClaudeBot download script files but never run them. They read the raw HTML your server returns, and move on. That is the opposite of Googlebot, which runs a full browser and renders the page.

The practical consequence: if your important content only appears after JavaScript runs (client-side rendering, content loaded on scroll, text injected by a framework), an AI crawler can see an empty shell. Server-render or statically generate anything you want quoted. Confirm AI crawlers can actually reach you before worrying about anything more advanced, because a blocked or unrenderable page fails at step zero.

Why structure decides what gets extracted

Once a crawler has your HTML, it has to work out which part is the content and which part is furniture: navigation, ads, cookie banners, related-post lists. It does that with semantic structure.

• Landmark elements (main, article, nav, header, footer) tell a machine where your real content begins and ends, so it can isolate it from the chrome.
• A clean heading hierarchy (one h1, then h2s and h3s that don't skip levels) works like a table of contents. It lets a retrieval system chunk your page into sections it can quote in context.
• Descriptive link text and real alt text turn links and images into information instead of dead ends.

None of this is exotic. It is ordinary, well-built HTML. But most sites don't have it: the 2026 WebAIM Million report found detectable structural failures on the large majority of top homepages, with missing link text and unlabeled controls among the most common. Every one of those failures is a place a machine loses the thread.

The new reader: AI agents and the accessibility tree

Answer-engine crawlers infer structure from your HTML. AI agents go one step further. The newer breed of browsing and shopping agents read the accessibility tree directly, the same simplified model of your page a screen reader uses. It strips the page down to roles, names, and states: this is a button called "Add to cart," this is a link called "Pricing," this is a labeled email field.

They favour it because it is cheap and unambiguous. A screenshot costs a vision model a lot of tokens and still leaves it guessing which pixels are clickable. The accessibility tree just says. OpenAI tells developers plainly that making a site more accessible helps its agent understand it.

The failure mode is literal: an unlabeled button or an empty link is a control the agent cannot identify, so it cannot click it, so it cannot recommend or buy through it. In 2026, Cloudflare Radar measured automated traffic passing human traffic on the web. The majority reader of your site is now a machine that depends entirely on structure.

So is AI discovery "based on accessibility"? Honestly, partly.

It is easy to over-read this. Accessibility and semantic HTML are necessary, not sufficient. They make your content extractable; they do not by themselves make you recommended. A perfectly structured page that nobody cites, and that answers no real question, will still lose.

The things that actually move AI recommendations are the ones in the rest of this library: being cited on the sources AI trusts, publishing answer-shaped content for the questions your buyers actually ask, and clear entity signals like structured data. Structure is the plumbing. Authority is the water. You need the pipes, but the pipes are not the point.

A short structural checklist

Fix these once and they keep paying off:

• Server-render or pre-render any content you want quoted. Assume JavaScript will not run.
• Use native elements: a real button, a real link, a real label, not a div wired up with JavaScript.
• One h1 per page, then a heading hierarchy that does not skip levels.
• Give every link and button text that makes sense out of context ("See pricing," not "click here").
• Real alt text on meaningful images, empty alt on decorative ones.
• Structured data. Add JSON-LD so machines can resolve what the page is about.
• Do not block AI crawlers in robots.txt unless you mean to. Confirm which bots you allow.

You can test most of this in your browser. Right-click, Inspect, and open the Accessibility panel to see the exact tree an agent reads. If a heading is missing, a button has no name, or your main content is not in the tree, that is what the machine sees too.

Where Geofound fits

Geofound does not grade your markup line by line, and it should not: your browser's dev tools and plenty of free checkers already do that. What Geofound does is close the loop around it.

• It shows you whether AI crawlers actually reach your site. Connect in one click through Cloudflare, a WordPress plugin, or a snippet, and watch which AI bots crawl which pages, so "we server-rendered it" becomes "GPTBot fetched it on Tuesday."
• Its crawlability audit checks the fundamentals a machine needs: whether robots.txt is blocking AI bots, whether the page responds, and whether it carries JSON-LD, a real title, and a meta description.
• Most importantly, it measures the outcome. Structure is a means to an end, and the end is being named in the answer. Geofound asks the AI assistants your buyers use whether they recommend you daily, and tracks it before and after you change something, so you know whether the structural work actually moved your visibility or just felt productive.

That last part is the whole point. You can do everything in the checklist above and still not know if it worked. The only proof is the answer itself.

The checklist tells you whether AI can read you. The free scan tells you whether it recommends you: three of your buyers' questions, and the names ChatGPT gave back.

FAQ

Does my site need to be WCAG compliant to be cited by AI? No. Accessibility compliance and AI-readability overlap heavily, but they are not the same goal. You are optimizing for the same underlying thing, clean and machine-parseable structure, and improving one tends to improve the other. Chasing a compliance score is not the objective. Being extractable, and then being cited, is.

Will AI crawlers ever render JavaScript like Google does? Some may get more capable over time. But building on the assumption that they will not is free insurance: a server-rendered page works for every reader, human or machine, today and later.

Is this just technical SEO? It shares a foundation with technical SEO, since both reward crawlable, well-structured pages. The difference is the target: SEO wants your page to rank, GEO wants your brand named inside the answer. See what GEO is for the full distinction.