Agents are visiting your site and reading it differently

Google now points site owners toward agent-friendly practice, and the reason is straightforward: a growing share of visits are software reading your page on someone's behalf. Those visitors do not see your design. They read the document object model and the accessibility tree.

Which means a page that looks immaculate to a human can be close to unreadable to an agent, and you would never know from a screenshot.

The traffic is small and converts unusually well

Keep both halves of that sentence together.

AI referral volume is still around 1% of sessions on most sites, and ChatGPT referrals grew 206% year on year from January 2025 to January 2026. Semrush puts AI-referred visitors at roughly 4.4 times the conversion rate of organic. Seer's client data shows ChatGPT-referred traffic converting near 16% against about 1.8% for Google organic.

The volume does not justify a rebuild. The conversion rate justifies not being invisible.

What agents actually read

An agent parses the served HTML and, when it is driving a browser, the accessibility tree. That is the same structure a screen reader uses: roles, names, states and relationships.

So the things that break for agents are the same things that break for a blind user, which is a useful shortcut. If your site passes a serious accessibility review, it is largely agent-ready already.

The specific failures, in the order they show up:

Icon buttons with no accessible name. A button containing only a magnifying glass graphic has a name of nothing. A human sees search. An agent sees an unlabelled button. Add an aria-label.

Divs acting as buttons and links. A div with a click handler has no role, cannot be focused and does not appear as an actionable control. Use real button and anchor elements.

Text rendered in canvas or as an image. Invisible. This includes pricing tables built as graphics and phone numbers set as images to defeat scrapers.

Content behind hover. An agent has no cursor. A dropdown that only opens on hover, with no keyboard or click path, contains information that does not exist as far as the agent is concerned.

Form fields with placeholder text instead of labels. The placeholder disappears on focus and is not a reliable accessible name. Label every field properly.

Critical content that arrives after several seconds of client-side rendering. Some agents wait. Many do not.

What Google says not to bother with

This is worth reading carefully, because a lot of paid advice contradicts it. Google published a guide on optimising for generative AI features in May 2026, and it is unusually direct.

On llms.txt files: Google Search does not use them. They neither help nor harm.

On structured data: not required for generative AI search. Useful for rich results, which is a separate reason to have it, but not the entry ticket people are selling.

On chunking: there is no requirement to break your content into small pieces for machine consumption.

On writing style: no need to write differently for AI. Manufacturing brand mentions across the web is described as not particularly helpful.

The one hard requirement is the boring one. The page must be indexed and eligible to appear with a snippet. If you have suppressed snippets anywhere, through a meta tag or a data-nosnippet attribute applied too broadly, you have suppressed AI eligibility on those pages at the same time. Check that before anything else.

Google also names the content type it does not want, using the example of a piece called something like seven tips for first-time homebuyers. That is the format a great deal of content marketing still produces.

A one-hour audit anyone can run

1. 1.Open your top service page and disable JavaScript. Whatever remains is roughly what a non-executing agent sees. If it is a spinner, that is your finding.
2. 2.Run the page through Chrome DevTools and open the accessibility tree in the Elements panel. Walk it. Every control should have a name that tells you what it does.
3. 3.Tab through the page with the keyboard only. If you cannot reach the enquiry form, neither can an agent.
4. 4.Confirm your prices, service descriptions, locations and contact details exist as text in the served HTML. Not in an image, not in a component that loads on scroll.
5. 5.Check for a stray noindex or nosnippet on any page you want cited.

None of that requires a developer to start, and most of it produces a short list that a developer can clear in a day.

The framing that makes this an easy decision

You are not building for a new audience. You are fixing the things that were already broken for screen reader users, keyboard users and anyone on a poor connection, and getting agent readability as a consequence.

That is the rare piece of work where the compliance case, the usability case and the visibility case all point the same direction, and where the fix list is short and cheap.

Build the page so it can be read without being seen, and everything that reads it will find you.