llms.txt or WebMCP? Test Your Site With an AI Agent First

October 27, 2026

· By Tarek Smiri

llms.txt or WebMCP? Test Your Site With an AI Agent First

Short answer

In May 2026, 97% of the llms.txt files tracked by Ahrefs received no requests at all, and Google states that its Search ignores them. WebMCP, which lets a page declare tools that an AI agent can call, is more promising. But it is still a W3C Community Group draft, on trial in Chrome and called by very few agents. Before you invest in either, run a five-minute test: ask an AI agent to add a product to the cart or send your contact form, using test data, and note where it stops. That is where your real work is.

What the data says about llms.txt

llms.txt is a text file at the root of a site that summarises the site and lists its key pages for language models. We define it in our AEO and GEO glossary for agencies.

The largest study so far comes from Ahrefs, published in June 2026. It covers the 137,210 domains in Ahrefs Web Analytics that received traffic in May 2026:

28% of those domains publish a valid llms.txt file, about 38,000 sites.
97% of those files received no requests in May 2026, from bots or from people.
The remaining 3%, about 1,100 domains, received all the traffic measured. 96% of those requests came from bots.
Only 19.5% of requests came from identified AI tools, with GPTBot first and Claude-Code second. 12% came from tools that study llms.txt (checkers, GEO and AEO tools, researchers) rather than use it.

Ahrefs notes that its customers are more technical than the web at large, so the 28% adoption figure is an upper bound.

What Google says, in its own words

Google leaves no room for doubt. Its guide to optimizing for generative AI features lists llms.txt among the things you can ignore: you don't need to create machine-readable files, AI text files or Markdown to appear in Google Search, including its generative AI features, "as Google Search itself doesn't use them". The guide adds that maintaining an llms.txt file for other services "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them."

That position covers Google Search. Other systems can read the file, and the Ahrefs data shows that some do, rarely. We cover what llms.txt can and cannot do in our AEO and GEO guide.

Reading isn't the problem. Acting is.

llms.txt tells an agent what to read. Agents can already read your pages: that is basic crawler work. What they do badly is act. To fill in a form or pick a size, an agent has to guess what each button and field is for, then simulate clicks. Chrome's documentation puts it this way: without guidance, an agent has to guess "the purpose of an interface element" (Chrome for Developers).

That is the problem WebMCP wants to solve.

WebMCP, in plain terms

WebMCP is a JavaScript interface that lets a page declare tools: a function, a natural-language description and a schema of the expected input. An agent visiting the page can then call the "add to cart" tool directly instead of hunting for the button.

Three points matter for decision-makers:

Status. WebMCP is a "Draft Community Group Report" published by the W3C Web Machine Learning Community Group. The latest version is dated 20 October 2026. The document says it plainly: it is not a W3C Standard, nor is it on the W3C Standards Track (WebMCP specification). Its editors include people from Microsoft and Google.
Confirmation. The specification does not require human confirmation before every action. It provides an annotation, consequentialHint, that flags a tool with real-world consequences, such as booking or paying. The agent then decides whether to ask for your confirmation. ChatGPT, for example, asks before purchases, sending messages or sharing personal information (OpenAI).
Stability. The text is still moving. The "declarative" version, which turned an HTML form into a tool with a few attributes, was removed from the specification on 8 October 2026 "for now", while remaining available in Chrome's trial.

Where WebMCP stands in October 2026

Who	Status	Source
W3C	Community Group draft, not on the Standards Track, version of 20 October 2026	Specification
Chrome	Origin trial since Chrome 149 (June 2026), plus a developer flag. Extension requested on 28 September 2026, no decision to ship	Chrome, blink-dev
Edge	Its own trial until 30 March 2027. Microsoft states the feature may never be enabled beyond the trial	Microsoft Edge
ChatGPT	The built-in browser in the ChatGPT desktop app calls WebMCP tools ("site tools"), JavaScript API only	OpenAI
Shopify	WebMCP tools live on every Liquid storefront since 5 August 2026 (catalog, cart), and at checkout since 28 September 2026, with nothing to install	Shopify, Shopify
Gemini in Chrome	Support announced at Google I/O 2026, not shipped yet	Chrome

In short: one major agent calls these tools today, Chrome and Edge are only testing, and the text is still changing. That is no reason to rebuild a site. It is a reason to look at how agents already cope without it.

The 5-minute test

This test needs no technical tools. It shows what an AI agent can do on your site today, with or without WebMCP.

1. Pick an agent that drives a browser

For example Claude in Chrome, an extension available on paid Claude plans, or ChatGPT, whose agent fills in forms in a cloud browser or in the desktop app's browser. Plans and countries vary: use the one you have access to.

2. Prepare test data

Never test with real information. Use a name like "AI Agent Test", a test email address, a low-cost product or a staging site if you have one. Warn the team that receives contact requests, and stop before any payment.

3. Give a short instruction, with no help

Phrase it the way a customer would, without saying where to click:

"On [your site], add [product] in size M to the cart and go to the checkout page. Don't pay."
"On [your site], send the contact form with this test data: …"
"On [your site], find the price of [service] and the delivery time."

4. Watch, don't step in

For each instruction, note:

What to note	Example
The step where the agent stops	Size selection
What it says	"I can't find the size selector"
Repeated attempts	Three clicks on the same block
Silent errors	Phone number typed into the "Company" field

5. Fix, then run the test again

The same test, with the same instructions, after each round of fixes gives you a simple measure: does the agent get further than before?

Where agents get stuck, and what to fix

Most blockers need neither a new standard nor a new file. They are accessibility and HTML issues that also get in the way of people who browse with a keyboard or a screen reader (WCAG).

Where the agent stops	Why	What to do
The captcha or bot protection	That's intended: a captcha exists to stop automation	Keep it. Check that everything before it works, and that the person taking over understands what is left to do
The cookie banner	It covers the page or blocks scrolling	Visible "Accept" and "Reject" options as real buttons, and a banner that doesn't hide the content
Fields without labels	Placeholder text doesn't say what the field expects once it's filled in	A `<label>` tied to each field, plus `autocomplete` attributes (name, email, phone)
Fake controls	Sizes, colours, dropdowns or calendars built from plain blocks, with no button or list role	Native elements (`<button>`, `<select>`, radio buttons) or the right ARIA roles
A greyed-out button with no explanation	The agent doesn't know what is missing	A visible message saying which field is blocking
Errors not tied to fields	"Error" at the top of the page, without saying where	A message under the field concerned, linked with `aria-describedby`
Content that only appears after an action	Prices, lead times or variants loaded on hover or by infinite scroll	Key information in the page HTML, and real pagination links
Pop-ups	A newsletter sign-up interrupts the flow	No pop-ups during a cart or form flow

For more on page structure, see our method for designing an answer-first page and our 30-minute AI visibility audit.

So what should you do?

llms.txt: don't buy it as a product. It is a ten-minute file. We include it at no extra cost in our GEO and technical setups, and we publish one ourselves on gnsdf.com, because it costs almost nothing and a few bots, such as GPTBot and Claude-Code, do read some of these files. But it will not change your visibility in Google, and nobody can promise it will get you cited by ChatGPT.

Your user journeys: fix what the test reveals. Forms, selectors, error messages. These fixes serve AI agents, your visitors and your accessibility at the same time.

WebMCP: don't rebuild anything. Two exceptions:

You're on Shopify. WebMCP tools for catalog, cart and checkout are already live on Liquid storefronts, with nothing to do. Run the add-to-cart test: your theme is what gets exercised. See our Shopify development page.
You have one high-value flow, such as a quote request or a booking. A limited prototype can make sense, as long as you first check that the browser supports the API, flag committing actions with consequentialHint, and keep the code in one place, because the specification is still changing.

Re-run the test every quarter. Agents change quickly. For the bigger picture (structured data, agentic commerce, a 20-point checklist), see our guide Is your website ready for AI agents?

Common questions

Do I need an llms.txt file in 2026?

It is not required. Google states that its Search ignores llms.txt files, and according to Ahrefs, 97% of those files received no requests in May 2026. It is cheap to publish, but it should come after technical SEO and content, not instead of them.

Is WebMCP a W3C standard?

No. WebMCP is a draft published by the W3C Web Machine Learning Community Group, with the latest version dated 20 October 2026. The document states that it is neither a W3C Standard nor on the W3C Standards Track.

Which agents use WebMCP today?

As of October 2026, the built-in browser in the ChatGPT desktop app calls WebMCP tools. Chrome, from version 149, and Edge, from version 150, offer only a trial. Gemini in Chrome has been announced but not launched.

Should an AI agent be able to get past my captcha?

No. A captcha exists to stop automation: if the agent stops there, the captcha is doing its job. Keep it, and check instead that everything before it works, so the person taking over has only one step left.

Is my Shopify store already WebMCP-ready?

Yes, for Liquid storefronts. Shopify turned on WebMCP tools for catalog and cart on 5 August 2026, then for checkout on 28 September 2026, with nothing to install. Agents only call them if they support WebMCP.

Run the test with us

Genesis Digital Factory builds and optimises Webflow, Shopify, WordPress and PrestaShop sites, directly or white-label for agencies, at a fixed price. If you want to know where an AI agent stops on your site or a client's, get in touch: we run the test, list the blockers and quote the fixes at a fixed price.

Further reading: our AEO services, our agentic AI services and white-label GEO for agencies.