Agent-ready website. What is already a rule, what is a proposal on trial, and what to do first.
An AI agent booking or filling in a form for a customer gets stuck where a person gets stuck: on the unnamed button, on the unlabelled field. Here is the ten-point checklist, with the maturity of every line.
A website is ready for AI agents when an assistant can use it as well as read it: find the button, fill in the form, reach the confirmation. The foundation is accessibility, which already serves people.
The rest is young. WebMCP, the interface through which a page offers tools to agents, is a draft from a W3C Community Group on trial in Chrome and Edge; the Agentic browsing category in Lighthouse is experimental, as Google itself says.
Across 200,000 heavily visited domains, 78% have a robots.txt and 4% state how they want to be used by AI, according to a Cloudflare study.
The ten-point checklist can be downloaded below, in Word and PDF, with a maturity label on every line.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The 10-point agent-ready checklist, to download
A sheet with ten checks, ordered from the most solid to the most experimental, each with its maturity label and how to verify it, in Word to fill in and PDF to print. The first four lines can be checked without paid tools, using a browser and a phone.

| File | Contents | Link |
|---|---|---|
| Wordto fill in, in Italian | The ten checks with maturity and verification, and what each label means. | |
| PDFto print, in Italian | The same checklist, to tick off with a pen. |
| N. | What to check | Maturity |
|---|---|---|
| 1 | Buttons and links are real buttons and real links, and every field has a label. | Basic |
| 2 | The page does not shift while loading: no banners pushing the buttons down. | Basic |
| 3 | Prices, opening hours, address and contact details are in the HTML, readable without JavaScript. | Basic |
| 4 | The contact or booking form can be completed through to confirmation, without asking for details a foreign customer does not have. | Basic |
| 5 | The robots.txt says what AI search and training crawlers may do. | IETF standard (RFC 9309) |
| 6 | The llms.txt file, if present, responds without server errors. | Emerging convention, optional |
| 7 | JSON-LD structured data describes the company, the offer, prices and opening hours. | Shared vocabulary (schema.org) |
| 8 | The main forms are exposed as WebMCP tools. | Proposal, on trial |
| 9 | The ard.json manifest lists the domain's MCP servers, agents or skills, if there are any. | Specification 0.91 |
| 10 | The /agents.md file for shopping agents, which Shopify publishes automatically on every store. | A platform's product |
On 7 September I published an analysis of the websites of 142 Italian companies that also sell abroad. In 61 of them a foreign buyer had no real way to get in touch. The form was broken, or missing, or asked for details a foreign buyer does not have: an Italian VAT number, a certified email address (PEC).
The same day, on the same sites, I counted which ones had the file that explains the company to an assistant: 28 out of 143, and in 24 a plugin had written it. I closed the post like this: “An assistant doesn’t browse a site. It reads it, and reports what it finds written there”. The full argument is in the post on the 142 sites, in Italian.
An agent writing to that producer for a German buyer would get stuck on the same PEC field. So line 4 of the checklist is a check done by hand, from a phone, pretending to be a foreign customer. No new file replaces it.
What an agent-ready website means
Google separates two moments: the agent that searches, where SEO rules apply, and the agent that uses the site on behalf of a person, which is what this page covers. Chrome for Developers says so in its toolkit for agent-ready sites of 22 June 2026. In the first moment the site has to be found; in the second it has to be usable.
For the first part there are already guides on generative engine optimization and on robots.txt for AI crawlers. This page covers the second: an agent opening the page, choosing a button, filling in a form and reaching a confirmation. ChatGPT’s agent does this, for example, in the browser of the desktop app, where OpenAI moved the features of Atlas, the browser it shut down in August 2026 (details in the guide on how to show up in ChatGPT).
So how does an agent see the page? web.dev, Google’s site on web quality, puts it in one line: “Agents can view your website in 3 primary ways: screenshots, raw HTML, and the accessibility tree”. Each of the three views has its blind spots, and the checklist starts from there.
The screenshot sees what a person sees, but goes wrong if a transparent element covers the button or the layout moves. The raw HTML sees the text before JavaScript runs, so it misses prices and contact details written by a script: the subject of the page on JavaScript and AI crawlers. The accessibility tree sees the names of buttons and fields, and a button with no name is a mute rectangle to the agent.
What is WebMCP, and where does it stand?
WebMCP is a proposed browser interface through which a page offers AI agents tools in JavaScript: functions with a description and a data schema. Instead of guessing where to click, the agent calls the tool “book a table” with a date and a number of guests. The specification is written by Brandon Walderman and Samuel Dallstream of Microsoft, and Khushal Sagar and Dominic Farolino of Google.
The document states it plainly: “It is not a W3C Standard nor is it on the W3C Standards Track”. The document is a draft report of the W3C Web Machine Learning Community Group, regenerated with every change. Presenting it as an approved standard gets several steps ahead of a process that has not even begun.
There are two interfaces. The imperative one registers tools in JavaScript with document.modelContext.registerTool(). The declarative one, based on ordinary HTML forms, is still being explored according to the project’s explainer.
WebMCP vs MCP: similar names, different places. An MCP server lives behind the scenes and talks to the agent over its own channel. WebMCP tools live inside the open page, share the session and the interface with the user looking at it, and need no separate server.
| Where | WebMCP status as of 10 November 2026 | Source |
|---|---|---|
| Chrome | Origin trial live from Chrome 149, announced on 9 June 2026. Before that, behind a flag in Chrome 146 Canary. | Chrome for Developers |
| Edge | Origin trial live in Edge 150. | Implementation status |
| ChatGPT Desktop | Support announced. | Implementation status |
| Brave | Experimental support in the Leo assistant. | Implementation status |
| Meta Ray-Ban Display | Listed as coming. | Implementation status |
| Firefox e Safari | No implementation, only an open request for a standards position. | Implementation status |
How to test WebMCP today
From Chrome 149 a site can sign up for the WebMCP origin trial: a time-limited programme in which Google lets a new feature be tried on real sites, possibly with usage limits. The origin trial token is added to the page, and the feature switches on for visitors with the right browser.
The use cases Chrome lists are concrete: complex forms, for example when one field wants the full name and another wants first and last name separately, diagnostic tools, and calendars. For a small business the calendar is the closest case: a booking, an appointment, a winery visit.
Before touching the code, ask an honest question. WebMCP does not exist on Firefox and Safari, and the origin trial has an end date. A form that works well for people and for the accessibility tree already works for today’s agents; a WebMCP tool adds a shorter route for those that support it.
The Lighthouse audit for agentic browsing
From Chrome 150, Lighthouse, Google’s tool for measuring page quality, has a new category, Agentic browsing, with deterministic checks on accessibility, layout stability and WebMCP integration. Google itself warns: “at the time of publication, Agentic Browsing is informational and unbenchmarked”.
The result is a fraction, how many checks pass, with pass or fail and a few warnings. There is no mark out of 100, and no comparison with competitors. The documentation adds that the category and WebMCP support are experimental and based on proposed standards; the WebMCP checks require signing up for the origin trial.
| Agentic browsing check | What it looks at | Checklist line |
|---|---|---|
| Accessibility for agents | Every interactive element has a name in the accessibility tree. | 1 |
| Layout stability | Page shifts (CLS) that make the agent click the wrong thing. | 2 |
| llms.txt | Server errors on the file. If it is missing, the check is Not Applicable. | 6 |
| Registered WebMCP tools | Whether the page offers tools to agents. | 8 |
| Forms missing declarative WebMCP | HTML forms that do not use the declarative version. | 8 |
| Schema validity | Whether the tool descriptions are well formed. | 8 |
| ard.json | The manifest of agent resources, validated against version 0.91. | 9 |
For a small business the two lines that really count are the first: accessibility and stability. The other five, on a site with no WebMCP and no MCP server, return “not applicable” or a warning, and nothing is wrong. A low result on checks for a proposal on trial does not mean the site is broken.
What is agents.md? The file for coding agents and Shopify’s file
Two different files go by the same name. AGENTS.md is an open Markdown format for guiding coding agents inside a repository: commands, code style, tests. The project’s site describes it as “A simple, open format for guiding coding agents, used by over 60k open-source projects”.
It grew out of work by OpenAI Codex, Amp, Google’s Jules, Cursor and Factory, and today it is stewarded by the Linux Foundation’s Agentic AI Foundation, which made it a founding project alongside Anthropic’s MCP and Block’s goose. This file is not published on the website: it lives in the code, and the agents writing that code read it.
The second belongs to Shopify. Since 28 May 2026 every Shopify store has a public /agents.md file for shopping agents, with the addresses for reading products, collections and search, the store’s policies and the entry points for commerce protocols. How to customise it, together with llms.txt, is in the guide to llms.txt on WordPress and Shopify.
Shopify itself warns that it is a different file from the AGENTS.md in the theme root, which serves developers and is not published. Outside Shopify, no source read for this page asks a website to add an /agents.md. Selling it as compulsory for everyone confuses the two files.
llms.txt and ard.json: what Google checks
Lighthouse calls llms.txt an emerging convention and treats it as optional: if the file is missing, the check comes out Not Applicable. The documentation puts it this way: “the audit is marked as Not Applicable (N/A), as providing the file is optional at the moment”. The check flags server errors only. What to put in the file is in the guide to llms.txt.
ard.json is the manifest of Agentic Resource Discovery, a specification for listing a domain’s agent resources: MCP servers, A2A agents, skills. Lighthouse looks for it in four places: the Agentmap: directive in robots.txt, a <link rel="ard">, the Link header or /.well-known/ard.json.
A small business with no MCP server and no agents of its own has nothing to list, and does not need the file. Line 9 of the checklist is there because the check exists and anyone opening Lighthouse will see it, not because it has to be filled in.
How many sites are ready: the Cloudflare study
In April 2026 Cloudflare scanned the 200,000 most visited domains in its Radar observatory, leaving out categories where readiness does not matter, such as redirects and ad servers. This is a study by a vendor that sells the tools it describes, and should be read as such. But it is the only large-scale public figure.
| What Cloudflare measured | What it means | Websites |
|---|---|---|
| robots.txt present | Almost everyone has one, but written for traditional search engines. | 78% |
| Content Signals | robots.txt lines stating how the site wants to be used by AI. | 4% |
| Markdown on request | The page answers in text/markdown if the agent asks for it. | 3.9% |
| MCP Server Card e API Catalog | Cards describing servers and APIs for agents (RFC 9727). | fewer than 15 sites |
The figure says two things. The new pieces are almost absent even on the largest sites in the world, so a small business skipping them is behind nobody. And robots.txt, line 5, is almost everywhere but rarely tells agents anything.
Where a small business starts
With lines 1 to 4, which are accessibility and common sense and already help both people and agents today. Then line 5, robots.txt, and line 7, structured data. Lines 8 to 10 are worth a look once a quarter, because they change faster than the rest.
I learnt this on my own site before writing it here. In August ChatGPT could not find the founder of itria.io, and I already had the file for assistants and the structured data: the usual advice. Among the faults I found later, one sat lower down: the email address on the contact page, which did not exist for anything reading the HTML.
The missing email address. Cloudflare’s obfuscation removed the address from the HTML and put it back with a script: to anything reading the raw HTML, the contact page had no contact details. The same fault on 14 out of 121 producers’ sites, in the piece on the email address AI assistants can’t see and in the study of producers’ websites.
The form that stops. Out of 142 sites of companies selling abroad, 61 with no real way for a foreign buyer to make contact, from the post of 7 September 2026.
The file written by someone else. On the same sites, 28 out of 143 with the file for assistants, 24 generated by a plugin.
The result. Between 11 October and 7 November 2026 Bing’s AI Performance report counted 268 citations of itria.io pages in Copilot, and Search Console 3,550 impressions in AI Overviews and AI Mode.
The AI crawler glossary collects the names to recognise in the logs, and the index of guides is in marketing for small businesses. How all this is built on a real company, with the right questions to own, is on the page about presence in search engines and AI assistants.
Questions and answers
What does it mean for a website to be ready for AI agents?
An agent, an assistant acting on behalf of a person, can use the site: it finds the right button, fills in the form, reads the price and opening hours, and reaches the confirmation. Google calls this the stage where the agent uses the site, as distinct from the stage where it searches.
The foundation is accessibility: real buttons, labelled fields, a page that does not shift.
What is WebMCP?
A proposed browser interface that lets a web page offer tools to AI agents, meaning functions with a description and a data schema, in place of clicks on the screen. Editors from Microsoft and Google write it in a W3C Community Group.
It is not a W3C standard and not on the standards track. As of 10 November 2026 it is on trial in Chrome 149 and Edge 150.
How do you enable WebMCP in Chrome?
From Chrome 149 a site can sign up for the WebMCP origin trial, a time-limited programme in which Google lets the feature be tried on real sites, possibly with usage limits. Before that it existed only behind a flag, in Chrome 146 Canary.
Firefox and Safari have not implemented it.
Does a website need an agents.md file?
Depends what you mean. AGENTS.md is a file for coding agents inside a code repository, and it is not published. Shopify, on the other hand, has published an /agents.md on every store since May 2026, for shopping agents, and does so automatically.
Outside Shopify no official source asks you to add one.
Does Lighthouse give an agent-readiness score?
Not a mark out of 100. The Agentic browsing category, from Chrome 150, shows how many checks pass, with pass or fail and a few warnings. Google describes it as experimental, informational and unbenchmarked.
Use it to find faults, not to compare yourself with competitors.
Notes on sources
- W3C Web Machine Learning Community Group, WebMCP, Draft Community Group Report, version of 10 November 2026: editors and status of the document, source of the quoted sentence. Explainer and implementation status, read on 10 November 2026.
- Chrome for Developers, WebMCP origin trial, 9 June 2026, and A developer toolkit to make your website agent-ready, 22 June 2026, CC BY 4.0 licence.
- Chrome for Developers, Lighthouse documentation: Agentic browsing scoring, the llms.txt audit and the ard.json audit, read on 10 November 2026, CC BY 4.0 licence.
- web.dev, Build agent-friendly websites, updated 1 April 2026, CC BY 4.0 licence: the three views of the agent.
- AGENTS.md and Linux Foundation, formation of the Agentic AI Foundation: the founding projects and the more than 60,000 open-source projects.
- Shopify, agents.md.liquid and changelog of 28 May 2026.
- Cloudflare, Introducing the Agent Readiness score, 17 April 2026: a vendor study, the 200,000 most visited domains according to Cloudflare Radar. Reread on 10 November 2026.
- IETF, RFC 9309, Robots Exclusion Protocol, September 2022, on the standards track.
- The cases come from the post of 7 September 2026 and from the itria.io labs; the itria.io figures from Search Console (generative AI report) and Bing Webmaster Tools, window 11 October to 7 November 2026.
The first four lines of the checklist take an afternoon. Getting chosen when the agent compares you with your competitors is the work that follows. Positioning is built starting from the market.
Mattia Esposito, who wrote this guide, also works with Itria on positioning: in-depth market analysis to find where competitors are exposed and take that ground. He did it on his own site first. In August, asked who had founded itria.io, ChatGPT couldn’t find him. From 11 October to 7 November 2026 itria.io pages appeared 13,803 times in Google and Bing results and 3,818 times in AI answers: Google’s AI Overviews and AI Mode, and Microsoft Copilot.