What is WebMCP, and which AI tools actually use it?
Hi, I'm Evren, an SEO and GEO consultant in Berlin. Over the past few weeks I tested WebMCP in depth on a large content site. Since then, one question keeps coming up: "Is this the new SEO?"
Short answer: no. The long answer is this post. I'm writing it because most of what I found online was either a summary of the spec or pure hype, and almost nothing looked at it from the search side. So here is what WebMCP is, which AI tools actually use it today, and where it does (and doesn't) touch your visibility.
This space changes every few weeks. I retest and update this page when something changes. You'll find every update in the changelog at the end.
What WebMCP is, in plain words
Today, when an AI agent visits your website, it behaves like a tourist who doesn't speak the language. It reads the page, takes screenshots, guesses which button does what and clicks around until something works. Sometimes it gets there. Often it takes a while.
WebMCP gives your site a way to hand that tourist a menu. The page registers a short list of tools in the browser, each with a name, a description and the inputs it expects. The agent reads the list, picks the right tool and calls it directly. No guessing which button is the search box.
A few facts worth knowing:
- It is a proposed web standard, developed in the W3C Web Machine Learning Community Group, with Google and Microsoft behind it.
- It runs in the browser, inside the page, with the visitor's session. That is the main difference from a classic MCP server, which lives on your backend.
- Tools only exist while the page is open. Close the tab and they are gone.
- There are two ways to define tools: the imperative API (JavaScript) and the declarative API (attributes on HTML forms).
Here is what a simple read only tool looks like with the JavaScript API. The example site and tool are made up, but the shape is real:
JavaScriptif (document.modelContext?.registerTool) {
document.modelContext.registerTool({
name: 'search_articles',
description: 'Search published articles on example.com by keyword. ' +
'Returns title, URL and publish date. Does not return full text.',
inputSchema: {
type: 'object',
properties: {
query: { type: 'string', description: 'Words to search for' }
},
required: ['query']
},
annotations: { readOnlyHint: true },
execute: async ({ query }) => {
const results = await searchArticles(query); // your existing search logic
return JSON.stringify(results);
}
});
}Two things in there matter more than they look. The description is what the agent reads to decide whether to use your tool, so it deserves the same care as a title tag. And readOnlyHint: true tells the agent that the tool only reads information, so it doesn't need to stop and ask the user for permission before calling it.
Which AI tools actually use it today
This is where the hype and reality drift apart. Plenty of articles say "AI agents can now use your website's tools". In practice, as of November 2026, one consumer product does it reliably.
| Tool or browser | Status | Notes |
|---|---|---|
| ChatGPT desktop app | Yes | Called "site tools". Only in the app's built in browser, in ChatGPT Work or Codex. Not in the web version, not in Chrome. |
| Claude | No | No WebMCP support announced. In my tests, Claude in Chrome did not discover WebMCP tools. |
| Gemini in Chrome | Announced | Google said at I/O in May 2026 that support is coming "soon". No date yet. |
| Chrome | Experimental | Testable with a flag. Origin trial started with Chrome 149. This is the API for developers, not an agent by itself. |
| Microsoft Edge | Origin trial | Runs its own WebMCP origin trial, listed until March 2027. |
| Safari (WebKit) | Opposed | WebKit's official standards position is "oppose", citing privacy, security and API design concerns among others. |
| Firefox (Mozilla) | Neutral | Mozilla's official standards position is "neutral". |
So if someone tells you their WebMCP tools are "used by AI", the honest version today is: used by ChatGPT desktop users in the right mode, plus developers testing in Chrome.
The ChatGPT details that matter
Since ChatGPT is the only real consumer of WebMCP right now, its rules decide whether your tools get used at all. These are the ones I'd check first:
- Desktop app only. Site tools run in the built in browser of the ChatGPT desktop app. The browser version of ChatGPT doesn't use them.
- Work or Codex mode. OpenAI's documentation names ChatGPT Work and Codex as the modes that discover and use site tools.
- Model matters. OpenAI only says a "supported model" is needed and doesn't list names. In my tests, GPT 5.6 Terra and GPT 5.6 Sol used site tools; GPT 5.6 Luna did not.
- Register on the top level page, in JavaScript. OpenAI's developer page mentions HTML form attributes and tool discovery in iframes, but its help center says tools provided only by embedded content are not supported. Until both agree, the safe choice is JavaScript on the main page.
- Users can switch it off. In the desktop app: Browser settings, Permissions, Enable site tools.
- You can see what was used. An arrow in the address bar shows the page's tools: gray when tools are available, blue while ChatGPT is using them. A "recently used" view shows which tools were actually called.
One detail that confused my first tests: ChatGPT runs each task in its own browser tab, which you may not see. If you say "this page" in your prompt, it may not mean the page you are looking at. Give it the full URL.
WebMCP is not SEO
Let me say it plainly: WebMCP will not help you rank. Not in Google, not in AI answers. And I say that as someone who would love a new ranking factor to sell.
Here is how I think about it, in five points.
1. Search brings the agent; WebMCP helps once it is there
Picture the journey of an AI agent working for a user:
- The user asks a question or gives a task.
- The agent searches or retrieves sources. This is where SEO and GEO decide whether your page is found, chosen and cited.
- The agent opens your page in a browser.
- This is where WebMCP decides how quickly and accurately the agent can do the task on your page.
Step 2 happens before your tools even exist. If your page doesn't make it through step 2, your WebMCP tools are never seen.
2. There is no ranking or citation signal in it
Neither OpenAI's developer docs nor its help center connect site tools with ranking, search visibility or citations. Google presents WebMCP as a browser API for agents, not as a search feature. Tools only exist in a live browser tab with the API available; the crawlers that build search indexes don't run your tools, so there is nothing for them to rank.
3. What it can improve: task success, speed and accuracy
Where WebMCP does make a difference is after the click. In my tests, when a site had no tool for its main job, the agent had to read page after page, paginate, and sometimes guess things the site knew perfectly well, like which content is free and which needs a login. That took tens of seconds per question. A well designed tool returns the same answer in one call, with the facts stated instead of guessed.
That is a user experience gain for people who use agents. It is real. It is just not SEO.
4. Where SEO and WebMCP do meet
There is overlap, and it is in the boring, solid parts of technical SEO:
- One source of truth. What your tool returns should match your page content and your structured data. If schema says an article is free and your tool says it needs a subscription, the agent gets two answers.
- Clean HTML. Unique ids, hidden elements that are properly marked as hidden, menus that don't pretend to be content. Agents, screen readers and crawlers all benefit from the same hygiene.
- Knowing your content inventory. Canonical URLs, access levels, dates. You need these for good tools and for good SEO.
5. Measuring it is a separate project
Agents work from the user's own browser, so in analytics their visits look like normal visits. If you want to know how often your tools are called, you need your own logging. And if you log tool calls, update your privacy policy first.
Should you add it now?
My rule of thumb:
Worth trying now if your users already do multi step tasks on your site (search and filter a catalog, compare options, look things up in documentation or a dashboard), and an agent visibly struggles with them.
Can wait if your main goal is traffic or citations, if nobody can maintain the tools when the site changes, or if your interactive parts live inside iframes.
If you do start:
- Start with read only tools and mark them with
readOnlyHint: true. - Keep the list short. Three clear tools beat fifteen vague ones.
- Write descriptions for an agent that has never seen your site.
- Never add a tool that types personal data into a form.
- Make tools tell the truth. If an action fails, the tool must say so.
How to test your own site
The quickest check takes two minutes in Chrome:
- Open
chrome://flags/#enable-webmcp-testing, set it to Enabled and relaunch Chrome. - Open your site and the DevTools console.
- Paste this to list the page's tools:
const tools = await document.modelContext.getTools();
console.table(tools.map(t => ({ name: t.name, description: t.description })));An empty table means the page registers no tools. A long table full of things like "close popup" means you have a different problem. I wrote a separate, step by step guide covering the console commands, ChatGPT prompts and the mistakes I see most often: How to test WebMCP on your website with Chrome and ChatGPT.
Try it on this site
This site has WebMCP tools too, so you can try everything above right here. There are two, both read only:
search_blog_postssearches and lists the posts on this blog.get_servicesreturns the SEO and GEO services I offer, as structured data.
All their data sits in a single JavaScript file on this site. Calling a tool makes no network request, sets no cookie and logs nothing. If your browser doesn't support WebMCP, the file does nothing at all.
With the Chrome flag enabled, open the console on this page and run:
Chrome consoleconst tools = await document.modelContext.getTools();
const services = tools.find(t => t.name === 'get_services');
console.log(await document.modelContext.executeTool(services, { area: 'geo' }));Or, in the ChatGPT desktop app in Work mode, try: "Open https://ntegrative.com/blog/ and use the site tools to list all blog posts. Tell me which tool you used."
Changelog
- First version. All statuses checked against the sources below.
Sources
- OpenAI: Site tools (developer documentation)
- OpenAI Help Center: Using site tools in the ChatGPT desktop app
- Chrome for Developers: WebMCP
- Chrome for Developers: WebMCP origin trial
- Chrome for Developers: Chrome at Google I/O 2026
- Microsoft Edge: WebMCP origin trial
- WebKit standards position on WebMCP
- Mozilla standards position on WebMCP
If you test WebMCP on your own site and see something different, I'd like to hear about it. That is exactly how a page like this stays accurate. Questions are welcome too: write to [email protected] or find me on LinkedIn.