The short version: Google announced WebMCP at I/O 2026, an open standard, backed by Google and Microsoft and discussed at the W3C, that lets a website expose typed tools an AI agent can call directly instead of screenshotting and clicking its way through a page. It is experimental, running in origin trials in Chrome 149, not yet a stable standard. I have spent two years getting sites cited by AI search instead of just ranked by it, and WebMCP reads to me like the next layer of the same shift: the web quietly reorganizing itself around a second kind of visitor that does not have eyes or a mouse.
I have spent the last two years teaching websites how to talk to AI search: writing the FAQ schema that gets pulled into an AI Overview, structuring pages so ChatGPT and Perplexity can cite them cleanly, building the llms.txt files nobody asked for a year ago and everybody wants now. So when Google announced WebMCP at I/O 2026, my first reaction was not surprise. It was recognition. This is the same shift, one layer down, moving from "can a machine read this page" to "can a machine act on this page."
WebMCP is still young. It is in origin trials in Chrome 149, which means developers can experiment with it on real sites, but it is not stable and it is not widespread. It still has to work through the W3C process before anyone can call it a finished standard. I am not writing this as a how-to, because there is not much to build yet that would survive the standard changing shape under it. I am writing it because the direction is worth sitting with now, while it still tells you something about where building for the web is headed.
What WebMCP actually changes
Right now, when an AI agent uses a website on someone's behalf, it mostly does what a very patient, very literal human would do with no hands: it takes a screenshot, tries to figure out where a button is, clicks a coordinate, waits, takes another screenshot, and reevaluates. It is scraping the DOM and guessing at intent from pixels and markup that were designed for a person's eyes, not a program's parser. WebMCP proposes something more direct. A site can expose a set of tools, each with a name, a plain-language description, and a JSON schema describing exactly what it needs and what it returns. An agent reads that tool list and calls the tool it needs with structured data, the same basic idea behind Anthropic's Model Context Protocol from late 2024, brought down to the browser and the individual website.
There are two ways a site can offer this. The Declarative API adds attributes to HTML you likely already have, like an existing form, so a simple interaction becomes agent-callable without a rewrite. The Imperative API is JavaScript, for anything more dynamic: a booking flow, a filtered search, a multi-step checkout. Neither replaces your site's interface. They sit next to it, as a second, structured way in.
Two Ways an Agent Gets Something Done on a Site
This is the shift in one picture: how an AI agent interacts with a website today, next to what WebMCP proposes instead.
- 1Agent opens the page
- 2Takes a screenshot
- 3Locates the cursor position
- 4Clicks, then waits for the page to respond
- 5Takes another screenshot
- 6Re-evaluates what changed, then repeats
- 1Site declares its tools: name, description, JSON schema
- 2Agent reads the tool list
- 3Agent calls the tool directly with structured data
- 4Site returns an instant, structured result
Early figures from Google and early coverage of Chrome's origin trial, not an independent benchmark. WebMCP is experimental and not yet a finished standard.
The pattern I keep seeing
The pattern I keep seeing in my own AEO and GEO work is that the sites that get cited are not the ones with the cleverest copy. They are the ones with the clearest structure: real heading hierarchy, schema that matches the content, answers written so a system can lift them cleanly instead of interpreting them. When I build a static site for a client now, I do not just ask whether the copy converts a human. I ask whether the markup would make sense to something that has never seen a browser or a mouse cursor. WebMCP takes that instinct one step further, from "can an agent read this" to "can an agent do something here without guessing."
[PLACEHOLDER: Dahlia adds real detail, e.g. a specific before/after observation from her own AEO/GEO audits on how structured markup changed AI citation behavior on a client or personal site.]
I run a multi-brand studio, so I am usually building the same kind of infrastructure for several sites at once: consistent schema, consistent semantic HTML, consistent content architecture. None of it was built with WebMCP in mind, because WebMCP did not exist when I built most of it. But it is the same groundwork. A site already legible to a language model reading it is closer to a site that could expose a tool to an agent calling it. The standard changes; the discipline of building something a machine can parse without friction does not.
A traffic number worth reading carefully
Here is a number that gets thrown around to make the agent-first web sound more advanced than it currently is: automated traffic passed human traffic on the web for the first time in 2024, at roughly 51 percent, according to Imperva's 2025 Bad Bot Report. It is real, and it is a meaningful marker. But it is easy to misread. Most of that 51 percent is not helpful shopping or booking agents doing errands for people. It is scrapers, credential-stuffing bots, and other bad automated traffic that has nothing to do with the kind of agent WebMCP is designed for.
What that number actually tells me is narrower: the web already spends enormous effort telling bad automated traffic apart from real visitors. WebMCP is not solving that problem directly, but it is an early attempt at giving good automated visitors a sanctioned front door, instead of leaving every agent, helpful or not, to get in the same way: by pretending to be a human clicking around.
The web spent three decades being designed for a person with a mouse. WebMCP is an early, unfinished attempt at also designing it for something with neither.
What this means for creators and small businesses
If you run a small business site or a creator page, the honest answer right now is: do not rebuild anything for WebMCP specifically. It is an origin trial in one browser, not a requirement. What is worth doing is the same thing that has been worth doing since AI search started citing sites instead of just ranking them: clear headings, real FAQ and Article schema, content organized around actual questions, and a structure that does not depend on JavaScript tricks or hidden text to make sense. That foundation is what any future agent-tooling layer, WebMCP or whatever eventually wins, will build on. I wrote the practical version, the actual steps for making a site agent-ready today, as the practical how-to on making your site agent-ready on the agency site.
For creators specifically, the more interesting shift is psychological before it is technical. For years the mental model was: build something a person will land on, get curious about, and click through. An agent-first web adds a second model running in parallel: build something a program can understand well enough to act on correctly, on the first try, with no one watching over its shoulder to catch a wrong click. Both will coexist for a long time. But the second is new enough that most people building small sites have not started thinking in it yet.
How I am thinking about my own work
I am treating WebMCP the way I treated the early AI search shift two years ago: not rebuilding around it yet, because it has not earned that, but paying attention and keeping my own sites structurally honest so they are not starting from zero once the standard, or whatever replaces it, matures. The pattern I trust most from this work is that sites already legible to a machine reader adapt fastest when the next machine-facing standard shows up. Structure now. Specifics later, once there is something stable enough to be specific about.
The takeaways, in one place
Frequently Asked Questions
WebMCP is a proposed open web standard, announced at Google I/O 2026 and backed by Google and Microsoft, discussed at the W3C. It lets a website expose typed tools, each with a name, a natural-language description, and a JSON schema, that an in-browser AI agent can call directly. It brings the idea behind the Model Context Protocol, which Anthropic introduced in late 2024, to the website layer itself.
It is experimental. WebMCP is running in origin trials in Chrome 149, which means developers can test it on real sites, but it is not a stable or widespread standard yet. It still has to go through the W3C process before it becomes something every site is expected to support.
No, and this gets misread often. Imperva's 2025 Bad Bot Report found that automated traffic passed human traffic for the first time in 2024, at roughly 51 percent. Most of that 51 percent is bad bots: scrapers and attackers, not helpful AI shopping or booking agents. The stat is a signal that non-human traffic is now the majority, not proof that agentic commerce has already taken over.
The groundwork is the same structural clarity that helps AI search cite you: clean semantic HTML, real schema markup, and content organized so a machine can parse it without guessing. That work pays off regardless of which agent standard eventually wins. For the practical, step-by-step version of this, I wrote a companion piece on the agency site.
Sources & Credit
- Chrome for Developers. WebMCP documentation, the primary source for the Declarative and Imperative APIs and the tool schema model.
- Chrome for Developers. Chrome at Google I/O 2026, on the announcement and the Chrome 149 origin trial.
- InfoQ. WebMCP: a web agent standard for Chrome, on Microsoft's backing and the W3C discussion.
- Imperva. 2025 Bad Bot Report: How AI Is Supercharging the Bot Threat, the source for the 51 percent automated traffic figure and the bad-bot majority within it.
