Updated
SEO got you found. GEO got you cited. This is step three: when an AI agent arrives on your site to enquire, book or buy, can it finish? I test that journey with real agents, then build the WebMCP tool layer that makes it reliable.
Agent readiness is whether an AI agent can complete real tasks on your website: request a quote, book a call, buy the right product. WebMCP is the proposed web standard behind it. A site exposes its actions as structured tools with natural language descriptions and input schemas, so agents call them directly instead of scraping the DOM and guessing at buttons and forms.
This is the whole journey. A customer types the job into their AI assistant. The assistant opens your site in its own browser, finds the tools your site exposes, and does the work. The customer just confirms. Pick a site type and watch the demo run.
Get me a conveyancing quote for a flat in Leeds. Email is sam@buyer.co.uk.
Smith & Vale Solicitors keep coming up for Leeds conveyancing. Opening their site to get you a real quote.
Quote request is filled in and ready. Confirm and it goes to the firm.
Buy me an internal oak fire door, 838mm wide, FD30 rated.
Opening oakdoor-store.co.uk to find that exact variant.
In the basket: oak, 838mm, FD30, £189. Confirm to check out.
Book me a hygienist appointment next week, mornings only.
Opening brightsmile-dental.co.uk to check real availability.
Tuesday 09:30 is held for you. Confirm and it is booked.
The customer's half happens in whichever assistant they already use. ChatGPT documents WebMCP support in its browser today, and Chrome's origin trial covers agents built on Chrome. The website half is the layer I design and build. Where the buying side of this is heading, including the Stripe and OpenAI payment protocol, is covered in my agentic commerce guide.
This is not one vendor's experiment. The two biggest agent surfaces publish their own WebMCP documentation, and OpenAI is running a public WebMCP Challenge for developers building agent-ready web apps. The two implementations cover different halves of the standard, and that detail decides how your site should be built.
Chrome runs WebMCP as an origin trial with two routes for tool registration: an imperative JavaScript API and a declarative API that turns annotated HTML forms into tools. Google also ships a free Model Context Tool Inspector extension that lists and test-calls a page's tools.
Read Chrome's WebMCP docsChatGPT's browser implements WebMCP as Site tools. Its own docs state the form-based API is not supported, so tools must be registered in JavaScript on the top-level page of your application. Sites that only annotate forms are invisible to it.
Read ChatGPT's Site tools docsThe spec is experimental: it lives in the W3C Web Machine Learning community group on GitHub, is actively maintained, and can change. That is why I sell agent readiness, with WebMCP as one ingredient in it.
Read the WebMCP explainerA browser agent on an ordinary website reads the page, simulates clicks and improvises a workflow. That works until it does not. These are the journeys where guesswork breaks, and every one of them is a lost enquiry that never shows in analytics.
| Journey | Where the agent stalls | What agent readiness changes |
|---|---|---|
| Quote requestLaw firms, trades, agencies | Multi-step forms, vague field labels, a submit button named something clever. | One named tool with described inputs. The agent knows exactly what to send. |
| BookingClinics, dentists, consultants | Date pickers and slot grids built in JavaScript that an agent cannot read reliably. | A check-availability tool returns real slots. A booking tool takes a chosen one. |
| Product choiceComplex ecommerce with variants | Compatibility rules, size guides and variant pickers spread across tabs and popups. | A find-the-right-variant tool answers from your data, not from page scraping. |
| Account tasksSaaS dashboards and portals | Actions hidden behind menus and state the agent cannot see from the page. | Tools scoped to login state, with confirmation steps on anything sensitive. |
Not a mock-up. In a browser where WebMCP is enabled, the tools below are live on this page right now: two registered in JavaScript, one declared on the enquiry form itself. The status line is read from your browser.
Returns the three service tiers with prices in GBP, so an agent can answer cost questions from source instead of summarising the page.
{ "inputSchema": { "type": "object", "properties": {} } }
Returns what each tier includes, who provides the service and how to get in touch, as structured text an agent can relay accurately.
{ "inputSchema": { "type": "object", "properties": {} } }
The enquiry form below is annotated with the declarative API. In Chrome, an agent can fill it as a tool call. Sending still needs a human press of the button, on purpose.
{ "required": ["name", "email", "message"] }
If the status line says WebMCP is not enabled, the page still works for you the normal way. The tools exist for the agents, and this site's origin trial registration switches them on for normal Chrome visitors with no flag needed. That is the point of the service.
Every engagement starts with the audit, because it delivers value even if you never add a single tool: it shows where real agent journeys break today. Implementation and monitoring follow only where the audit proves they are worth it.
I run AI agents against your site's money journeys: quote, booking, purchase, variant choice. You get a report of where they stalled, what they misread, which actions have no agent-usable path, and a prioritised fix list.
I design and ship the tool layer: which 5 to 15 actions to expose, tool names, descriptions, JSON Schemas and outputs, wired to your site's existing logic, with confirmation steps on sensitive actions, tested against both the Chrome and ChatGPT implementations.
Site deploys silently break tools. Monthly, I re-run the agent journeys, check every tool still registers and works, watch for new agent platforms and send a short report of what changed.
WebMCP is a conversion layer. An agent has to reach your site before tools matter. Acquisition is what AI search optimisation and GEO are for, and this pairs with them.
WebMCP is a proposal in an origin trial, not a finished standard. That is why the service is agent readiness, with WebMCP as one ingredient. The audit findings stay useful whatever happens to the spec.
Agent traffic is small today and growing. The honest pitch is first-mover positioning plus an audit that pays for itself in ordinary conversion findings, not a promise of agent-driven revenue this quarter.
Nobody controls what an agent platform does next. What I control is that your actions are exposed clearly, tested against the implementations that exist, and re-tested when either side changes.
This form is itself a WebMCP tool. If an agent brought you here, it can fill the fields. Either way, a person presses send, and I reply to the email address given.
Prefer a conversation first? Use the main contact page and mention agent readiness.
WebMCP is a proposed web standard that lets a website expose its actions as structured tools for AI agents. Instead of an agent guessing at buttons and forms, the site declares what can be done, what inputs each action needs and what comes back. Chrome runs it as an origin trial and ChatGPT implements it as Site tools.
An agent readiness audit runs AI agents against your website and tests whether they can complete the journeys that make you money: request a quote, book a call, buy the right product. The report shows where agents stalled, what they misread and which fixes matter most, in priority order.
No. WebMCP is a conversion layer, not a discovery layer. An agent has to reach your site before tools matter, which is what SEO and GEO are for. What WebMCP changes is whether the visit ends in a completed enquiry, booking or order instead of an abandoned one.
The agent readiness audit is a fixed 450 pounds. WebMCP implementation projects start at 1,800 pounds depending on how many actions need a tool layer. Ongoing monitoring, which re-tests agent journeys after site changes, starts at 300 pounds per month.
Two implementations are documented by their vendors: Chrome, where WebMCP runs as an origin trial with a JavaScript API and an HTML form API, and ChatGPT, whose browser implements the standard as Site tools using the JavaScript API only. The standard itself is still a proposal and can change.
Not necessarily. Agents can still drive an ordinary website by reading the page and clicking, and the audit tests exactly that. Tools make the difference on the journeys where guesswork fails: ambiguous labels, multi-step forms, product variants and anything where a wrong click costs money.
For 450 pounds I will find out, journey by journey, and hand you the fix list. The tool layer comes after, and only where it earns its keep.