Your pricing calculator, your demo booking flow, and your product finder are the highest-intent pages on your entire site. To an AI agent browsing on a buyer's behalf, they are currently invisible. The agent can see the pixels. It cannot reliably use the tool.
That changed this week. OpenAI shipped WebMCP support in ChatGPT's desktop browser, and it is the first agentic web standard built specifically for the pages marketing teams own, not the backend systems engineering owns. Most B2B sites will not touch it for a year. That delay is going to cost them pipeline.
What WebMCP Actually Does
Model Context Protocol solved a real problem: it gave AI agents a standard way to call your backend systems, your CRM, your analytics warehouse, your internal tools. But MCP was never built for the public web. It assumes a server, credentials, and a defined integration.
WebMCP solves a different problem. It lets any webpage register JavaScript functions as tools an agent can call directly, complete with a name, a description, and a structured input schema. No backend integration required. No API key exchange. Just a script tag that tells a visiting agent exactly what actions are available on this page and how to invoke them correctly.
Compare that to what happens today. An agent lands on your pricing page and has to infer intent from a DOM tree never designed to be read by a non-human visitor. It guesses which button is "Get a quote" versus "Talk to sales." It fills form fields based on label proximity, not certainty. When your dev team ships a redesign, every one of those guesses breaks silently, and you have no idea your agent traffic just went to zero.
| DOM Guessing (Status Quo) | WebMCP Tool Call (New Standard) |
|---|---|
| Agent infers intent from visual layout | Agent reads a declared function schema |
| Breaks on every frontend redesign | Survives redesigns as long as the tool contract holds |
| No way to log what the agent actually did | Structured, loggable tool invocations |
| Works only as well as your HTML semantics | Works exactly as well as you define it |
end
Why Marketing Owns This, Not Engineering
It is tempting to route this straight to the dev backlog and forget about it. That would be a mistake for the same reason schema markup was a mistake to hand off without marketing input: the tools you expose are messaging decisions.
When you register a "compare plans" tool, you are deciding which comparison logic an agent sees: list price, or the discounted price after the objection-handling step your best reps use? When you register a "check availability" tool for a demo booking flow, you are deciding whether an agent can book directly or has to hand off to a human, and at what point in that flow you want a person, not a bot, taking over.
This site has already covered why AI purchasing agents are evaluating B2B sites right now and why checkout flows built for humans fail agentic buyers. WebMCP is the piece that closes the gap between "agents can visit your site" and "agents can transact on your site" without your team building a custom integration for every AI platform that shows up. ChatGPT supports it today. Given how fast Anthropic and Google have matched OpenAI on agent protocols this year, expect Claude and Gemini browser agents to follow within a quarter, not a year.
What to Expose First
Do not try to wrap your entire site in tools on week one. Start with the handful of pages where an agent completing the action correctly is worth real revenue, and where getting it wrong is worth real risk.
- Pricing and plan comparison: expose the actual current pricing logic, not a stale cached number
- Demo or meeting booking: let the agent check real availability and confirm a slot, not just find a form
- Product or use-case finder: turn your interactive quiz or configurator into a callable tool with a clean schema
- Content search: expose your docs or resource library search so agents cite the right page instead of a 2023 blog post
- Lead capture: define exactly what fields are required and what happens after submission, so an agent does not retry and create duplicate leads
end
Notice what is missing from that list: your entire marketing site. Blog posts, about pages, and career listings do not need tool schemas. They need to stay readable HTML. WebMCP is for the pages where an agent is trying to complete a task, not read a page.
The Governance Question Nobody Has Answered
Exposing a tool means an agent can now take an action on your site without a human clicking anything. Before you ship the first one, your team needs answers to three questions: What happens when an agent calls a tool with malformed input? Who gets notified when tool-driven bookings or leads spike or drop? And what is your rollback plan if an agent platform starts calling your pricing tool in a way that surfaces a price you did not intend to show?
None of that is exotic. It is the same production discipline this site has argued marketing needs for every other AI agent surface, applied to the one surface marketing actually controls end to end: its own website.
The Takeaway
WebMCP is live in ChatGPT's desktop browser today, and OpenAI is running a challenge with cash prizes through September 4 specifically to seed compatible sites. That is a short window where being early is genuinely differentiating, before "agent-ready" becomes the new "mobile-friendly," a baseline nobody notices until you fail it.
Pick three pages: your pricing page, your booking flow, and your highest-intent product finder. Define the tool schema for each with the same rigor you would apply to an API contract, because that is exactly what it is. Ship it behind a flag, watch the invocation logs for two weeks, then expand. The teams that treat this as a content project will publish a blog post about it. The teams that treat it as infrastructure will be the ones an agent actually recommends.
Tags
LETSGROW Dev Team
Marketing Technology Experts
Ready to Apply This Insight?
Schedule a strategy call to map these ideas to your architecture, data, and operating model.
Schedule Strategy Call
