GEO · AI agents

WebMCP: the new entry point for AI agents on your website?

Hubert de Cartier·26 JUNE 2026·11 MIN

For years, brands have optimised their websites for two audiences: the internet user and Google. The first needed to understand, click, buy or request a quote. The second needed to crawl, index and rank. Then came a third challenge, GEO (Generative Engine Optimisation), with a question that has become central for marketing leaders: is my brand cited by ChatGPT, Claude, Gemini or Google’s generative answers?

A new stage is already taking shape. Tomorrow, a customer may no longer look for your contact page. They may ask their AI assistant to compare three Belgian providers, check who delivers in Brussels or Liège, read the returns policy, then prepare a quote request. In this journey, your homepage is no longer necessarily the entry point: your website becomes an infrastructure the agent needs to understand.

This changes marketers’ priorities, shifting from “generating traffic to my website” to “being present where my customers and their agents search”. Simple on paper, more complex in practice.

This is where the WebMCP protocol comes into play. With a strong promise, but one that needs to be viewed without naivety.

WebMCP, a user guide for websites for AI agents

To understand WebMCP, we need to go back to MCP (Model Context Protocol). Introduced by Anthropic in 2024, MCP aims to create an open standard for connecting AI assistants to the systems where data lives. The goal: to move away from a world of fragmented integrations, where each tool has to be connected separately to each AI. In less than two years, MCP integrations have become very popular with users of Claude, ChatGPT or Copilot, making it very simple to connect Google Analytics, HubSpot, Google Ads, Google Search Console, Slack, Notion, Zapier/Make/n8n, as well as Belgian tools such as Leexi, Odoo, TeamLeader or Showpad.

Definition

WebMCP, Web Model Context Protocol

According to the Chrome documentation, it is a proposed standard allowing websites to expose structured “tools” to AI agents, through JavaScript and the annotation of HTML elements, so that these agents know precisely how to interact with the features of a page.

Put more simply: today, an AI agent often has to guess what a website allows it to do. It observes an interface, interprets a button, identifies a field, clicks, waits for a result, then starts again. This approach sometimes works, but it remains fragile: a design change, an A/B test, a poorly labelled form or a pop-up can be enough to make the task fail.

WebMCP proposes another path: the website explicitly declares what it can do. Search for a product. Filter an offer. Check availability. Submit a quote request. Add an item to the basket. Book an appointment. Track an order. Open a ticket.

In marketing terms, this is a major shift: the website no longer only says what it is, or even what it is about. It says what it enables people to do, in a structured way that agents can understand.

The project is still young, however. The WebMCP specification is published within the W3C Web Machine Learning Community Group, following a joint initiative by Google and Microsoft. But it is not yet a mature W3C standard, nor a stable part of the web.

From GEO to agentic: being cited will no longer be enough

The parallel with GEO is obvious. Since the arrival of generative engines, brands have been learning to ask whether they are included, mentioned, recommended, or forgotten in AI answers. That is already a significant change: a brand’s authority is no longer built only on its own website, but also through the consistency of its mentions across the web, press, reviews, comparison sites, forums, knowledge bases and third-party content.

But agentic adds another layer. Being cited in an answer is one thing; being selected by an agent to carry out an action is another.

The same logic applies to a retailer, a mutual insurance provider, a bank, an estate agency, a tourism brand or a B2B company. In Belgium, the complexity is even greater: several languages, several regions, networks of points of sale, local stock, and conditions that sometimes differ depending on the channel or market. For an AI agent, this complexity has to be readable. Otherwise, it could look elsewhere.

There is also a parallel with Core Web Vitals. In recent years, SEO, UX and tech teams have learnt to measure experience quality through speed, visual stability and interactivity. With agentic, a new form of experience emerges: readability and actionability for agents. It is already integrated into Google’s Lighthouse tool, through an “Agentic Browsing” audit logic that notably recommends working on WebMCP, the representation of page structure (accessibility tree), semantic HTML, page stability and the possible presence of an llms.txt.

The subject becomes even more strategic with e-commerce. OpenAI and Stripe have launched the Agentic Commerce Protocol to allow agents, users and merchants to collaborate around a purchase. Google, for its part, is promoting the Universal Commerce Protocol, designed as an open standard to make agentic commerce journeys smoother. Visa has announced a collaboration with OpenAI to integrate its payment network into these experiences, while Mastercard emphasises the need for standards that guarantee user intent, credential security and the verifiable identity of agents.

The direction is clear: agents will not only be recommendation engines. They will gradually become intermediaries for decisions, and then for actions.

Starting now, without believing in a magic wand

Let us be clear: WebMCP will not save a bad website. It will not compensate for an unclear offer, inconsistent prices, a poorly maintained catalogue, a weak content strategy or a lack of brand awareness. Nor does it guarantee that an AI agent will recommend your brand. And at this stage, it is not a new magic ranking factor. But preparing for it is essential, because the subject will evolve quickly.

The first task is not to “do WebMCP” in order to tick an innovation box. Above all, putting yourself in the user’s shoes, and thinking about what they should be able to achieve on your website, makes it possible to consider concretely what agents should be able to accomplish. For an SME, this can start with three simple things: understanding the offer, finding the right contact page, submitting a quote request. For an e-commerce business, it will instead mean searching for a product, checking a price, confirming availability or consulting the returns policy.

Then, you need to return to the fundamentals. An “agent-ready” site is often first and foremost a good website: clean HTML, properly labelled forms, explicit buttons, stable navigation, hierarchical content, up-to-date product data, coherent multilingual pages, and contact and delivery information that is easy to interpret. Google also reminds us that making a website more agent-friendly often comes down to the basic principles of a structured, accessible and semantic web, the very fundamentals of good technical SEO.

Where to start, in practical terms

1. 1

   Test the “Agentic Browsing” audits

   The new audits integrated into Lighthouse assess the readability and actionability of your pages for agents.

2. 2

   Create an llms.txt

   Not as a talisman, but as a readable summary of the site’s key content. Its impact on GEO remains very limited according to a recent Ahrefs study, but its potential for agentic is real.

3. 3

   Document a simple action

   Describe precisely a task that agents should be able to accomplish on your website.

4. 4

   Experiment with WebMCP

   On a search form, an appointment booking flow or a quote request.

5. 5

   Measure, correct, expand

   Iterate gradually based on the first results observed.

SEO made websites findable. GEO makes them citable. WebMCP could make them actionable.

For Belgian brands, the topic deserves to be integrated into digital strategic thinking from now on, without hype, but also without waiting passively. Because if consumers tomorrow delegate part of their searches, comparisons, purchases and bookings to their favourite assistant, the customer journey may no longer begin on your homepage, or even on Google. It will begin in a conversation, within agents that transform “traditional” customer journeys.

And in that conversation, your brand will need to be more than a name. It will need to be understandable, reliable and usable.

Frequently asked questions

What is WebMCP?+

WebMCP (Web Model Context Protocol) is a proposed standard that allows a website to declare to AI agents the actions it can perform, such as searching for a product, checking availability or submitting a quote request, through JavaScript and the annotation of HTML elements. Instead of guessing the interface, the agent has a structured “user guide” for the page.

What is the difference between SEO, GEO and WebMCP?+

SEO makes a website findable in search engines. GEO (Generative Engine Optimisation) aims to make it citable in generative AI answers. WebMCP could make it actionable: allowing AI agents to carry out concrete tasks on it.

Is WebMCP already an official standard?+

Not yet. The specification is published within the W3C Web Machine Learning Community Group, following a joint initiative by Google and Microsoft, but it is neither a mature W3C standard nor a ranking factor. The project remains young, even though it is evolving quickly.

How should you prepare your website for the age of AI agents?+

By first returning to the fundamentals of a good website: clean HTML, properly labelled forms, stable navigation, up-to-date data and coherent multilingual pages. Only then come the specific layers: “Agentic Browsing” audits, an llms.txt file and experimentation with WebMCP on simple, low-risk actions.

What are the risks of WebMCP?+

The specification identifies several risks: prompt injection, ambiguity about the true intent of a tool, data leakage through over-parameterisation, or the triggering of sensitive actions in a session that is already authenticated. This is why human confirmation for sensitive operations, traceability of interactions and strict GDPR logic are essential.