What happened

On 28 July 2026, the Model Context Protocol (MCP) shipped a new specification version — its largest revision since the protocol appeared in 2024.

Two things to remember. First, the core was rewritten to be stateless. The official specification lists “Stateless, self-contained requests” among the first properties of the base protocol. In plain terms: an AI agent used to establish a connection with an external service before it could start talking; now every request carries its own identity and context. That change serves one purpose — letting MCP run behind load balancers, across machines, across regions. This is groundwork for production scale.

Second, the new specification writes several extensions into the official documentation. One is MCP Apps, described by the spec as “interactive UI elements (charts, forms, video players) rendered inline within conversations.” The other two are Tasks (asynchronous execution of long-running operations, with polling, mid-flight input and durable handles) and Skills over MCP (structured workflow instructions an agent can read directly).

Governance moved too. MCP now sits under the Agentic AI Foundation (AAIF) at the Linux Foundation, whose platinum members include AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI. The specification did not fork and the maintainers did not change — what changed is the legal and governance wrapper around it.

Why it matters to you

Start with something you can verify yourself. Open any AI assistant and ask it which moving company in your city has good reviews. You get a paragraph and a few cited links. You click through, browse the site, fill in a form, wait for a call.

Along that path, your website is at least still on screen.

Here is the next step MCP Apps describes: the AI grows a comparison table inside the conversation, with a quote form beside it and a date picker you can drag. The user never leaves the chat window. And that form, that comparison table, was supplied by the handful of providers already wired into the agent.

What happens to everyone else? Your content may still get read — as the reference layer underneath. The screen the user is actually looking at and actually clicking belongs to someone else.

For two years GEO has been a fight over whether the answer mentions you. The next stretch, already written into the spec, is a fight over that screen.

An honest caveat: what the specification permits is not what ships tomorrow. No major AI product has opened this kind of interface to ordinary websites at scale, and nobody can hand you a date for when — or in what form — it becomes common. Anyone promising you a timeline is selling something.

But look at that platinum list again — AWS, Google, Microsoft, OpenAI, Anthropic, Cloudflare. Companies that fight each other everywhere else are sitting at one table pouring one foundation. That turns the question from whether into when does my industry come up.

Should you act now

Start with what to avoid: do not rush to wrap your company website into an MCP server. That is an engineering project, and a harder question sits in front of it — should your service be callable by an automated agent at all? A function that lets an agent place orders or book appointments is a very different proposition from a static description page.

What deserves attention is sequence. Before an agent slots you into a workflow, three things have to be true: it can find you, it knows what you do, and it trusts what you say. All three overlap exactly with the fundamentals you should be working on now — content a crawler can actually retrieve (client-rendered sites often hand back an empty div), company facts that agree across your site and off it, structured data that matches the content it describes.

Brands without that layer will not even qualify for the queue on the day these interfaces open up. Brands with it are already sorted and waiting to be picked up.

For the last few standards we covered, the verdict was mostly “worth knowing about, no rush.” This one sits differently. It does not ask you to start building today — it shrinks the number of months you have to take your time.

To see where you currently stand, book a demo. We will run a backtest first and show you how the mainstream AI engines describe you right now, and what crawlers actually retrieve from your site — once you can see it, we can talk about handing the ongoing watch to us. This layer shifts week to week; you are welcome to watch it yourself, though doing it by hand gets exhausting fast.

Further reading: MCP and A2A: the two pipes under the agentic web, The GEO brand knowledge system