What is an MCP? Two ways it affects search marketers

An MCP is a connector that lets an AI assistant, such as ChatGPT, use an outside source of data or a tool. MCP stands for Model Context Protocol, the open standard these connectors follow. An MCP server is the piece a company runs to make its data or tools available that way: it controls what the AI can see and do.

It's one of those terms I keep hearing dropped into meetings, and it often means two different things. Sometimes people mean making a store readable and buyable for AI shopping agents.

Sometimes they mean connecting a tool like Pi to ChatGPT so their team can ask it questions. Add UCP and ACP to the mix and you can see why there's confusion.

Both use the same protocol, but they're different jobs with different owners and next steps. This post separates them.

What is an MCP server?

I describe MCP as the USB for AI: one standard connector instead of a custom cable for every device. A company that runs an MCP server decides what an AI can do through it, such as read product specifications, check stock or pull a report.

Anthropic open sourced MCP in November 2024 and donated it to the Linux Foundation's Agentic AI Foundation in December 2025. ChatGPT, Claude, Gemini and Microsoft Copilot all support it.

How is that different from an API? An API gives developers a structured way to request specific data. An MCP server makes data and actions understandable to an AI, so it can work out what to use while it answers a question in plain language.

In this short glossary video, we explain what MCP is and how it lets an AI assistant connect securely to your data.

Use 1: the store side (being found and bought by AI agents)

The first use is about your customers' agents, not your team.

AI shopping is moving from answering questions to acting on them. Brands are setting up business agents, conversational assistants that represent the brand inside Google Search and Gemini.

Shoppers are starting to use personal agents that research, compare and, increasingly, buy for them. For any of this to work, the agent needs reliable facts about your products: specifications, pricing, availability, delivery and returns.

In the past, an agent would have had to scrape your site to find them. MCP lets you present that information in a secure, structured way instead. As I put it in one of our shorts, the business agent gets it "through an MCP layer. It's like the USB of AI, a connector that makes your business agents really super smart."

In this short video, I explain both uses of MCP in under a minute, starting with the product information business agents need.

MCP isn't a commerce protocol on its own. It sits alongside two that handle the shopping itself:

The simplest way I can split it: UCP and ACP handle the business logic and the transaction, and MCP handles the information. That's how a business agent can pull detailed product information mid-conversation, or check whether you're already a customer before offering a discount.

Watch me and Richard Barrett unpack MCP, UCP and ACP in our July webinar (from 31:36).

What to do about it

1. Get your product data complete and consistent. Feeds, Merchant Center attributes, schema, prices, stock and policies. If that feels like a lot, start with your top 50 products.
2. Build one source of truth for the facts agents ask about. Policies, FAQs, sizing, delivery and returns. Whatever you expose through MCP is only as good as how that data is structured.
3. Find out where your data already goes. Check what reaches Google through Merchant Center and what reaches ChatGPT through ACP feeds or a provider such as Shopify, before you build anything new.
4. Ask whether you need your own MCP server. That's a conversation with your development team or commerce platform, not a job for SEO alone.
5. Measure whether agents recommend you. Our AI Shopping Tracker shows whether your business is suggested as the seller, not just whether your product appears.

If you sell in the UK, most of the checkout features above are US first, so the immediate job is data readiness rather than plugging in a buy button. For the full picture, including business agents and where each platform stands, read What is agentic commerce?

Use 2: the team side (getting work done with your own data)

The second use has nothing to do with shoppers. Here MCP connects an AI assistant your team already uses, such as ChatGPT, to the tools and data you work with. You ask for analysis in plain language instead of opening reports one by one.

This is the use I rely on most day to day. One morning on the train I had a question I'd been trying to solve about some complex data. I connected ChatGPT to Pi through our MCP and simply asked it.

As I said in the short I recorded afterwards, "I had a conversation with the data." Within minutes I'd built a dashboard and seen the data in ways I hadn't seen before.

In this short video, I explain how I connected ChatGPT to Pi through our MCP and had a conversation with the data.

That's what our MCP Server does. It connects compatible AI assistants to the search intelligence in your Pi account, so you can ask about your brand, competitors and markets across AI and traditional search. For example, you can ask:

You can follow up, dig into a finding and ask for an output such as a table, chart, dashboard or presentation, depending on what your assistant supports. The assistant only works with the data available to you in Pi. You don't need to be a developer: our help article shows how to connect to the Pi MCP, and our MCP Template Library gives you pre-built workflows to start from.

It gets more useful when you connect more than one tool. In our July webinar, Richard Barrett of Lumar explained how he can take a Lumar crawl and ask what Pi data exists for those pages in the same conversation, without opening a new window.

This isn't the same as our GPTs, such as Forecasting Demand or the 12-Month Content Planner. Those open in ChatGPT as ready-made assistants for a specific task. MCP is the connection that lets your assistant reach your Pi data for whatever question you bring.

Which one applies to you?

If you sell online, you probably need both, but treat them as separate projects. One makes your brand legible to other people's agents. The other makes your own data easier to work with.

Want to see where this is heading? Join me and Peter Longton on December 2, 2026, at 3pm GMT for "Google's agentic Search future: will your business make the shortlist?"