Do Websites Need MCP to Be Ready for AI Agents?
Most websites do not. The Model Context Protocol (MCP) lets AI applications call a service's tools and data directly, which matters when agents need to do something inside your product: search a catalogue, check an account, book or buy. But when agents research, compare and cite, they mostly read ordinary web pages. In a study Vercel published in August 2026 with the ora.ai research lab, 82.5% of the agent answers that could be traced came from a page the agent fetched. For most sites, readable server-rendered pages and accessible buttons and forms come first. An API or MCP server is worth building when agents must act or query live data, and commerce sites get most of it through their platform.
What does an agent need from your site?
It depends on the task:
| Agent task | What it uses | What you need |
|---|---|---|
| Answer a question and cite you | Fetched pages, often converted to Markdown | Server-rendered HTML that answers the question; fast, reachable pages |
| Operate your site in a browser | Screenshots, HTML and the accessibility tree | Real buttons, labelled forms, stable layouts; optionally WebMCP |
| Act through your product | API calls or MCP tools | A documented API, an OpenAPI file, an MCP server, working authentication |
| Buy a product | Commerce protocols | UCP or ACP support, usually through your platform, plus accurate feeds |
Most business websites mainly need the first two rows. SaaS products, marketplaces and booking services also need the third.
How MCP, ARIA, and UCP make websites easier for AI agents to use
Each addresses a different layer:
- MCP is an open protocol for connecting AI applications to tools and data. A remote MCP server, often at a path like /mcp, lists tools such as “search_products” or “create_invoice” with descriptions and input schemas. An AI client that the user has connected to it can call those tools directly, instead of reading pages. Authentication uses OAuth metadata published under /.well-known, as covered in what goes in a site's .well-known directory for agents.
- ARIA and semantic HTML help agents that operate a site in a browser. Google's web.dev guidance on agent-friendly sites says agents combine screenshots, raw HTML and the accessibility tree, and recommends real button and link elements, labelled form fields, visible controls and stable layouts.
- UCP, Google's Universal Commerce Protocol, gives shopping agents a standard way to discover a store's capabilities at /.well-known/ucp and complete checkout. Shopify stores support it natively and publish an MCP endpoint for it, so most merchants get it without building anything.
- WebMCP, a proposed web standard from the Chrome team, sits between these. A page registers tools, such as “checkout” or “filter_results”, in the browser, so an agent calls the tool instead of clicking through the interface. Chrome offers it in an origin trial, and it is described as a progressive enhancement, not a requirement.
Can AI access structured data through an API or MCP server?
Yes, but only in specific situations. An AI assistant can call your API or MCP server when a user or developer has connected it, for example as a connector in Claude or an app in ChatGPT, or when a coding agent is told to use your API. In that case the agent gets live, structured answers, such as current stock, pricing or account data, rather than a snapshot from a page.
AI search does not work that way. The systems behind ChatGPT search, Google's AI Overviews and Perplexity do not query each site's MCP server for a normal question. So an MCP server does not make you more visible in AI answers on its own. How Claude decides to recommend an integration is covered in what an MCP server needs to say for Claude to recommend it.
Structured data on pages has limits too. Vercel's study did not measure any improvement in agent accuracy from JSON-LD, and its controlled tests found that adding llms.txt, JSON-LD, sitemaps or Markdown mirrors did not change whether agents could retrieve an answer already in the server-rendered HTML. Google's own AI optimisation guide says no special markup is required. Some tools bridge the two: Microsoft's open-source NLWeb turns a site's Schema.org data into a conversational interface, and every NLWeb instance is also an MCP server.
What did the agent study find?
Vercel and ora.ai observed 1,033 agent runs across 25 sites:
- Docs pages carried most answers. Agents reached documentation in 83% of runs.
- Markdown was requested often. When a fetch specified a format, it asked for Markdown 95.7% of the time.
- Sitemaps were rarely read directly, in 4% of runs.
- OpenAPI files were used. On seven sites that published one, agents fetched it in 21% of runs.
- .well-known paths were reached in 22.7% of runs.
Six sites accounted for two thirds of the runs, and most were developer products, so the results fit SaaS and documentation sites best. Vercel also launched is-agentic.com, a free checker that scores a domain against agent readiness tests.
Who should build an MCP server?
- Build one: SaaS products whose users want agents to act in them, such as creating records, running reports or managing settings; developer platforms; data providers; booking and marketplace services where live availability matters.
- Use your platform's: online stores on Shopify or another platform with UCP support; documentation sites on hosts that offer an MCP server.
- Skip it for now: brochure sites, local service businesses, blogs and most B2B marketing sites. Readable pages, accurate listings and clear forms do more.
The claim that sites without MCP will not exist to agents is not supported by how agents behave today. Most agent traffic still reads pages.
What should you do first?
- Check that agents can fetch your key pages. Request a page with an agent user agent, such as
curl -iL -A 'Claude-User/1.0', and confirm it returns the content in HTML. - Make buttons and forms accessible: real elements, labels and stable positions, as covered in how to make websites readable to AI agents.
- Publish an OpenAPI file and self-serve API keys if you have an API.
- Then consider an MCP server for the actions agents most need, and WebMCP if agents operate your site in the browser, such as for bookings, as covered in how to structure a site so agents can complete bookings.
What each AI provider publishes about optimisation is collected in whether AI search providers publish guidelines.
We have not measured the effect of adding an MCP server on AI visibility. This page is based on Vercel's published study, Google's web.dev and Chrome documentation, and published protocol documentation.
What might this look like in practice?
Consider an illustrative scheduling SaaS, Ruddby, whose founder heard that sites without MCP would be invisible to AI. Its team first checked that its docs and pricing pages rendered without JavaScript for agent user agents, and fixed two pages that did not. It published an OpenAPI file and added a remote MCP server with three tools: check availability, create a booking link and list upcoming meetings. Customers connected it to their AI assistants, and support saw fewer “how do I” tickets. The marketing site itself got no MCP server, and its visibility in AI answers kept coming from the docs and comparison pages.
What do people ask about MCP and websites?
Does an MCP server help my site appear in ChatGPT or AI Overviews?
Not directly. AI search answers come from pages the systems crawl or fetch, not from each site's MCP server.
What is the difference between MCP and WebMCP?
MCP connects AI apps to a server's tools. WebMCP lets a web page register tools in the browser for agents operating that page.
Do Shopify stores need to build MCP?
No. Shopify stores support UCP natively and publish an MCP endpoint for it automatically.
Is llms.txt a substitute for MCP?
No. llms.txt is a list of pages to read. MCP exposes actions and live data.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources