Connect Claude to your SEO data

You ask Claude where your site ranks for "emergency plumber Leeds" and get a careful paragraph about how rankings vary and how you should really check Search Console. It isn't being evasive. An assistant has no window onto Google's results unless you give it one, and the MCP connector is that window: a small set of live lookups Claude can run mid-conversation, so its answers rest on measurements rather than on pattern-matching.

Why an assistant needs a tool, not a better prompt

A language model produces text from two things: what it learned during training, and whatever is in the conversation in front of it. Neither includes today's results page for your keyword in your market. Ask it your position and it has nothing to measure, so it reasons from general patterns and hedges, or worse, produces a number in a confident tone that it never saw anywhere.

MCP, the Model Context Protocol, is an open standard that lets an assistant call outside tools. When a client such as Claude connects to an MCP server, it asks the server what tools it offers, keeps their descriptions in mind, and calls one whenever a question needs it. The reply comes back as data, and the assistant writes the sentence around the data.

That division of labour is the whole point. The measurement comes from a tool that reads the live results page; the explanation comes from the assistant. You get a number you can check, with the reasoning beside it.

The three lookups the connector gives Claude

SearchBlueprint's MCP server exposes three read-only tools. The connector page lists them, and your assistant sees the same list the moment it connects.

1. rank_check. Takes a domain, a keyword and a country, and returns the domain's live Google organic position and the URL that holds it. A position of zero means the domain isn't in the top 100 results, which is the range the tool's own description says it checks. This is a snapshot for one market at one moment, not an average.
2. page_meta. Takes a URL and returns the page's title, meta description, H1 and headings. It's the quickest way to see what a page actually says about itself, which is often not what you remember writing.
3. ai_presence. Takes a domain and a keyword and reports whether Google shows an AI Overview or an AI Mode answer for that query, and whether your domain is among the sources it cites. A "no" here is a real measurement, because many transactional and local queries trigger no AI surface at all.

Just as important is what the connector doesn't do. The server exposes lookups only: running audits, generating content and anything else that spends real credit stays inside the product. The endpoint is public, no account secrets pass through it, and it can't read your workspace data, so anything Claude knows about your site is what you paste in or what a lookup returns. Tool results match what the product's own pages would show for the same question, and the same rate limits apply, so a conversation that fires off lookups by the dozen will be told to slow down.

Connect it in three steps

1. Open the MCP connector page for the connection details. The server speaks JSON-RPC over HTTP at the /api/mcp endpoint, and there is no SDK to install.
2. Add that URL to your assistant's MCP configuration as a remote server. Where the setting lives depends on the client, so follow your assistant's own documentation for adding a remote MCP server.
3. Ask a question that needs one of the tools. The assistant discovers them automatically on connection and you don't have to name them; "where does my site rank for X in the UK" is enough.

If the assistant says the server won't connect, check two things. The URL must be the /api/mcp endpoint exactly, not the product's homepage. And your client must support remote MCP servers; some only support servers running on your own machine.

Read the answers like an analyst

Tool responses are data and the assistant phrases them. Most of the time the phrasing is faithful, but an assistant can round, summarise or misread, so when an answer smells wrong, ask it to show the raw tool result before you act on it.

Treat a rank check as a single reading, in one market, at one moment. Position three in Singapore and not found in the United States are both true at once, and the country you pass decides which one you see. Google also personalises the results in your own browser by location and history, so a neutral lookup that disagrees with what you saw this morning isn't necessarily wrong.

Some lookups are cached briefly so that repeated questions stay cheap. If you've just changed a title and want to confirm the new one is live, ask again with a different phrasing, or wait a few minutes before re-checking.

Finally, keep the assistant honest about what it measured. "Your site ranks ninth for this keyword in Australia right now" is a measurement. "Your rankings are improving" is a claim no single lookup can support, and the next lesson builds the weekly routine that can.

What to take away

• An assistant can't measure Google's results on its own; the MCP connector gives Claude live lookups so its answers rest on data you can check.
• The connector offers three read-only tools: a live rank check, a page's metadata and headings, and whether Google's AI surfaces cite your domain for a query.
• Connecting takes the /api/mcp URL and your assistant's remote-server setting; the tools are discovered automatically.
• Every lookup is a snapshot for one market, so ask for the raw tool result whenever a phrased answer looks off.

Next

With the connector working, put it on a schedule in A weekly SEO review with Claude.