How to Connect Your SEO Data to Claude

Pasting a keyword export into a chat window works, but it is slow, it truncates, and the assistant only knows what fits in the message. Connecting your data properly means Claude fetches what it needs, when it needs it — and can go back for more without you doing anything.

This guide covers what you can connect, how to do it, and how to check it is actually working. Most of it takes under five minutes. The free options alone will change how you work.

(If you came here for the Opinly MCP specifically, the launch announcement covers what shipped and why. This page is the practical setup guide, and it applies to any SEO data source, ours included.)

How the connection works

Claude connects to external data through the Model Context Protocol. A service runs an MCP server exposing its capabilities as tools; you connect Claude to that server; Claude can then call those tools mid-conversation.

Two flavours, and the difference matters for setup:

Remote servers are a URL. You paste an endpoint, sign in through OAuth, and you are done. Works everywhere you use Claude, including mobile.

Local servers run on your own machine as a process Claude launches. More setup — a runtime, a config file, credentials — and they only work on that machine.

What you can connect

Your own data, free

Google Search Console. The single highest-value connection, and free. Real impressions, clicks, positions and indexation status for your actual site. Google does not publish an official server for it — the most widely used community one is mcp-gsc, covering search analytics, URL inspection and sitemaps in around twenty tools. Local setup. (New to Search Console itself? Start with our beginner's guide.)

Google Analytics 4. Google ships an official, open-source MCP server here — released in May 2026 and actively maintained, which is the opposite of the Search Console situation. Search Console tells you how people arrived; GA4 tells you what they did next. Local setup, free.

If you connect nothing else, connect these two. They are ground truth, they cost nothing, and no paid tool can generate them for you.

Competitor and market data, paid

Your own data cannot tell you what competitors rank for — you would need their Search Console. That is what third-party indexes are for. Ahrefs, Semrush, SE Ranking and Serpstat all ship remote MCP servers on paid plans, and DataForSEO offers a pay-as-you-go raw data layer.

These numbers are modelled, not measured. Search volumes are estimates and traffic figures are inferences. Excellent for direction and comparison; do not treat them as counts.

Servers that can also act

Most SEO MCP servers only read. A few also write, so the assistant can change something rather than just report on it — drafting a post, scheduling it, tracking a keyword. Frase and Opinly both work this way. Useful when you want the analysis and the follow-through in one conversation; unnecessary if you only want answers.

Connecting a remote server

The whole flow, using the Opinly MCP as the example — any remote server follows the same steps:

Open Claude and go to Settings → Connectors.
Choose Add custom connector.
Paste the server URL. For Opinly that is https://mcp.opinly.ai/mcp.
Save. Claude opens an OAuth window.
Sign in with your account for that service and approve access.
Start a new conversation. The tools are available.

In Claude Code it is one line:

claude mcp add opinly --transport http https://mcp.opinly.ai/mcp

No API keys to generate or store. OAuth means the connector gets scoped access you can revoke from the service, and your credentials never sit in a config file.

Connecting a local server

Google's servers are local, so they need more work. The shape of it:

Install the runtime. Most Google MCP servers are Python. You will need Python 3.10+ and uv or pip.
Create credentials. In Google Cloud Console, create a service account, enable the Search Console or Analytics API, and download the JSON key.
Grant access. Add the service account email as a user on the Search Console property or GA4 account. This step is the one people miss — the credentials exist but have no permissions.
Configure Claude Desktop. Edit claude_desktop_config.json to register the server command, arguments and the path to your key file.
Restart Claude Desktop. Config is read at launch; it will not hot-reload.

Budget an afternoon the first time if you have not made a service account before. It is a one-off.

Verifying it worked

Do not assume a green connector means good data.

Ask what tools it has. "What Opinly tools do you have access to?" You will see the real surface, which is often broader than the marketing page suggests, and you will spot immediately if a server is thinner than advertised.

Ask something you already know. Check a figure against the platform's own interface. If Claude reports a different number, find out why before trusting it for anything that matters — you may be looking at a different date range, market or property.

Check which source it used. With several servers connected, Claude picks the tool it thinks fits. If you have both first-party and estimated data available, say which you want: "using Search Console data, not estimates."

Common problems

Tools do not appear. Start a new conversation — existing ones do not pick up newly added connectors. For local servers, restart Claude Desktop entirely.

OAuth loops or fails. Usually a browser blocking the popup, or being signed into a different account than you expect. Try in a normal window, signed out of other accounts.

Local server never starts. Almost always an absolute-path problem in the config, or the runtime not being on the PATH Claude sees. Claude Desktop's logs will name the failure.

Permission errors after successful auth. For Google servers, the service account probably has not been added to the property. Credentials and access are separate steps.

Responses are slow or expensive. SEO data is verbose. A backlink profile can be thousands of rows, and every row entering the context is billed. Ask for filtered, limited results — "the top 20 by traffic" rather than "the keywords" — and prefer servers that summarise rather than dump. Some servers also cap rows or time out long queries without saying so, so ask whether a large result came back complete when the answer matters.

What to ask first

Questions that show what the connection is worth:

"Which queries am I getting impressions for but almost no clicks?" — surfaces title and snippet problems, from Search Console.
"Which pages lost the most clicks in the last 28 days versus the previous 28?" — decay detection, hard to see in the interface.
"Which keywords do my competitors rank for that I do not?" — needs third-party data.
"Summarise my site audit and list the top technical issues by impact."
"Which sites link to my competitors but not to me?"

The pattern that pays off is asking for a judgement rather than a list. "Give me my keywords" returns a table you then have to read. "Which three keywords should I prioritise this month, and why?" makes the assistant do the analysis. For a fuller set of these, see seven SEO workflows you can run from ChatGPT — they work identically in Claude.

Frequently asked questions

Do I need a paid Claude plan? Custom connectors are available on paid plans. Free-plan availability changes, so check current terms.

Can I connect several servers at once? Yes, and it is the normal setup — first-party data plus a competitor index is the common pairing. The practical limit is context: many servers with hundreds of tools between them makes tool selection less reliable.

Is my data safe? Remote servers use OAuth, so the connector gets scoped access you can revoke at the source, and no keys are stored locally. Read what each tool can do before approving — well-built servers declare whether a tool reads, writes or deletes, and Claude will prompt before anything changes.

Does this work in ChatGPT and Cursor? Any client supporting remote MCP servers with OAuth connects the same way, with the same endpoint. Local servers depend on the client supporting local processes.

Can Claude change things, or only read? Depends on the server. Most SEO servers are read-only by design. Some also expose write tools, in which case each tool declares its behaviour so your client can ask for confirmation first.

Start with the free ones

The instinct is to connect the most expensive tool first. Do the opposite. Search Console and GA4 cost nothing, describe your actual site rather than a model of it, and will answer most of the questions you have this week.

Add paid data when your questions turn to competitors — that is the point where you genuinely cannot get the answer any other way. When you reach it, our comparison of SEO MCP servers covers how to choose between them.