Cituna
GEO

How to query Google Search Console from Claude

Search Console holds the answer to most SEO questions and buries it three menus deep. Connecting it to Claude with an MCP server turns it into something you can just ask, and opens one question no other setup can answer.

By Rahul AUpdated August 3, 20268 min read

See which of these you are already failing.

On this page
  1. Why put Search Console in the chat
  2. What MCP actually is
  3. Setup, in three steps
  4. Questions worth asking
  5. The join a GSC connector cannot do
  6. Limits worth knowing
  7. FAQ

Google Search Console knows which queries brought you clicks yesterday, which pages are slipping, and where you sit on page two waiting for a nudge. It is the most useful free dataset in SEO, and using it is a chore: every real question means a fresh filter, a date-range change and a squint at a chart that shows the shape but not the answer.

Putting it behind Claude changes the ergonomics completely. You ask the question you actually had, in the words you had it in, and get the numbers back with the reasoning attached. This guide covers the setup, the questions worth asking, and the one analysis this arrangement can do that a Search Console dashboard cannot.

Why put Search Console in the chat

Because the interface is the bottleneck, not the data. Most SEO questions are comparative (what changed, against what, and where), and Search Console answers them one filter at a time. You know what you want to know. Getting there costs six clicks and a mental note about which date range you left selected.

In a chat window the question and the analysis collapse into one step. Ask which queries lost ground this month and you get the list, the deltas, and a read on whether it looks like a ranking drop or a seasonal dip, because the thing fetching the numbers is also the thing interpreting them. You can follow up in the same breath: now show me the pages those queries land on. That follow-up is where the time actually goes in a dashboard, and it is free here.

What MCP actually is

The Model Context Protocol is an open standard for giving an AI assistant access to a specific system: a database, an API, a set of files. An MCP server is a small program that exposes that system as a set of tools the model can call, with names and inputs it can reason about. Claude does not get blanket access to anything. It gets exactly the tools the server declares, and nothing else.

That matters for Search Console. Rather than pasting exports into a chat and hoping the model reads them correctly, the model queries the real API and gets structured rows back. The numbers it reports are the numbers Google returned, not a summary of a screenshot.

Setup, in three steps

The whole thing takes a few minutes, and you never handle a Google OAuth flow yourself.

1. Connect Search Console to Cituna. In the app, open Integrations and connect Google Search Console. You authorise once, against your own Google account, and the backend keeps the refresh token encrypted at rest. This is the only time Google is involved.

2. Add the connector. In Claude, open Settings, then Connectors, then Add custom connector, and paste:

https://mcp.cituna.com/mcp

3. Approve access. Claude sends you to Cituna to sign in and shows you what it is asking for. Approve, and you are connected. Nothing is installed, there is no config file, and no key changes hands. That is the whole setup, and because the server is hosted it works in Claude on the web and on your phone as well as in Desktop and Code.

If you would rather run the server on your own machine, for a script or CI job, it is also published on npm. Generate a token from Integrations, then in Claude Code:

claude mcp add cituna --scope user \
  --env CITUNA_API_KEY=cituna_sk_your_key \
  -- npx -y cituna-mcp

Or in Claude Desktop, open Settings, Developer, Edit Config, and add the server to claude_desktop_config.json:

{
  "mcpServers": {
    "cituna": {
      "command": "npx",
      "args": ["-y", "cituna-mcp"],
      "env": { "CITUNA_API_KEY": "cituna_sk_your_key" }
    }
  }
}

Restart Claude Desktop and the tools appear. Ask is Search Console connected? to confirm the link before you rely on any numbers.

Questions worth asking

Once it is wired up, the useful questions are the ones that were tedious before, not the ones a dashboard already answers well.

  • Movement with a cause. “Which queries lost the most clicks versus the previous 28 days, and did their average position move or just their impressions?” That distinction separates a ranking loss from a demand dip, and it is two different reports in the Search Console UI.
  • Striking distance. “Which queries sit between position 8 and 20 with over 100 impressions?” This is the standard prioritisation exercise, and it is a filter combination nobody remembers.
  • Page-level truth. “For this URL, which queries does it actually rank for, versus the one I wrote it to target?” The gap between intent and reality is where most content fixes start.
  • Splits that expose a problem. “Break traffic down by device, then by country, and tell me if anything looks off.” Mobile collapsing while desktop holds is a technical problem wearing a content problem’s clothes.

The pattern: ask the comparative question directly instead of assembling it. The model picks the dimensions, runs the query and reads the result back with the caveats attached.

The join a GSC connector cannot do

Everything above is useful and none of it is unique. Several tools can put Search Console numbers in front of a model. The reason to run this particular server is the second dataset sitting beside them.

Cituna tracks how six AI answer engines (ChatGPT, Perplexity, Gemini, Claude, Grok and Google AI Overviews) answer your buyer questions, and keeps what each one actually said, including which brands it cited and which sources it leaned on. That lives in the same MCP server as your Search Console data. So you can ask a question neither dataset can answer alone:

The question that needs both

“Which queries does Google already send me impressions for, where the AI engines never cite me?”

That list is the highest-value target in AI search work, because it separates two problems that look identical from the outside and need opposite fixes. If Google sends you nothing and AI cites you nowhere, your constraint is indexation and coverage: write and get crawled. If Google already ranks you and AI still cites someone else, coverage is not the problem. The engines are building their answers out of third-party roundups, review sites and comparison posts you are absent from, and no amount of on-page work moves that. The fix is earning a place in those sources.

Answering that with a Search Console connector alone is impossible: it has half the evidence. Answering it by hand means running your buyer questions through six engines, reading every answer, and reconciling the result with an export. That is the work the join removes. There is more on the shape of that analysis in our guide to AI visibility tools versus SEO tools.

Limits worth knowing

Three honest caveats, because finding them out mid-analysis is worse than reading them now.

Search Console lags about two days and its windows are UTC. Ask about yesterday and you will get a thin or empty answer. That is Google’s reporting delay, not a broken connection.

Read and write are gated differently. Every Search Console tool, and every tool that reads your existing visibility data, works on Starter at $39/mo. The write tools (running a fresh scan, changing your action queue, queueing an article) need Pro. The 3-day trial is app-only: it has no MCP access at all.

Zero is a real answer. On a new domain the honest result is often an empty table, and the server is built to report that plainly rather than dress it up. If your totals come back at zero, that is the finding: the constraint is coverage, and no amount of re-querying changes it.

Where this leaves you

Search Console in a chat window is a genuine ergonomic win on its own: the questions you actually have, answered in the words you had them in. The reason to care beyond convenience is the pairing. Once the same conversation holds what Google sees and what the AI engines say, you stop guessing about which of the two problems you have, and the answer comes with the receipts attached.

The server is open source, hosted at mcp.cituna.com and also installable with npx. The Cituna MCP server page covers the full tool list, and plans start at $39/mo with all six engines tracked daily on every tier.

Frequently asked questions

Can Claude access Google Search Console?

Not on its own. Claude has no connection to your Search Console account until you give it one, and the standard way to do that is an MCP server: a small program that exposes your data to Claude as tools it can call. Install the Cituna MCP server, authenticate once with an API key, and Claude can read your clicks, impressions, CTR and average position, break them down by query, page, country or device, and compare date ranges, all in plain English.

Is there a Google Search Console MCP server?

Yes. Cituna publishes one on npm as cituna-mcp. It exposes Search Console through two tools: gsc_overview for the headline summary over a trailing window, and gsc_query for arbitrary Search Analytics queries with your own dimensions, filters and date ranges. The same server also carries your AI-visibility data, so one connection covers both what Google sees and what the AI engines say.

Do I need to be technical to set this up?

No. For Claude Desktop it is a short block of JSON pasted into a settings file, and for Claude Code it is a single command. You never touch Google OAuth yourself: you connect Search Console once inside Cituna, and the server calls the backend, which holds the refresh token. Nothing sensitive lives on your machine beyond one revocable API key.

Is my Search Console data safe?

The MCP server is a thin bridge. It holds no database and no Google credentials: it authenticates as you with an API key you generate in the app and can revoke at any time, and the backend keeps the Google refresh token encrypted at rest. Requests go from your machine to the Cituna backend over HTTPS. Revoking the key in Integrations cuts access immediately.

How much does it cost?

The MCP server is included from the Starter plan at $39/mo. Read tools, which is everything covered in this guide, work on Starter. The write tools, which run a fresh scan or change your action queue, need Pro. The 3-day trial is app-only and has no MCP access.

See how AI engines see your brand

Start a free 3-day trial and see the exact buyer prompts you lose across ChatGPT, Perplexity, Gemini, Claude, Grok and Google AI Overviews, with a prioritized AEO, GEO and SEO action plan and the fixes to win them.

3-day free trial · Card required, cancel anytime · Works with ChatGPT, Perplexity, Gemini, Claude, Grok and Google AI Overviews

Start free trial