Connect Google Search Console to Claude

If you manage SEO for a site, you've probably opened Search Console, squinted at a table, and wished you could just ask it a question. Now you can — search-console-mcp-server is an open-source MCP server that connects your real Search Console data straight into Claude, so you can ask things like "which pages lost the most clicks this month" and get an answer pulled from your actual numbers, not a guess.

This guide covers the full setup, in order, plus the first few things worth asking once it's connected.

Why connect Search Console to Claude at all

Search Console has the data. It doesn't have the reasoning. You can see that a page's clicks dropped 40% in the UI, but you still have to cross-reference query changes, check if a competitor started ranking, and figure out whether it's a content problem or an indexing problem — by hand, across multiple tabs.

An MCP server hands that same data to an LLM that can already reason about SEO. Claude can pull the real numbers, compare periods, cross-reference queries against pages, and tell you what's actually going on — with your data, not a generic answer.

The catch with most "AI + Search Console" tools is the setup: create a Google Cloud project, enable the Search Console API, generate OAuth credentials, download a JSON key, wire it into some config. That's 20+ minutes before you've asked a single question. search-console-mcp-server skips all of it.

The 30-second install

You need Node.js installed (most developers already have it — check with node -v). That's the only prerequisite.

Step 1: Sign in.

npx search-console-mcp-server login

Your browser opens, you pick your Google account, you approve read-only access to Search Console. That's it — no Google Cloud console, no API key, no service-account JSON file. Tokens are issued by Google straight to your machine and saved to ~/.search-console-mcp/.

Step 2: Add it to Claude.

If you're using Claude Code:

claude mcp add gsc -- npx -y search-console-mcp-server

If you're using Claude Desktop, open claude_desktop_config.json (on macOS: ~/Library/Application Support/Claude/claude_desktop_config.json; on Windows: %APPDATA%\Claude\claude_desktop_config.json) and add:

{
  "mcpServers": {
    "gsc": {
      "command": "npx",
      "args": ["-y", "search-console-mcp-server"]
    }
  }
}

Restart Claude Desktop. That's the whole setup — from an empty terminal to a working connection in well under a minute.

Using Cursor, Windsurf, or VS Code instead? The config file location and JSON key name differ slightly per client — see the full per-client setup guides for exact paths.

The first 3 prompts to run

Don't start by exploring. Start with these three — they're the fastest way to see whether anything on your site needs attention right now.

1. "Run a full SEO checkup and give me the top 3 moves."

This calls the built-in /gsc:seo_checkup analysis: overall trend, your biggest movers (up and down), and a ranked list of what's worth fixing first. It's the right first question because it tells you where to look before you go looking.

2. "Which queries sit at position 5–15 with real impressions? What's the fix for each page?"

This is "striking distance" — queries close enough to page one that a small on-page change can push them over. These are usually the fastest wins available, and most people never check for them because it means manually filtering a big table by position and impressions. Claude does the filtering and tells you what changed for each candidate page.

3. "Compare this month vs last month, by page. What fell, and why?"

Instead of eyeballing two exports side by side, Claude computes the deltas directly and can cross-reference the drop against query-level changes — did a specific query lose position, did impressions drop (an indexing or seasonality signal), or did CTR drop on the same position (a title/snippet problem)?

What the output actually looks like

You're not getting a chart. You're getting Claude's read on the numbers, in plain language, referencing your actual pages and queries — e.g., "/blog/pricing-guide lost 340 clicks month over month. Impressions held steady, so this isn't an indexing issue — CTR dropped from 4.2% to 2.1% on the same average position, which points to the title or meta description. The query pricing calculator moved from position 4 to position 9 over the same window; a competitor page appears to have overtaken it."

That's a diagnosis, not a dashboard. You still make the actual edit — the MCP reads Search Console, it doesn't touch your site.

Limits, honestly

This connects Search Console — not Google Analytics, not your CMS, not your backlink profile. It sees clicks, impressions, position, CTR, URL inspection status, and sitemaps: exactly what Search Console itself exposes, no more. Read-only means it also can't submit URLs for indexing or fix anything for you; every fix ships as advice you paste into your CMS yourself.

It's also a one-time ask-and-answer tool. It only has the data it can pull in the moment you ask, and it only checks anything when you open a chat and ask it to. If you close your laptop for two weeks, nothing happens on your site's behalf during that time — the server doesn't run when you're not asking it something.

When you outgrow ask-and-answer

The MCP is great for the moment you're actively investigating something. It's less good as an early-warning system — it can't tell you your rankings dropped last Tuesday if you don't happen to ask on Tuesday.

That's the gap Percy fills. Percy is an AI Head of SEO that watches your Search Console every week on its own — no prompting required — builds a prioritized list of what to fix, emails it to you, and then re-checks the live page afterward to confirm the fix actually took and measure the traffic result. The MCP answers when you ask; Percy never waits to be asked.

If that's more what you need than an on-demand chat tool, connect your site to Percy — it's free for 7 days, no card required.