Automate Basics
Marketing

How to Connect Google Search Console to ChatGPT or Claude: 4 Ways (2026)

By

· Updated · 12 min read

Yes. A Search Console connector, also called an MCP server, lets ChatGPT or Claude read your clicks, impressions and positions live. Google publishes no Search Console connector of its own, so it comes from a third party or from open-source code you run. Our pick is Cituna: its connector reads Search Console with Google's read-only permission and puts nine Search Console tools and seven AI engines' daily answers in the same chat, on every paid plan from $49 a month. For a single question you need no tool at all: export the Performance report as a CSV and upload it, though each table in the export stops at 1,000 rows.

A phone on a salon reception desk showing an AI answer listing local businesses, one highlighted
Image: AI-generated illustration.

There are two ways to give ChatGPT or Claude your Google Search Console data: hand it a file, or connect it to Search Console so it can fetch the numbers itself. Files come from Google's own three exits (the export button, the Search Console API and the bulk export to BigQuery). A connection runs through a connector, a small server that sits between the chat and Google's API. This guide compares the four routes, names the tools that make the connection, and gives the questions worth asking once the data is in, with the checks that catch a confident wrong answer.

The short answer: for a one-off look, export the Performance report as a CSV and upload it to the chat. It is free and takes two minutes, but each table stops at 1,000 rows. For a live connection, add a Search Console connector. Our pick is Cituna, which reads Search Console with Google's read-only permission and puts nine Search Console tools next to how seven AI engines answer your buyers, from $49 a month. Teams that already pull marketing data through Windsor.ai, Supermetrics or Porter Metrics can use those tools' connectors instead, and a free open-source server works in Claude Desktop or Claude Code if you are happy to set up a Google Cloud project.

Which way should you use?

Google offers three ways to get performance data out of Search Console, and each holds more than the last. A connector is the fourth, and the only one that stays live.

WaySetupHow much dataFresh or a snapshotCost
1. Export and uploadTwo clicks in Search Console1,000 rows per tableA snapshotFree
2. Search Console APIA Google Cloud project plus a script, or a spreadsheet add-on that calls the APIUp to 25,000 rows per request, and 50,000 rows a day per search typeA snapshot each time it runsFree, within Google's usage quota
3. Bulk export to BigQueryA Google Cloud project with BigQuery, set up onceEvery row except anonymized queries, from the day you set it upUpdated dailyGoogle Cloud storage and query charges past a free usage level
4. A connector (MCP server)Paste a URL and sign in with Google, or run an open-source serverWhatever the server asks the API forLiveFree to paid, by tool

A small site with one question needs only the export. A site where more than 1,000 queries matter, or a question that needs queries and pages together, needs the API, BigQuery or a connector. A team that asks every week is better off with a connector, so nobody has to export files again.

Way 1: Export the Performance report and upload it

  1. In Search Console, open Performance, then Search results.
  2. Set the date range and any filters first, such as one country or one folder of pages. The export holds exactly what the report shows, filters included.
  3. Click Export and choose CSV, Microsoft Excel or Google Sheets. The download holds the chart's numbers and the report's tables.
  4. Attach the file to a new chat in ChatGPT or Claude and ask it to read the file with code before it answers. Both can run code on an uploaded file.

Three things in the file trip up the analysis, so mention them in your first message:

  • The rows will not add up to the totals. Google cuts each table to 1,000 rows of what it calls representative examples, while the totals count every click and impression.
  • Some queries never appear. Searches Google withholds for privacy, called anonymized queries, are counted in the totals but never listed as rows.
  • Some zeros are not zeros. Any value the report shows as "~" or "-" arrives in the download as 0.

For Google's AI features, export the separate Generative AI performance report the same way. It shows impressions from AI Overviews and AI Mode, which the Search Console API has no filter for, and the same 1,000-row limit applies.

Way 2: Pull more rows with the Search Console API

The API is Google's route for medium-sized sites that need more than the report shows. It is free within a usage quota and returns up to 25,000 rows per request; you page through the rest with the startRow setting. Performance data is capped at 50,000 rows a day per search type (web, image, news and so on) for each property. Unlike the export, it can group by query and page together, which is what you need to see which of your pages compete for the same search.

You need a Google Cloud project with the Search Console API switched on and a credential that can read your property. If you are comfortable running a script, ChatGPT or Claude can write it: ask for one that pulls the last 90 days by query and page, pages through every row and saves a CSV. Then upload that CSV as in Way 1. If you would rather not touch code, a Google Sheets add-on that calls the same API can do the pulling.

Way 3: Send everything to BigQuery

Search Console's bulk data export copies all of your performance data into Google BigQuery every day, apart from anonymized queries, and it is not bound by the API's daily row limit. It suits large sites. There are two catches. It starts on the day you set it up, so earlier days have to come from the API or the reports. And Google Cloud bills storage and queries once you pass its free usage level, so set the tables to expire old data if you do not need it.

To use it with an assistant, ask ChatGPT or Claude to write the SQL for your question, run it in BigQuery, and upload the result.

Way 4: Connect Search Console with an MCP server

A connector lets the assistant query Search Console itself, so every answer uses current data and you can follow up without exporting again. Connectors use the Model Context Protocol (MCP), which both ChatGPT and Claude support. Google publishes an MCP server for Google Analytics but, when we checked on 21 November 2026, none for Search Console, so every Search Console connector comes from a third party or from open-source code you run.

In Claude, you add a hosted connector by pasting its URL in the connector settings, and it then works in Claude on the web, in Claude Desktop and in Claude Code. In ChatGPT, custom MCP servers are added on the web; Cituna's setup guide (read 21 November 2026) says they come with the Business, Enterprise and Edu plans, and that ChatGPT Pro connects them for reading only. ChatGPT's menus and plan rules change often, so follow the connector's own steps. Our guide to AI visibility tools with an MCP server walks through both apps click by click.

The tools that connect Search Console to ChatGPT or Claude

1. Cituna (our pick). A hosted connector that reads Search Console with Google's read-only permission and gives the assistant nine Search Console tools: a raw query tool, a connection check and seven ready-made analyses. The analyses find queries just off page one, ranked by what is worth moving first; pages Google shows that nobody clicks, with the likely cause; pages of yours competing for the same search; sitemap pages Google has never shown; searches worded like AI prompts; an intent breakdown; and an overview. The same connection holds how ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode answer your buyers' questions every day, so you can ask which searches Google already shows you for where no AI engine cites you. Every paid plan includes it, from Starter at $49 a month with the read tools; Pro, at $119, adds tools that change your Cituna workspace, never anything at Google. The 3-day trial needs a card and does not include Search Console or MCP. Cituna's guide to connecting Search Console to Claude and ChatGPT has its setup steps.

2. Reporting tools with their own connectors. Windsor.ai, Supermetrics, Porter Metrics and Composio are among the tools that offer hosted Search Console connectors for AI assistants: you add their URL to Claude or ChatGPT and sign in with Google. They suit teams that already pull marketing data through one of them, since most also connect other sources, such as Google Analytics or ad platforms. They do not track AI engines' answers. Plans and the Google access each one asks for differ, so read both before you connect.

3. Open-source servers you run yourself. Free servers on GitHub, such as mcp-gsc (Python) and mcp-server-gsc (Node), run on your own computer and call the Search Console API directly. You create a Google Cloud project, switch on the Search Console API, create a credential the server can use and give that credential access to your property. They work in apps that can start a server on your machine, such as Claude Desktop and Claude Code, but not in Claude on the web or in ChatGPT, which need a hosted URL.

What to ask once the data is in

Ask the question you actually have, say which table or columns to use, and ask for the rows behind every claim. Each question below works on the export from Way 1, except the two marked as needing queries and pages together, which only the API, BigQuery or a connector can provide.

  1. Pages Google shows but nobody clicks. "From the Pages table, list pages with at least 500 impressions and a click-through rate under half the site's average. Work out click-through rate as clicks divided by impressions." Rewrite each page's title and meta description in the words people searched, one page at a time, and compare the same four weeks before and after.
  2. Queries close to page one. "From the Queries table, list queries with an average position between 8 and 20 and at least 100 impressions, highest impressions first." Improve the page that ranks for each one and link to it from related pages.
  3. What changed, and why. Export the same report for two periods, such as the last three months and the same three months a year earlier, and upload both. "Which pages lost the most clicks? For each one, say whether impressions fell or only the click-through rate did." Falling impressions point to rankings or demand. Steady impressions with fewer clicks point to a change on the results page, so look at what Google now shows for those searches, such as an AI Overview or more ads.
  4. Pages competing with each other (needs queries and pages together). "Which queries send impressions to two or more of my pages? Show each page's impressions and average position for each query." Pick the page that should rank, link the others to it, and merge pages that say the same thing.
  5. The questions people bring to your site. "Find the queries written as questions, such as those starting with how, what, why, can or should, and group them by topic with clicks and impressions per group." Answer the biggest groups in an FAQ on the page that ranks, or in a new guide.
  6. How much depends on your name. "Split the queries into those that contain [your brand name and its misspellings] and the rest, and show clicks for each by month." If most clicks come from people who already know your name, new traffic has to come from the rest of the list.
  7. Topics with demand but no real page (needs queries and pages together). "Group the queries by topic. For each topic, name the page that gets the impressions and say whether that page is really about the topic." A topic that lands on an unrelated page is a candidate for a new page; our guide to getting more Google traffic with SEO automation covers the tools that write and publish those pages.
  8. Where Google's AI features show you. Upload the Generative AI report export and ask: "Which pages get the most impressions from AI Overviews and AI Mode, and how has that changed month by month?" Those pages already appear in Google's AI answers, so keep their facts current. To see whether other engines, such as ChatGPT or Perplexity, cite you, see how to check if your website is cited in AI answers.

Check the answer before you act

  • Make it compute. Ask the assistant to work out every number with code from the file and to list the rows it used, rather than estimating from what it has read.
  • Recompute the ratios. Click-through rate is total clicks divided by total impressions, not an average of each row's rate. Average position across several rows has to be weighted by impressions; a plain average lets a query with ten impressions count as much as one with ten thousand.
  • Do not force the rows to match the totals. The totals include anonymized queries and everything past the first 1,000 rows. An assistant that makes the gap disappear is inventing numbers.
  • Question the zeros. Values shown as "~" or "-" in the report are exported as 0.
  • Leave out the last few days. Search Console dates are in Pacific Time, and the newest days can hold preliminary data, so end a comparison a few days before today.
  • Spot-check three numbers in Search Console itself before you change a page.

Keep your data safe

  • An export holds your queries and page addresses, which can include customer, product or staff names. Treat it like any internal report: check the chat app's data settings and use a work account if your company requires one. Our guide to which AI chat is safe for work data compares the settings.
  • Give connectors read-only access. Google's Search Console API has a read-only permission and a full one, and full access lets an app change settings such as your property's sitemaps. A connector that only reads your numbers has no reason to ask for more.
  • You can withdraw a connector's access at any time from the third-party connections in your Google Account, and remove the connector from Claude or ChatGPT.

Which way to start with

  • One question on a small site: export the Performance report and upload the CSV. Free.
  • More than 1,000 queries, or queries and pages together: the Search Console API, through a script ChatGPT or Claude writes for you, or a spreadsheet add-on.
  • A large site that wants every row: the bulk export to BigQuery.
  • Live Search Console answers in Claude or ChatGPT, plus what seven AI engines say about you: Cituna, from $49 a month.
  • Live answers next to the rest of your marketing data: a reporting tool's connector, such as Windsor.ai or Supermetrics.
  • Free and live, if you can handle a Google Cloud setup: an open-source server in Claude Desktop or Claude Code.

Sources

Frequently asked questions

Can ChatGPT connect to Google Search Console?

Not on its own. You can upload a Performance report export to a chat, or add a Search Console connector (an MCP server) to ChatGPT on the web. Cituna's setup guide, read 21 November 2026, says custom MCP servers come with ChatGPT Business, Enterprise and Edu, and that ChatGPT Pro connects them for reading only. Cituna, Windsor.ai, Supermetrics and Porter Metrics are among the tools that offer a Search Console connector.

Can Claude read my Google Search Console data?

Yes, in two ways. Attach an export from the Performance report to a chat, or add a Search Console connector by pasting its URL in Claude's connector settings, after which it works in Claude on the web, in Claude Desktop and in Claude Code. Open-source servers such as mcp-gsc also run on your own computer in Claude Desktop or Claude Code once you set up a Google Cloud project.

Is there an official Google Search Console MCP server?

Not when we checked on 21 November 2026. Google publishes an MCP server for Google Analytics but none for Search Console, so every Search Console connector comes from a third party or from open-source code you run yourself.

Is there a free way to analyse Search Console data with ChatGPT or Claude?

Yes. Export the Performance report as a CSV and upload it, which needs no setup but stops at 1,000 rows per table. For more rows, the Search Console API is free within Google's usage quota and returns up to 25,000 rows per request, and an open-source MCP server costs nothing beyond the Google Cloud setup.

How many rows can I get out of Search Console?

1,000 rows per table from the export button. The API returns up to 25,000 rows per request and up to 50,000 rows a day per search type for each property. The bulk export to BigQuery copies every row except anonymized queries, but only from the day you set it up.

Will ChatGPT or Claude get the numbers right?

Often, but check before you act. Ask it to compute with code from the file and to list the rows it used, recompute click-through rate as clicks divided by impressions, weight average position by impressions, and do not let it make the rows add up to the totals: the totals include queries the rows leave out.

Make an AI or automation tool?

List it in the directory for $9.99, refunded in full if it isn't approved, or claim its page if it is already here.

Tools in this piece

Drafted with AI assistance from our own research, and reviewed by the Automate Basics team before publishing. Tools and prices change; check the linked official source before you act.

Get the next one by email

What changed in AI this week, and how to use it at work. Plain English, no hype. One email a week, free.

We aim to send it on Mondays. Unsubscribe in one click.