How to connect Google Search Console to ChatGPT: 5 options

ChatGPT cannot open Search Console by itself: something has to fetch the data from Google and hand it over. There are five practical ways to do that, from least to most set-up: upload an export, build a custom GPT whose action calls the Search Console API, run your own MCP server, add a hosted MCP server such as AutoPSEO's, or call the OpenAI API with a remote MCP tool for scheduled jobs.

The details below were checked against Google's and OpenAI's documentation on 12 October 2026.

The limits every route shares

Whichever route you pick, the data comes from the same place and carries the same limits:

• The interface exports at most 1,000 rows, and rare queries are left out of every table for privacy (Search Console data limits).
• The API returns up to 25,000 rows per request, with paging for more, but Google says it returns the top rows rather than guaranteeing every row, and the daily maximum is 50,000 rows per site per search type. Each site and each user can make up to 1,200 queries a minute, and queries grouped by both page and query are the most expensive (Search Analytics API, usage limits).
• The daily bulk export to BigQuery includes all performance data except anonymised queries, for sites that need everything (bulk data export).
• Data is two to three days behind, and days are counted in Pacific Time.

There is also a limit on the ChatGPT side. When the model receives raw rows, it has to add up, average and compare them itself. Ask it to state the date range it used and to show how it reached any total, and check a few numbers against Search Console.

Option 1: upload an export

How it works. In the Performance report, set the date range and filters, export the table, and upload the file to a chat. ChatGPT accepts spreadsheets as XLSX, XLS, CSV and TSV (OpenAI file uploads).

Good for: a one-off question about a small site.

Where it stops: 1,000 rows per export, a snapshot that is out of date the moment you ask about another period, and no way to export queries and pages together from the interface, so "which page ranks for this query?" needs a second export and some matching. Export the Queries and Pages tables separately, and say in your prompt which date range each covers.

Option 2: a custom GPT with an action

How it works. GPT actions let a custom GPT call an outside REST API, described by a schema, with authentication you configure (GPT actions). For Search Console you create a Google Cloud project, enable the Search Console API, set up an OAuth client and consent screen, describe the searchanalytics.query method in the schema, and connect the GPT's action to your OAuth client.

Good for: a developer who wants a private assistant and is comfortable owning the integration.

Where it stops: you maintain the schema, the OAuth set-up and any changes Google makes. If people outside your organisation use it, Google's OAuth rules apply: apps that request sensitive or restricted scopes generally need verification, and an unverified app is limited to 100 new users (Unverified apps). The GPT receives raw rows, so the arithmetic is still the model's job.

Option 3: run your own MCP server

How it works. The Model Context Protocol (MCP) is an open standard for connecting assistants to tools, and several open-source Search Console MCP servers exist. You run one with your own Google OAuth credentials at a public HTTPS address, then add it to ChatGPT as a custom MCP server. OpenAI also offers a Secure MCP Tunnel for testing a server that is not public yet (Connect from ChatGPT).

Good for: developers who want to read every line of the code and use one connection across several assistants, since Claude, Cursor and others speak MCP too.

Where it stops: you are responsible for hosting, security and updates, and most such servers pass the API's rows through without analysis, with the same row and quota limits as option 2. OpenAI's own documentation warns that it does not verify remote MCP servers, that a malicious server can hide instructions in its responses, and that official servers hosted by the service provider are the safer choice (Remote MCP risks).

Option 4: add a hosted MCP server

How it works. A provider runs the server and handles the Google connection; you add its address to ChatGPT and sign in. AutoPSEO's server lives at https://app.autopseo.com/mcp. According to OpenAI's developer documentation, you add a custom MCP server from ChatGPT's Plugins page: choose the plus button, then Add custom MCP server, give it a name, enter the server URL including the /mcp path, configure authentication, accept the risk notice, and create it. OpenAI notes that account and workspace policies apply, so in a company workspace an admin may need to allow it. ChatGPT has renamed these menus several times, so if your account shows something different, follow the step-by-step guide on AutoPSEO's MCP & ChatGPT page or the ChatGPT connection guide. You sign in with your AutoPSEO account; no Google Cloud project or OAuth app of your own is involved.

What ChatGPT gets. Besides plain queries over your data, the 19 Search Console and GA4 tools include finished analyses computed on AutoPSEO's side, so the model does not have to do the arithmetic: content_decay for pages losing clicks, cannibalization for searches split between your pages, ctr_benchmark against your own CTR curve, page_poaching for striking-distance searches, compare_periods, rank_changes, url_inspection and more. They cost no credits, and when you give no dates they use the last 28 settled days against the 28 before. The demo property sc-domain:demo.autopseo.app lets you try them before connecting your own site.

The same connection reaches AutoPSEO's research tools (which spend research credits) and its content engine. ChatGPT can take a topic from your site's queue, read its brief, write the article and submit it; the draft passes the same checks as AutoPSEO's own writer and uses none of your article quota. Drafts wait for review, and AutoPSEO's instructions tell the assistant to publish only when you say so (writing with your own AI).

Where it stops: you need an AutoPSEO account (a 7-day trial, then plans from $19 a month), and the connection only sees the properties your account can see, with the permissions your role allows. Calls are limited to 60 a minute per key and 1,000 an hour per account.

Option 5: the OpenAI API, for scheduled jobs

How it works. For reports that run without anyone opening ChatGPT (a Monday summary posted to a channel, say), the Responses API can call a remote MCP server directly. You add a tool of type mcp with a server_url and a server_label, optionally restrict it to named tools with allowed_tools, and decide with require_approval whether each call needs approval; by default OpenAI asks before sharing data with a remote server (Remote MCP in the Responses API). AutoPSEO's MCP & API page has a ready-made tool entry that authenticates with an API key, and a read-only key keeps the job from changing anything.

Good for: recurring reports and automations in your own code.

Where it stops: you write and run the code, and you pay for the tokens. If you would rather not use MCP at all, AutoPSEO's REST API at https://app.autopseo.com/v1 offers the same 19 Search Console and GA4 tools with the same keys.

Which one should you use?

Getting good answers once connected

• Name the property and the period. With several sites connected, say which one, and ask the assistant to state both date ranges it compared.
• Ask for the analysis by name. "Run the cannibalisation analysis for this site" maps onto one tool; "check for keyword overlap" might map onto several.
• Remember the delay. A question about yesterday reads as a drop because the newest days are incomplete.
• Act on the answer with a playbook. Once ChatGPT has the numbers, the Week 1 quick wins checklist and the content decay and cannibalisation playbooks turn them into changes.

If you are choosing between several MCP servers to put next to Search Console, the comparison of the best SEO MCP servers covers what each one connects to.