How to connect Search Console to ChatGPT: routes, limits and trust
ChatGPT does not read Search Console on its own: a remote MCP server has to call the API for you. This playbook covers the three routes, what OpenAI documents for developer mode and the API, the Google row limits every connector inherits, and a way to check that the numbers still match.
✓ Checked against OpenAI and Google documentation · 4 October 2026
What “connecting Search Console to ChatGPT” means
ChatGPT cannot read your Search Console by itself. To put the data in a conversation, something has to call the Search Console API for you. In ChatGPT that something is a remote server speaking the Model Context Protocol (MCP) which you add as an app or custom connector (developer mode guide). We found no Search Console connector published by OpenAI or by Google, so every route below depends on a third-party or self-built server, and you should decide how far you trust it before signing in.
The vocabulary has changed. OpenAI’s help says the app directory was migrated to the plugin directory on 9 July 2026 and that existing app connections are unaffected (Apps in ChatGPT). Treat “connector”, “app” and “plugin” in older tutorials as labels around the same idea and check what the server on the other end returns.
Three routes that carry the data
| Route | What you do | What to know |
|---|---|---|
| Directory app or plugin | install a published Search Console app and sign in | OpenAI’s help says availability differs by plan and that Free and Go accounts can discover plugins but not use plugin extensions; check the plugin’s own page |
| Custom MCP server | turn on developer mode and add the server URL | developer mode is available to Pro, Plus, Business, Enterprise and Education on the web; the path is Settings → Security and login; remote servers only, using SSE or streaming HTTP |
| Responses API | add an mcp tool with server_url to your own request | for scripts and agents; you pass the credential in the authorization field and resend it every time (remote MCP guide) |
The built-in file sources, such as a drive or chat tool, index documents and messages. Search Console is a metrics API, so exporting a report into a drive gives you a frozen snapshot, not live data.
Adding a custom server in developer mode
- Enable developer mode under Settings → Security and login. Settings move between releases, so search for “developer mode” if the path differs, and on business plans an admin may control it.
- Add a remote server URL. ChatGPT accepts remote servers; a local process started on your own computer will not register unless you deploy it behind a public HTTPS address.
- Choose authentication. OpenAI documents OAuth, no authentication or mixed authentication, and does not list API keys as an option here.
- Sign in once and read the tool list. Look at which tools can write, not only read.
- Expect confirmations. OpenAI says write actions require confirmation by default; review the inputs before approving, because an incorrect write can change or share data.
Decide whether to trust the server
OpenAI’s own guide tells developers to trust any remote MCP server they use, because a malicious server can exfiltrate data from anything that enters the model’s context. For Search Console that means data about your site and queries. Before connecting, check who runs the server, what Google scopes it requests, whether it can write, and whether it stores your data. Prefer read-only access for analysis and use the Responses API allowed_tools filter to expose only the tools you need.
The row limits behind every connector
A connector that calls the Search Analytics API inherits Google’s limits (query reference, getting all your data).
| Limit | Value | Effect |
|---|---|---|
| Default rows per request | 1 000 | silently short unless the connector sets rowLimit |
| Maximum rows per request | 25 000 | more needs startRow pagination |
| Rows per day | 50 000 per site per search type, sorted by clicks | a daily ceiling on what can be pulled |
| Detail dimensions | “some data may be dropped” | page or query detail can omit rows |
| Freshness | typically 2–3 days | the last days may be absent or incomplete |
Results are sorted by clicks, so a truncated pull keeps the head of your data and loses the long tail. That is also true of the CSV export in the interface, which is why a long-tail question asked on a truncated pull deserves a doubtful answer.
Make sure the numbers still match Search Console
- Compare one number. Ask for total clicks and impressions for a fixed date range and compare them with the Performance report (Performance report).
- Ask for the dimension and aggregation. Totals by property and totals by page differ; Google documents why (data discrepancies).
- Check the date range. Stay out of the last two or three days.
- Check row counts. Ask how many rows came back; a round number like 1 000 suggests a default limit.
- Do not average positions. Position is an impression-weighted metric, so ask for the reported position, not the mean of a list.
Prompts that stay checkable
- “List the 20 queries with the most clicks for the last 28 days, with clicks, impressions, CTR and position, and say how many rows you retrieved.”
- “Compare the last 28 days with the previous 28 and list the pages that lost the most clicks; show the numbers.”
- “Which queries are in positions 4–20 with at least 100 impressions? Show impressions and position.”
Each answer can be checked against the report, and each links to a workflow you can run by hand: striking distance, content decay and CTR recovery.
Troubleshooting
| Symptom | Likely cause |
|---|---|
| No developer mode switch | your plan is not one of those OpenAI lists for it, or a workspace admin turned it off |
| Server URL rejected | it is not a remote HTTPS endpoint, or it is a local process |
| Nowhere to paste an API key | the connector flow documents OAuth, not keys; use the Responses API when you must pass a credential |
| Totals differ from Search Console | truncated rows, a different dimension or aggregation, or a date range that includes the freshest days |
Using SEOKit’s MCP server
SEOKit publishes its own MCP server for the Search Console workspace, with scoped credentials and separate read and write tools. The setup page is MCP and the ChatGPT app. It is one third-party option among several, so apply the trust checks above to it as well.
ChatGPT and Search Console, answered
Is there an official Search Console connector in ChatGPT?
We found none from OpenAI or Google. Every route uses a third-party or self-built server.
Do I need a paid plan?
Developer mode, which adds a custom MCP server, is documented for Pro, Plus, Business, Enterprise and Education accounts on the web. Whether a directory plugin works on Free depends on the plugin; check its page.
Can I paste an API key into a custom connector?
OpenAI’s guide lists OAuth, no authentication and mixed authentication. For key-based access use the Responses API authorization field.
Why does ChatGPT show fewer rows than Search Console?
The API returns 1 000 rows unless asked for more, allows at most 25 000 per request and 50 000 per day per site per search type, and can drop some data in detailed queries.
Can I use a local open-source MCP server?
Not directly: ChatGPT connects to remote servers. Deploy it behind HTTPS with authentication, or use a client that runs local servers.
What to read next
Start with week 1 quick wins to know what a good answer looks like, or go back to the Brave Search playbook.