Google Analytics MCP: how to connect GA4 to Claude and ChatGPT (and when agencies need more)

The Google Analytics MCP server lets you ask an AI assistant questions about your GA4 data in plain English, and get answers built from live figures rather than a pasted export. Google’s own version is free, open source and read-only. The catch is that it is a developer tool: you install it on your machine and wire it up through Google Cloud.

I run a UK agency, so I tested this the way most account managers would want to use it: “What happened to organic traffic for this client last month, and why?” This guide explains what MCP is, what you can actually ask, how the official setup works at a high level, and the no-code route for agencies that need answers across many clients and platforms, not just one GA4 property.

What is MCP, in plain English?

MCP (Model Context Protocol) is an open standard that lets an AI assistant call tools to fetch data or take actions. You can read the spec at modelcontextprotocol.io. A tool is just a described function, such as “run a GA4 report for this property and date range”.

Without MCP, getting AI help with analytics means exporting a CSV, pasting it into a new chat and hoping nothing gets lost. With MCP, the assistant reads your question, picks the right tool, calls the Google Analytics APIs and explains the result. The assistant is the MCP client. The thing that connects to GA4 is the MCP server.

Is MCP the same as an API? No. The GA4 Data API is how software requests numbers from Google Analytics. MCP wraps those API calls so an assistant knows they exist and when to use them. Under the hood, a Google Analytics MCP server still talks to the Data API and Admin API.

What can a marketer ask GA4 through MCP?

You can ask anything the GA4 Data API can answer: traffic sources, landing pages, events, conversions, user behaviour over time, realtime activity and funnel steps. Google’s server also reads property configurations, so it can list custom dimensions and custom metrics or tell you which Google Ads accounts are linked.

The real gain is follow-up questions. You ask for sessions by channel, spot a drop in paid search, then ask which landing pages lost the most traffic, all in one chat. That is quicker than rebuilding an exploration in the GA4 interface every time.

Option 1: Google’s official Google Analytics MCP server

Google’s official route is the open source Google Analytics MCP server on GitHub, documented on the Google for Developers MCP page. The GitHub repo still labels it “Experimental”. Google’s page was last updated on 26 October 2026 and describes the server as read-only: it cannot create, modify or delete your Google Analytics configuration.

What the official server includes

The server exposes tools built on two Google Analytics APIs:

Tool group	Tools (as listed by Google)	What it lets you do
Admin API	get_account_summaries, get_property_details, list_google_ads_links, list_property_annotations	List your Google Analytics accounts and properties, check settings and linked Google Ads accounts
Data API	run_report, run_funnel_report, run_realtime_report, run_conversions_report, get_custom_dimensions_and_metrics	Run standard reports, funnels, realtime reports and conversion reports

The GitHub README also describes an optional --allow-write flag for creating property annotations, which needs the edit scope. Leave it off unless you have a clear reason.

What you need before you start

At a high level, Google’s setup needs:

1. Python 3.10 or higher and pipx (or uv) to run the analytics-mcp package.
2. A Google Cloud project with the Google Analytics Admin API and Google Analytics Data API enabled.
3. Credentials. Application Default Credentials with the analytics.readonly scope. You can use your own user credentials through an OAuth client and the gcloud command line, a service account JSON file, or service account impersonation.
4. Viewer access (at least) on each Google Analytics property you want to query.
5. An MCP client. Google’s page lists Claude Desktop, Claude Code, Cursor and Antigravity. It notes that Gemini CLI support is deprecated.

You then add the server to your client’s configuration, pointing it at your credentials file and project ID, and restart Claude Desktop or your client so it picks up the new tools.

Where the official server shines, and where it struggles

It is the best choice when one analyst wants deep, flexible access to GA4. Any dimension, any metric, realtime data and funnels, for free, with Google maintaining the code.

The friction shows up in agency life:

• It runs locally. Every team member needs Python, a terminal and their own credentials. That is fine for an analyst, less so for an account manager.
• It is GA4 only. Search Console, Google Ads performance, Meta Ads and your CRM need separate servers, each with its own setup.
• It does not plug into ChatGPT directly. OpenAI’s help centre says ChatGPT connects to remote MCP servers, not ones on your laptop.
• Credentials sprawl. Twenty clients means twenty properties under whichever Google login or service account you configured, and someone has to keep that tidy.

How to connect Google Analytics to ChatGPT

To connect Google Analytics to ChatGPT, you need a remote (hosted) MCP server, because ChatGPT does not run local servers. OpenAI’s help centre says custom MCP server connections and full MCP support are available on ChatGPT Business and Enterprise/Edu on the web, that Pro users can connect custom MCP servers with read/fetch permissions, and that only admins and owners can publish custom connections in a workspace. It also points to a Secure MCP Tunnel if your server sits on a private network.

In practice, most marketers use a hosted connector for this. Windsor.ai, for example, offers a ChatGPT connector for GA4, and Ryze AI lists ChatGPT among the clients for its hosted Google Analytics MCP. Details for both are in the comparison table below.

How to connect GA4 to Claude

Claude is the easiest assistant to start with because it supports both routes. For Google’s server, Claude Code and Claude Desktop are on Google’s supported list, and you run the server locally. For a hosted server, Claude lets you add a custom connector: you paste the server URL, sign in on the provider’s consent screen and grant access. No terminal needed.

Option 2: a hosted MCP for agencies with many clients

The no-code route suits agencies that need answers across many clients and platforms. This is where MYCONTENTLAB’s MCP server fits, and it is worth being clear about what it is and is not.

MYCONTENTLAB’s MCP connects your assistant to the data already in your MYCONTENTLAB workspace. You connect each client’s platforms once in the workspace (GA4, Search Console, Google Ads, Meta Ads, LinkedIn, TikTok, HubSpot and more are listed on the integrations page), and the MCP server lets the assistant read across them. According to the MCP setup guide, the tools are:

Tool	What it returns
list_clients	Active clients with company name, industry and website
get_client_metrics	Monthly performance metrics across each client’s connected ad and social platforms
get_report_summary	AI summaries, key findings and action items from reports
list_mentions	Brand mentions and sentiment from Social Listening
get_trends	Trends and opportunities linked to the client’s industry
get_crm_summary	Deal stages, total pipeline value and conversion metrics
list_content_ideas	Content Lab ideas, hooks, angles and opportunity scores

Setup. The recommended route is Claude Desktop or Claude.ai on the web: add a custom connector with the MYCONTENTLAB endpoint, sign in through OAuth 2.1 and click Allow Access. The account must have the owner or manager role. Claude Code works with a personal access token. The guide also lists Cursor and Antigravity, and the MCP page says other MCP-compatible assistants can connect with a token. ChatGPT is not on the documented list yet, so do not plan a client workflow around it until it is.

The honest trade-off. MYCONTENTLAB’s MCP works from monthly workspace data and saved reports. It does not run arbitrary GA4 Data API queries, so it cannot tell you what is happening on a client’s site right now, or break last Tuesday down by landing page. For that, Google’s server is the better tool. Where MYCONTENTLAB helps is the account manager’s question: “How did this client do last month across everything we run for them, and what should we do next?”

Google Analytics MCP options compared

Details below come from each provider’s own page, checked October 2026. Prices change, so confirm before you buy.

	Google Analytics MCP (official)	MYCONTENTLAB MCP	Ryze AI GA4 MCP	Windsor.ai	Stape GA4 MCP
Where it runs	Locally on your machine	Hosted	Hosted	Hosted	Requires Node.js v18+ (hosting not stated)
Setup	Python, Google Cloud project, APIs, credentials	Custom connector plus OAuth sign-in	Google sign-in	Install connector in ChatGPT and authorise	Google OAuth plus Node.js
GA4 depth	Full Data API reports, realtime, funnels, config	Monthly metrics and report summaries, no raw queries	Nine read-only GA4 tools	92 GA4 metrics and 390 dimensions listed	Reports, realtime, property details, annotations
Other platforms	GA4 only	Social, ads, SEO, CRM, mentions, trends, Content Lab	Google Ads, Meta Ads, TikTok, Search Console, Shopify	350+ data sources	GA4 only (separate GTM server)
Assistants named	Claude Desktop, Claude Code, Cursor, Antigravity	Claude (documented), Cursor, Antigravity, others by token	Claude, ChatGPT, Cursor, Windsurf, VS Code and more	ChatGPT (setup shown), Claude, Gemini, Perplexity	Claude Desktop named
Read-only	Yes (optional write flag for annotations)	Yes	Yes	Not stated	Not stated
Built for agencies	No	Yes, multi-client workspace	Workspace based	Priced by sources and accounts	No
Price	Free, open source	Plans from £47/mo; confirm MCP access for your plan on pricing	SEO Starter $129/mo, MCP included	Free plan, paid from $19/mo	Free

Sources: Google for Developers, GitHub, get-ryze.ai/google-analytics-mcp, windsor.ai/connect/google-analytics-to-chatgpt-integration, stape.io/solutions/mcp-server-ga4.

Which should you pick? If you are a solo analyst who lives in GA4, use Google’s server. If you need ChatGPT with GA4 specifically, a hosted connector such as Windsor.ai or Ryze AI is the realistic path. If you run an agency and want one connection that covers every client’s results, mentions, pipeline and content ideas in Claude, MYCONTENTLAB is built for that. Plenty of teams will use two: Google’s server for deep dives, a workspace MCP for weekly account work.

14 prompts agencies can actually ask

These are prompts I would hand to an account manager. Replace [Client Name] with a real client. The first group needs a server with raw GA4 access, such as Google’s. The second works with MYCONTENTLAB’s MCP.

With Google’s Google Analytics MCP server (one property):

1. “List my Google Analytics accounts and properties, then give me details about the property with [Client Name] in the name.”
2. “Compare sessions and key events by default channel group for the last 28 days against the previous 28 days.”
3. “Which landing pages from paid traffic had the highest conversion rate last month? Show sessions alongside.”
4. “What are the most popular events in this property over the last 90 days, and which are marked as key events?”
5. “How many users are on the site right now, and which pages are they on?” (realtime report)
6. “List the custom dimensions and custom metrics in this property and flag any that look unused.”
7. “Which Google Ads accounts are linked to this property?”
8. “Build a funnel from landing page view to form start to form submit for the last 30 days. Where is the biggest drop?”

With MYCONTENTLAB’s MCP (across clients and platforms):

9. “List my active clients.”
10. “Summarise last month’s marketing performance for [Client Name] across every connected platform.”
11. “What are the key findings and action items in the latest report for [Client Name]? Turn them into a three-line email for the client.”
12. “Show recent negative brand mentions for [Client Name] and group them by topic.”
13. “Summarise the CRM pipeline for [Client Name], including deal stages and total value, and compare it with last month’s ad results.”
14. “Show the highest-scoring Content Lab ideas for [Client Name] and suggest three priorities for next month based on what performed best.”

A tip from our own use: ask the assistant to show which tool it called and the date range it used. It makes checking far quicker.

Security, privacy and client data

Read-only access is the starting point, not the finish line. Here is the checklist we use.

• Use read-only scopes. Google’s server needs only analytics.readonly unless you switch on write tools. MYCONTENTLAB’s scopes are all read-only (read:clients, read:metrics, read:mentions, read:trends, read:crm, read:contentlab), and the MCP cannot create, edit or delete anything in the workspace.
• Scope access to the clients who need it. MYCONTENTLAB personal access tokens can cover all clients or only specific ones, expire after 30, 90 or 365 days, and can be revoked at any time. Revocation takes effect immediately. With Google’s server, the reach is whatever the Google login or service account can see, so keep that account lean.
• Know where data goes. When an assistant answers, the figures it fetched enter the conversation with your AI provider. Check that provider’s data use and retention terms, and whether your plan trains on chats.
• UK GDPR and client contracts. GA4 data is usually aggregated, but CRM and mention data can include personal data. As an agency you are normally a processor for client data, so an AI assistant is another subprocessor. Check your contracts, update your records and tell clients if needed. The ICO’s guidance on AI and data protection is a sensible starting point.
• Review access regularly. Look at last-used dates on tokens, remove connectors staff no longer use and remove leavers on day one.

Limitations: always check the numbers

AI answers about analytics data are a first draft, not a report. Four things catch people out.

Data freshness. Google’s data freshness guidance says standard GA4 properties typically process intraday data in 2 to 6 hours and daily data in about 12 hours, and that processing can take 24 to 48 hours. Ask about yesterday and you may get partial figures. Workspace tools like MYCONTENTLAB answer from data as current as the last sync, and monthly metrics only make sense once the month has closed.

Sampling and thresholds. GA4 can apply sampling to some large queries and data thresholds when Google signals are on. The assistant may not tell you. If a number looks oddly rounded or low, check the same report in GA4.

AI mistakes. Assistants can pick the wrong date range, confuse users with sessions, misread “key events” or invent a reason for a change. Ask for the exact query and date range, and do not accept explanations that are not backed by a figure.

Attribution gaps. GA4, ad platforms and CRMs count conversions differently. An assistant that blends them without saying so will give you a tidy, wrong story. Our guide to cross-platform reporting across GA4, ads, social and CRM explains how to present those differences to clients.

Where this fits in your reporting

A Google Analytics MCP server is brilliant for questions. It is not a replacement for a consistent monthly report that clients can trust and compare. We use MCP to investigate and draft, then put the checked conclusions into the client report. If you are rebuilding your process, our practical guide to digital marketing reporting covers the structure.

If you want to try the agency route, connect your workspace from the MCP page and follow the setup guide. Every plan has a 14-day free trial. Start by asking for your list of active clients, then ask for one client’s monthly summary and check it against the platforms. That one exercise will tell you more than any comparison table.