Connect Claude/ChatGPT to your marketing data with MCP
A working, honest walkthrough of connecting Claude and ChatGPT to GA4, Google Ads, Meta and a CRM over MCP — including the parts everyone under-counts: auth per tool, consent, approval gates, and the drift you'll have to chase.
The first tool is easy. A full harness is where it stops being a hobby. This guide walks you through what it actually takes to connect Claude and ChatGPT to the data your marketing runs on — GA4, Google Ads, Meta and a CRM — over MCP (Model Context Protocol). Not the idealized version, and not the scare story: the real one, including the parts every tutorial under-counts.
The honest headline up front: getting one AI model reading one source of marketing data is a quick win. Getting a maintained, multi-tool setup that your team can trust with real decisions is a job. Most write-ups stop after the easy bit. This one goes all the way to where the work actually starts.
Why MCP, and what it actually is
MCP is an open standard that lets an AI like Claude or ChatGPT call tools on external systems the way a browser calls APIs. Instead of copying a CSV into a chat prompt, the model can query your real GA4 property, live Google Ads campaigns or CRM records — and get current numbers, not whatever it guessed.
The protocol itself is the easy part. A server exposes tools; the AI client calls them. The work is everything around that, and each platform adds its own slice. To make the walkthrough concrete, here is the real job, tool by tool.
Tool 1: Google Analytics 4
Start here, because GA4 is the least painful and gives you the fastest win. To give Claude a read path into GA4 you need the Analytics Data API enabled, a Google Cloud project, OAuth credentials or a service account, and a server exposing GA4 metrics as MCP tools.
If you're running it yourself, that means: create the GCP project, enable the API, set up OAuth, download a key file, host the server, then handle token refresh and quota handling. It's a well-documented path and a reasonable afternoon for one property.
The quick win is real: within an hour or two, Claude can answer "which landing pages drove conversions this month?" from your actual data. That's the demo that makes the idea stick.
Tool 2: Google Ads
Google Ads is where the first under-counted cost shows up. The API speaks a query language (GAQL) you need to work with, the scopes are different from GA4's, and the audience you hand the credentials to matters — a service account created by one Google user does not automatically see another user's campaigns.
Auth per tool means a second OAuth flow, a second scope set, and a second thing that can be misconfigured. The setup is no longer an afternoon; it's an afternoon plus a few "why is this campaign missing?" debugging sessions that usually end with a permission level or a scope you forgot.
The payoff is worth it — Claude can audit wasted spend, compare campaigns, and draft where to shift budget. Just budget for the wiring.
Tool 3: Meta Ads
Meta raises the consent question. Facebook's and Instagram's ad APIs are gated behind app review and approval queues, and they carry a stricter idea of what your AI is allowed to read and — more importantly — to change.
This is the point where you have to decide what "connect" means. Read-only access to insights and campaign structure is one thing. Write access — pausing a campaign, moving budget, changing creatives — is a whole different consent conversation, and it's a policy decision before it's a code one. Who approves a write? What can never happen without a human in the loop?
A self-built server tends to start as a blank permission set you have to reason through yourself. That's fine for a single platform; it gets heavy quickly.
Tool 4: A CRM (HubSpot is the usual)
A CRM adds the cross-reference that makes everything else useful: did the people who came through those campaigns actually become customers? HubSpot is the common choice here, and it slots in the same way — its own OAuth flow, its own scopes for contacts, deals and pipelines.
The value only appears once more than one tool is connected. "Which source produced the most revenue per dollar" needs GA4 or the ad platforms and the CRM talking to the same model in the same conversation. That's where a single-tool server starts to chafe: you now have four servers, four auth flows, four version trackers, four failure modes — and your AI has to know which one to call just to answer a question that spans them.
The parts everyone under-counts
Every tutorial covers "add the MCP server, ask a question." Here's the honest list of what a working setup needs that the demos skip:
- Auth per tool, forever. Google access tokens expire in about an hour. Meta's and HubSpot's last longer but still expire. Every one of them needs refreshing, and a connector that doesn't do it silently fails — the AI just gets an error where it expected data.
- Consent and approval gates. Connecting AI to your accounts means deciding what can be written, by whom, and only after what. That's a policy layer — read-only by default, writes gated by explicit human approval — not a line in a config file.
- Drift. Marketing APIs change on their own schedule. Meta retires API versions roughly every six months; Google deprecates endpoints. If nobody is tracking it, a connector stops working — often weeks after the change, and quietly.
- Measurement. The point of connecting the data is to know whether a change worked — did the metric actually move, or didn't it. One connector is easy to keep aligned with that. Four tools, and the job — keeping auth alive, enforcing consent, chasing drift, and wiring the results into anything that measures them — is a standing expense. That is the multiplication problem: each new connector doesn't just add a server, it adds the full job of keeping that server alive and honest.
Where to go from here
None of this is an argument against MCP or against connecting your AI to your data — it's an argument for being honest about the scope. If you want to learn the protocol and you have one tool in mind, build it. It's a great way to understand what's under the hood, and it is genuinely achievable in an afternoon.
But if your marketing spans several platforms and you want your AI to work across them — with auth that stays alive, writes gated behind approval, drift handled for you, and every change measured — that's the point where a pre-built marketing MCP hub stops being a convenience and becomes the sensible call. That's the honest line, and it's the argument for buying the harness rather than building it: the first tool is easy, a full harness is ongoing work.
Read the full decision framework in the build-vs-buy marketing harness guide, and see what a marketing harness is if you want the definition. When you're ready to stop building the connector and start running the harness, connect your channels and see it measured.
Start free — no card required.