Connecting Claude or ChatGPT to your marketing data: step by step and the first ten questions

The MCP connector is a URL and an authorization. What separates people who use it from people who tried and gave up is what happens in the next ten minutes: picking the project, understanding what the AI sees, and asking questions it can answer well. This is the script.

The technical part of connecting an AI to an MCP server takes under two minutes. The part that decides whether the connection gets used takes ten: the first questions. Whoever starts with “give me a full report” gets something generic and gives up. Whoever starts with a small, verifiable question discovers what the AI can do and never stops.

Step 1: connect

1. 01In Claude (web, desktop or Cursor) or in ChatGPT, add a custom connector with the platform’s MCP server URL.
2. 02Authorize with your account. The consent screen shows the scopes — data read, resource write — and you can pick read-only.
3. 03Choose the active project. The token is bound to it: the AI cannot switch projects on its own, and reconnecting resets the choice.

A detail that confuses people: after a connector update, the list of tools the AI sees is cached. If a new tool does not show up, disconnect and reconnect — and pick the project again.

Step 2: discover what exists

Before any business question, an inventory question: “which sources and tables can you see in this project?”. The AI lists the connected sources (Hotmart, Meta Ads, Google Ads, pixel, CRM), each one’s tables and columns. That calibrates the rest: you stop asking about data that does not exist and start asking about what is there.

Step 3: the first ten questions

1. 01How many approved sales did I have in the last 7 days, per product? (one source, verifiable in the Hotmart dashboard)
2. 02How much did I spend on Meta and Google in the same 7 days? (two sources, verifiable in the managers)
3. 03What was the cost per real sale per channel? (the first question no dashboard answers)
4. 04How many new leads came in through the site and how many have an e-mail? (measures form capture)
5. 05What are the 5 ads with the most spend and zero CRM contacts? (paid media + CRM)
6. 06How many people opened checkout and did not buy within 7 days? (pixel + Hotmart)
7. 07What is the average time between first click and purchase? (the ruler of the attribution window)
8. 08How many of this month’s buyers were already students of another product? (identity)
9. 09How many subscriptions have an overdue installment? (Hotmart)
10. 10Show me the timeline of the lead with e-mail X. (the profile, in text)

The first three have answers you can check elsewhere. Check them. That is how you learn to trust — or distrust — before asking what nobody else answers.

Step 4: check and correct

The AI shows the query it used. Read the rule: did it filter approved status? Remove renewals? Convert currency? When the rule is wrong, say so — “renewals do not count” — and the next answer already comes right. What we learned about checking the AI’s answer applies here: the number without the rule is worthless; the number with the rule is worth something even when you disagree with it.

Step 5: turn a question into a routine

The question you repeat every week should not be repeated: ask the AI to publish it as a report that updates itself, with a link for the team. And the question whose answer demands action — “who abandoned checkout” — can become an audience, created by the same connector with a simulation before publishing. Then the AI stops answering and starts configuring; it is the shift we describe in when the AI stops answering and starts configuring.

In CrazyLeads the MCP connector is part of the product: a single URL, authorization with scope choice, project locked in the token, over a hundred read and configuration tools, and authoring guides the AI itself reads before creating an audience, a delivery, a journey or a report.