#032 | I gave Claude access to all our marketing data

My team and I built something pretty cool this summer.

It's a data mart that aggregates from our main ad platforms (Google, Bing, and Meta), Salesforce, and Gong into Supabase, which we have connected to Claude through MCP.

If "data mart" and "MCP" mean nothing to you, that's totally fine. Here's the short version: a data mart is one central spot where data from a bunch of different tools gets stored together, so it can actually be compared.

MCP is the connector that lets Claude query that data directly, the same way you'd ask a coworker.

GitHub runs python scripts daily to keep this dataset fresh, and we've backfilled nearly 13 months worth of historical data.

Let me break it down!

• Google Ads, Bing Ads, Meta: what we spent and which ads ran

• Salesforce: which leads became pipeline and revenue

• Gong: what prospects actually said on sales calls

• GitHub Actions: runs the scripts every day so nobody exports anything

• Supabase: the database where it all lives, connected together

• Claude: sits on top, so anyone on the team can ask questions in plain English

It feels like magic

It feels like magic when you can ask in plain text something like…

Or even questions outside of paid like…

And get actionable outputs for the team.

Why this matters

The cost of a question drops to near zero. When an answer takes two weeks, you ask maybe two questions a month. When it takes two minutes, you ask twenty. And you start asking the follow-up questions too, which is usually where the real insight is.

Everyone works from the same numbers. Marketing, sales, and leadership query the same connected dataset, so the conversation starts at "what do we do about this?" instead of "where did that number come from?"

Your sales calls become a research asset. Hundreds of hours of customer conversations used to sit in Gong, listened to once. Now they inform ad copy, content, and positioning every week.

Decisions get made on evidence. When the data is one question away, there's no reason to go with a hunch.

How this can be applied with other data

This setup works well beyond marketing. Every version of it comes down to the same five steps:

1. Pick the sources that, if combined, would produce your answer

2. Build a pipeline that pulls the data from each source

3. Schedule it to run on its own (daily, hourly, whatever the decision needs)

4. Store it all in one central database, structured so it connects

5. Connect Claude on top so anyone can ask questions in plain English

Change the sources and the questions change with them. A few examples:

• Churn early warning: billing, product usage, support tickets, renewal calls → "Which accounts renewing soon look most like last year's churned accounts?"

• "Who are we losing money on?": invoices, time tracking, project management, payroll → "Which clients look profitable on paper but aren't once delivery hours count?"

• Zombie software audit: company card spend, logins, renewal dates → "Which tools are we paying for that most seats haven't touched in 60 days?"

• Competitor radar: competitors' pricing pages, job postings, ad libraries, G2 reviews → "What is Competitor X starting to emphasize, and are they hiring for it?"

• Multi-location scorecard: point-of-sale data, Google reviews, staffing schedules, weather → "Why did Location 7 underperform compared to similar locations?"

Cheers,
Matt