AI agent skill / Paid media analytics
Paid Media Intelligence
Meta, Google Ads, GA4 and the CRM rarely agree on what happened. I built a set of AI-agent skills that lines them up, checks the numbers, and answers in plain words: what changed, why, how sure it is, and what to do next.
- My contribution
- Concept, analysis method, playbooks and build
- Timeframe / scope
- Open-source skills for AI coding agents. Published October 2026.
- Current stage
- Version 1.0.0, pre-release. Checked on fictional data; not yet connected to live ad accounts.
At a glance
- Skills, from a connection check to a cross-platform investigation
- 8
- Fictional test cases run end to end
- 23
- It never changes an ad account, a campaign or a GA4 property
- Read-only
Where the idea started
Every platform tells its own story. Meta says leads are up. Google says conversions doubled. GA4 shows something else, and the sales records show something else again. Working out what really happened takes hours of exports and spreadsheets.
I wanted that checking to happen the same careful way every time, and I wanted the answer in plain words.
The approach
I treated it as a reconciliation problem before a reporting one. The skills bring Meta Ads, Google Ads, GA4 and your own orders or CRM records together, then check the things that usually go wrong: ad clicks against website visits, platform claims against your own records, changes over time, and whether two periods are even fair to compare.
Platform claims are shown side by side and never added up. Your own sales or CRM records are the reference. And it stays read-only: it never changes an ad account, a campaign or a GA4 property.
How I built it
It runs inside the AI agent you already use, as a set of skills. A small Python core, standard library only, does the arithmetic, the checks and the charts. The agent does the reasoning, guided by playbooks.
You ask a normal question, like "Why did our cost per lead rise this month?" You get a short answer, a self-contained HTML report, and an Excel file where every chart sits above its data table. Change a number and the chart follows.
Claims the evidence can't support are blocked. It won't tell you "this proves fatigue", and it won't tell you to raise the budget by 20%.
Where it stands
It's installed and run end to end in Claude Code on 23 fictional cases, from creative fatigue and lead quality to cross-platform UTM checks. The Codex skill files are generated but not yet tested.
Live Meta, Google Ads and GA4 accounts weren't connected while I built it. Every check so far used invented businesses and numbers, and the repository says so.
What I learned
Most reporting arguments are really about which numbers are allowed to sit next to each other. Writing those rules down, including the ones that end in "we can't know that yet", did more for the answers than any chart.
Tools used
- Python
- Claude Code
- Codex
- Meta Ads
- Google Ads
- GA4
- Excel