How to set up an AI marketing stack in an afternoon

What does "set up an AI marketing stack" actually mean?

An AI marketing stack is a pre-built harness that connects your marketing accounts — Google Ads, Meta Ads, GA4, Search Console and more — to an AI that can read the data, recommend a change, and then ship and measure it. The point is speed: you are not building an integration pipeline from scratch, you are plugging accounts into something that already works. If you are still weighing build versus buy, the build-vs-buy marketing harness guide covers the background. Here we get it running.

How fast is it really?

Honest split, because not everything is a two-minute click:

• Truly quick (about 10–15 minutes): create the workspace and connect your main ad accounts and GA4. Each connection is a browser OAuth handshake.
• Needs a little setup time (30–60 minutes): deciding what to measure first and what your AI may change. This is the part people skip and then regret, because it decides whether the verdicts later mean anything.
• Everything after that: asking what to fix first, shipping one change, and reading the result. That is the loop that runs long after the afternoon.

The promise is not "zero setup". It is "fast time-to-value": you see a first recommendation the same day, not after a three-week build.

Step-by-step: the one-afternoon setup

1. Create the workspace and connect your accounts (10–15 minutes)

Create a workspace in Cogny. Then connect the platforms you actually use. Start with the channel that spends the most, because that is where a wrong budget or a wasted keyword costs you real money fast. Google Ads and Meta Ads each take a couple of minutes. Add GA4 and Search Console so the AI can see the whole journey, not just the clicks.

2. Decide what to measure first (30–45 minutes)

Your changes only mean something if you measure the right number. Pick the one metric that determines whether the business is working — cost per conversion, ROAS, or organic visits that turn into revenue. Keep it small: one primary number. The AI proposes next steps against that number, ships the change you approve, and reads the result a few days later. If you measure everything, you measure nothing.

3. Grant your AI the tools it may use (15 minutes)

The harness can only act with the tools you grant it. This is the safety and the honesty control in one: it writes only where you say it may, every write is logged with who authorised it, and it cannot use a tool you did not grant. Choose the accounts and actions you are comfortable with. Block the rest.

4. Ask what to fix first (one chat)

Ask: "What should I fix first?" You get one specific next step, with the numbers behind it — not a wall of generic advice. A budget shift, a paused keyword, a title rewrite. You approve it and it ships.

5. Read the verdict (three days later for ads, seven for content and SEO)

Every shipped change becomes a ticket. Three days later for ads, seven for content and SEO, Cogny compares the periods in your own data and closes the ticket as worked, partial, didn't work, or inconclusive. Misses are included. Nothing is claimed as a win without the numbers to back it, and a change that did not help stays on the record as did not work.

What is the honest catch?

A measured setup makes for boring, reliable progress rather than a magic-feeling demo. The first change is usually small — a keyword pause, a budget trim — because the AI starts conservatively until it has a baseline. If your accounts are messy, the first week is partly cleanup. And if a change does not move the number, the verdict says so: the whole point of the harness is that you learn what does not work, not just what does.

Can you set this up from your own AI, not the app?

Yes. Connect Cogny over MCP and run the same steps from Claude, ChatGPT, Cursor or Codex. The harness and the ledger are the same, and your own AI client asks before any write. This is the setup most builders end up liking: the stack is already connected, and they work from the assistant they already use.

When should you not use this approach?

If your only goal is a one-off analysis of an exported CSV, you do not need a stack at all — a chat window is fine. A stack earns its keep when you want ongoing reads of live data, scheduled reruns, and a measured record of what you shipped. If you do not yet know which two or three numbers matter to your business, sort that out first; the harness will not tell you your business goal. It will, however, measure whatever you pick.

Ready to spend the afternoon on it: Start free — no card required. By the time the first verdict is in, you will know whether a pre-built harness gets you more speed than building the same thing from scratch.