Vibe Code Your
Ad Reporting Stack

I asked the question I always ask, because most founders never get a straight answer to it: why hand money to Mark Zuckerberg at all? Why not do the free thing?

They did the free thing. For three years.

Here is the part worth stealing, and it is not “paid beats organic.” It is why paid beat organic for them. Not reach. Not cost. Feedback.

Organic gave them an outcome. Paid gave them an outcome plus the reason for it. One of those you can build a machine on top of. The other you cannot, which is the whole argument of this article arriving eight minutes into the conversation.

Marcus came at it from a different angle — the old one.

I put my favourite bit of marketing folklore to them — traffic is like sex, only losers pay for it— and Marcus answered without a pause: “then I think we are the biggest losers.”

Two people running a profitable business in ten countries, happily accepting the label. If you want the honest version of the channel debate, that is it. Paid is not a compromise here. It is the channel that told them the truth.

If you have never mapped your own channels honestly, the traffic channel compass in the Codex is the same exercise, done on paper first.

Every direct-to-consumer brand runs roughly the same three tools: Shopify, Meta, and Triple Whale sitting in the middle doing cross-channel attribution. Paul called it the basic stack. It is what you install because everyone installs it.

They ripped the middle one out and rebuilt it themselves.

The reason is the single most useful sentence in the episode, and it is true for your business too:

So they compute the number twice. Meta says a campaign returned 3.5. Then they take the actual units sold — that product, that country, that window — straight from Shopify, divide by the ad spend, and get the real one.

“Because this is your actual truth.”

One campaign, one product, one country. That constraint is what makes the second number computable at all, and it is the piece most people skip. If your campaigns each sell five products across four countries, no dashboard on earth can untangle them.

Shopify comes in through a Shopify app they built and never listed — you can keep one private to your own store. I asked what it was called.

Meta comes in through the Meta Ads MCP, or the Graph API directly. Marcus's verdict on how much backend work the rebuild actually took is the part that should make you uncomfortable if you are paying a monthly subscription for this:

He is fair about the trade. They gave up the tracking pixel, so they gave up some visibility — and he is explicit that when you are still experimenting, still testing new angles and new products, a tool like that earns its money. They only cut it because after three years they already knew which channels work.

Buy the dashboard while you are learning. Own it once you know what you are looking at. That is a better rule than either “never pay for tools” or “always buy the standard stack,” and it only shows up if you ask people who have done both.

The Codex chapters on what to measure and how to report it are the same decision, one level up.

Here is the whole machine, described by Paul in about fifteen seconds. No database. No BI tool. No engineer.

Three skills and a schedule. That is the entire replacement for the tool their industry treats as mandatory. There is no stored history, either — no warehouse quietly accumulating rows. It runs, it looks at where things stand right now, it tells them what to change.

We are at a genuinely strange moment in software here, and it is worth naming: the app is now the harness, and all you do is connect it to your data. As I put it to them — give it the arms and the legs.

But a schedule is not the interesting part. This is:

Read that path again, because it is the most copyable thing in this article and it costs nothing to start:

Judgment you already have → spoken out loud → a document → a skill.

He did not sit down to architect an agent. He talked about how he decides, had the model tidy it up, and kept sharpening it until it was executable. The rules existed in his head for years. The only new thing is that they are now written somewhere a machine can read.

Which means the ceiling on your automation is not your technical skill. It is how much of your own judgment you have bothered to write down.

The Codex covers campaign structure and explore and scale — if you want the rules before you write them down, start there.

So what has actually changed in how they spend a working day? Marcus gave the answer that made me want to publish this conversation:

Now — the honest part, and the reason I trust the rest of it.

Neither of them has gone fully automatic. Paul lets the agent create campaigns: he says which ads, and it goes. But he still holds its hand on the thing you would most expect to be automated — moving budgets up and down day to day.

That is the cleanest definition of the approval line I have heard from anyone. The leash extends exactly as far as the rules are complete. Not as far as the model is capable, and not as far as your nerve holds — as far as you have written it down. Where there are blank spots, a human stays in the loop, and the job is to close the spots, not to summon more courage.

Marcus keeps his hand in for a different reason, and he is refreshingly blunt about the contradiction:

He admits it makes fewer mistakes than he does — and still supervises every big step. That is not irrational. The average is better; the tail is worse, and the tail is what takes your shop down.

Paul draws the line in the most concrete way possible, and it is a good test for anyone deciding how much rope to give:

Automations that append to a Notion log can be wrong occasionally. Checkout cannot. The leash is not one setting — it is per system, and the question is always the same: what does it cost when this is wrong?

Ask most people how AI makes ads and you get: describe your product, get a video. That is not how it works here, and the difference is where all the value sits.

They start in the Meta Ad Library, looking at competitors' creatives that have been running for a long time — because an ad that has run for three years is an ad that is working.

Then comes the step that everyone building an “AI creative pipeline” gets wrong. They do not hand the winning ad to a model and ask what makes it good.

So the humans watch the ad and break it into hypotheses — the hook, the angle, the lighting, the person, the music, the voiceover. Each one a guess at why it works.

That is the whole division of labour, and it is the opposite of the one people assume. The machine does not supply the judgment and leave you the typing. The machine supplies volume. You supply the reason to point it somewhere.

Only then does the automation start:

• An engine turns one idea into a very long, very detailed prompt for a single video scene.
• That prompt generates in bulk — around ten videos, not one — currently on Sora 2, though they swap models whenever something better lands.
• A human picks the one that survived.
• Most finished ads are several scenes stitched together, not one clip.

The stitching is the clever bit. They use Creatomate, which lets you build a video template and then export it as JSON.

A video becomes a data structure. Once it is JSON, an agent can vary any slot in it — swap the hook, swap the voiceover, swap the headline — without a human touching an editor. That is how you get from one ad to a hundred variants of one ad.

They drive it two ways, and the reason for the first one is a good lesson in itself: a plain HTML page with input fields wired to n8n workflows, built “just to save on context, to save on tokens” because the process is so repeatable. When something runs the same way every time, it does not need an agent thinking about it. Marcus's alternative, for when they do want to experiment with cutting times and voiceovers, is Codex skills rendering the videos locally — voiceover, captions, the whole ad from a single prompt.

The ad creatives chapter is where we keep our own version of this.

I pitched Paul an idea I have been circling for a while: point an agent at the highest-performing videos in your niche, pull them down on a schedule, slice out the hooks, and let validated data suggest your next creative. Would it work?

His answer reframed the question, and it is the sentence I would put on the wall:

A hook that prints for one product, in one market, for one audience may do nothing for the next. So stop trying to be right up front. Build the thing that can be wrong cheaply, often, and on the record.

I told them where I am still stuck, because I am: I still think in straight lines. Campaign idea, make creatives, launch, optimise, end of line. Not a closed loop that keeps going on its own — generating the next batch, reading early signals like click-through rate, pushing further in whatever direction is working, exploring a little at the edges, round and round.

Paul's answer was the most honest moment in the hour. They already run a version of it — prove an angle in one market, then replicate the exact same ad, not roughly the same, into another language for another market. And then he graded it himself:

Nobody in this conversation has the closed loop. Three people who do this every day, agreeing it is the right direction and that none of us has got there yet. I would rather publish that than a tidy answer.

If you want to run the disciplined version of this while the loop is still manual, ad testing is the chapter.

I asked Marcus to expand on something he had said to me before we recorded, because I think it is the most important observation either of them made:

He does not claim that gap is permanent. He said so plainly: in twelve or twenty-four months the twenty percent might shrink to five, and then everyone can do what they are doing. Nobody knows.

Which is exactly why his conclusion is the right one — “it's worth staying at the edge of what's possible.”

And notice what this whole conversation was actually about. Two marketers learned to build, and the first thing they built was not the product. It was the machine that sells it. Every piece of this article — owning the numbers, hardening judgment into skills, the leash, the variant factory — is a marketer's instinct that finally got arms and legs.

Four moves, if you want the compressed version: own your numbers, because Meta is telling you seventy percent of the truth. Write your judgment down until it becomes a skill. Extend the leash only as far as the rules are complete. And spend your human hours on the why, never on the volume.

One ask before you go — theirs and mine. We are trying to get a small group together of people who are deep in both AI and performance marketing: analytics, creative, landing pages, the whole conversion experience. Nothing paid, nothing organised, just practitioners comparing notes. There are not many of us yet. If that is you, say so.

Paul and Marcus are on stage at Vibe Coding Cologne on 2 September — talking about exactly this, with a laptop open and the real dashboard on screen. Free, at STARTPLATZ. Come and ask them the question I didn't.