The AI Media Buyer: Can AI Run Your Facebook Ads?

Your ROAS is probably lying to you.

Not maliciously. Not because the platform is out to get you. It’s lying the way a scale lies when someone’s leaning a hand on it — the number is real, the measurement is rigged. And an ads agent that trusts that number will spend your money faster and more confidently than you ever could by hand. That is the whole danger, and it is also the whole opportunity.

Here is the thing about media buying: everyone wants the agent to be the gunslinger — the thing that scales the winners and kills the losers while you sleep. That is exactly the wrong first job. The first job is forensic accountant.

The Old Math

For years, the media buyer was a role you either hired or wore yourself. If you wore it yourself, it went like this: log into Ads Manager, stare at a wall of campaigns, trust the ROAS column, and make moves off it. Push budget at the green rows. Pause the red ones. Launch a new thing when growth stalled.

The problem is that the ROAS column is a story the platform tells you about itself. And every platform grades its own homework generously.

I have the receipts, and they’re mine. We pulled the full lifetime history of Rosebud Woman’s Meta account — account 537924426672937, “Rosebud Woman 10-15-18.” Meta only retains about 37 months, so this is 2023 forward; the years before that are gone. What Meta can still see is roughly $780K of lifetime spend across 70 campaigns, at a blended ROAS of about 1.3×.

1.3× isn’t a disaster. On a blended basis, with a brand that has real repeat purchase behavior, you could look at that and decide things are basically fine.

They were not basically fine.

What the Audit Actually Found

The autopsy took an afternoon. That’s the part I want you to sit with. Two agents, working in parallel through Meta’s official connection, pulled five years of an account no human had fully reconciled — every campaign, every objective, every dollar — saved it to file, and did the accounting. What would have been a week of a specialist’s life squinting at exports was done before lunch. This is the work agents should eat: precise, rule-bound, enormous, and soul-crushing to do by hand.

And here is what the blended number was hiding.

That flagship campaign — a $267K giant, targeting the top of the funnel with a conversions objective — showed 1.48× on Meta’s omni-attribution. On a pixel/website basis it was 0.72×. Underwater. Meta was counting sales it did not cause. The omni-attribution number was crediting the ads for purchases that would have happened anyway, from people who came in through email or Google or simply already knew us. That’s the mask: the blended figure looked fine precisely because it was absorbing demand the ads didn’t create.

Then the leaks. $21K went to a Traffic-objective campaign — optimizing for link clicks, not sales — and returned 0.03× ROAS. Read that again. Three cents back on the dollar. Another $5K went to an Awareness campaign that produced zero purchases. And the cruelest detail: the only currently active campaign on the account was one of those Awareness campaigns. Zero purchases, 0.13% click-through. The account had drifted into spending on the one objective guaranteed not to sell anything.

None of that is a stupid-person mistake. It’s an everyone mistake. It’s what happens when you’re one person wearing six roles and the ads column looks green enough to leave alone. Nobody had the afternoon to do the forensic pass. So it never got done.

That’s the precondition lesson, and it’s the spine of this whole chapter: an ads agent is only as honest as its measurement. If you point an optimizer at a broken conversion signal, it doesn’t fix anything. It just gets you to the wrong answer faster and with more conviction.

What Actually Worked

The audit wasn’t all wreckage. Buried in the same account was a clear picture of what worked — and it’s boring in the best way.

Broad targeting beat clever targeting. The best campaign at scale was “Broad Campaign–Active, Affluent,” a Sales-objective build that returned 1.80× on $68K. Not a lookalike stack, not a fifteen-interest Frankenstein. Broad audience, sales objective, let the machine find the buyers.

Creative was always the real lever. The single best creative the account ever ran was “March ‘25 Grayce Creative” at 2.07× and 2.82% CTR — and it was pulled deliberately, once. The best-performing product theme was Arouse at 1.66× on $26.8K. And the winning copy formula wasn’t a discount or a clever hook; it was problem-first plain language anchored to a named, real review. The top ad by click-through — 4.06% on 97K impressions — said “Say Goodbye to Intimate Dryness!” and quoted a customer by name: “Pain Relief in a Jar!… on my third jar.” — Angela L.

Retargeting quietly did its job. The warm, middle-of-funnel bucket returned about 1.27× on $124K — the steadiest bucket in the account after the winners above. And one number reframes all of these: Rosebud has roughly a 38% repeat rate, which means a first-order return of 1.7–1.8× is genuinely profitable once you count the lifetime value behind it. The bar is lower than the dashboard makes it look — but you only get to use that math if the underlying attribution is honest.

So the pattern the agent surfaced was legible: broad targeting, review-anchored creative, and retargeting worked. Traffic and Awareness objectives burned money. Manual over-segmentation split the learning across 70 fragmented campaigns. That is not a hunch. That is what an afternoon of honest accounting produces when nobody’s leaning on the scale.

What the Seat Senses, Drafts, and Never Does

So what does a media-buyer seat actually do, once it’s built? Three verbs, and the third one is a wall.

It senses. Spend, ROAS, CPA, frequency — pulled straight from the account, dated, no zero-filling a gap into a fake zero. And before it trusts any of it, it checks the conversion signal itself. A ROAS built on a broken pixel isn’t a yellow flag to the seat; it’s a hard stop. It won’t optimize on a number it can’t trust. Signal health is a precondition, not a footnote.

It drafts. Budget shifts, pauses, new campaign structures, killing the dead Awareness campaign, moving spend off Traffic objectives it will never recommend again. It writes the move and the reasoning and puts it in front of you.

It never spends on its own. This is not a nice-to-have; it is the architecture. Every new campaign the seat creates arrives paused. Every budget move, every launch, every dollar decision routes through one approval queue — the same gate every other seat uses — and waits for a human. The campaign budget itself is the spend cap; the seat cannot exceed what you approved. There is no autonomy rung on this seat where money moves without a person saying yes. The seat’s strongest possible output is a well-reasoned card sitting in your queue.

I’ll say it plainly, because it’s the load-bearing rule of the whole book: the agent that can spend money without asking is not a media buyer. It’s a liability with API access.

The Policy Lens (Or: The Foria Lesson)

There’s a second wall, and it’s specific to regulated categories — wellness, supplements, CBD, intimate health. Whatever you sell that a platform has an opinion about.

Meta’s rule for sexual and reproductive wellness is not what most people assume. It is not “these products are banned.” It is this: an ad is permitted when the focus is health or medical efficacy, and prohibited when the focus is pleasure, arousal, or enhancement. Same product. Two ads. Opposite outcomes. The line is drawn on the claim in the copy, not on the product itself.

This is why Rosebud ran intimate-wellness ads for years — Honor to over 570K impressions — without the account being banned. “Dryness,” “menopause,” “vulva moisturizer,” “comfort,” “clinically tested” all sit on the permitted side. It’s the specific pleasure and arousal claims that get individual ads rejected, not the brand that gets erased.

Watch how a serious competitor threads it. Foria — the closest rival, running well over a hundred ad variants — sells a product literally called Awaken Arousal Oil. In their ads, it’s just “Awaken Oil.” The arousal claim lives on the product page, where the platform doesn’t gate it, and stays out of the ad copy, where it would. That’s not a loophole. That’s understanding exactly where the line is and building on the legal side of it.

So the seat carries a policy screen. When it drafts copy, it checks its own words against the permitted and restricted lists — dryness and menopause and comfort on one side; arousal, pleasure, enhancement, libido on the other. A restricted match does not get auto-reworded. The seat does not quietly soften “arousal” into something that’ll clear review and ship it. It flags it and routes it to a human to decide the reframe — the same posture a compliance officer would take. A machine should never be the thing that decides how to word a regulated health claim to a regulated audience. That judgment stays with a person, every time.

The Principle, Named

The media-buyer seat’s first job is forensic accountant, not gunslinger.

That’s the coinage, and it’s the whole inversion. Everyone reaches for the agent as a scaling machine — point it at the account, tell it to grow the winners. But an optimizer on top of dishonest measurement is worse than no optimizer at all, because it launders your bad numbers into confident action. The order of operations is fixed: audit before optimize. Signal health before ROAS. Structure fixes before scale. And every dollar through the gate.

The forensic pass is also the cheapest, safest, highest-value thing the seat will ever do. It touches no money. It creates nothing. It just tells you the truth about what you already spent — and on my account, the truth was an afternoon away the whole time, worth a specialist’s week, waiting for someone to do the accounting.

So What

If you’re running ads and thinking about pointing an agent at them, do it in this order.

First, make it audit before it touches anything. Have it pull your lifetime spend and objectives and find the leaks — the Traffic campaigns, the Awareness campaigns, the fragmentation. You will almost certainly find money that went nowhere. I did.

Second, make it check your conversion tracking before you trust a single ROAS figure. Ask for the pixel-basis number, not just the blended one. If they disagree the way mine did — 1.3× blended, 0.72× on the flagship’s pixel basis — believe the worse one until a lift test tells you otherwise.

Third, wire the gate before you wire the automation. New campaigns paused. Every dollar move in one queue. No rung where money moves without you. If a tool won’t let you do it that way, that tool is not built for a small team that can’t afford to be wrong.

And if you sell something regulated, make the policy screen the agent’s reflex, not your afterthought. Health claims clear; pleasure claims don’t; the machine flags and the human decides. Always that order.

The media buyer you couldn’t afford to hire is now something you can staff in an afternoon. But the first afternoon isn’t for scaling. It’s for finding out where the money actually went. Start there. The truth is cheaper than you think, and it’s been waiting the whole time.