A client called me in a panic three months ago. Their Performance Max campaign was reporting a 6.8x ROAS. Best number they'd ever seen. They wanted to double the budget immediately.
I asked one question before we touched anything: "What was your branded search ROAS doing before PMax launched?"
Nobody had checked. We pulled it. Branded search conversions had dropped 40% in the same window PMax "grew." PMax wasn't creating new customers — it was intercepting people who were already about to buy through branded search and taking credit for the sale. Same customers, same revenue, reassigned to a shinier-looking campaign.
We didn't double the budget. We turned on the New Customer Acquisition goal, excluded branded terms, and re-measured. Real incremental ROAS was 2.1x, not 6.8x. Still profitable — just not the miracle the dashboard was selling.
That's the story I keep coming back to when people ask me if AI has "changed everything" in paid media. It has. Just not always in the direction the platforms want you to believe.
Where AI Is Actually Earning Its Keep
I used to spend two hours writing 15 headline variations for a single ad set. Now I feed ChatGPT or Claude the offer, the audience pain point, and three examples of copy that's converted before, and I get 30 usable variations in ten minutes. Not final copy — first drafts I then cut down and edit myself.
The time saved isn't the point. The point is I can now test 12 headline angles in the same week I used to test 4, because writing them stopped being the bottleneck.
On accounts doing 100+ conversions a week, Meta's Advantage+ and Google's broad match with smart bidding genuinely find pockets of buyers a manually built audience would never surface. I've seen Advantage+ shopping campaigns pull CAC down 15–20% below a manually segmented equivalent once the algorithm has real volume to learn from.
The keyword there is volume. This only works when there's enough conversion data feeding it weekly. Below that threshold, you're not getting AI optimization — you're getting expensive guessing.
I manage several accounts at once. Pulling CAC, ROAS, and frequency trends across all of them used to eat half a Monday. Now I export the raw data and have a model summarize what changed week over week and flag anomalies — creative fatigue, a sudden CPM spike, a budget pacing issue. I still verify every flag myself before acting on it, but the first pass of "what actually needs my attention today" takes 15 minutes instead of 3 hours.
Where It's Quietly Making Things Worse
This is the big one, and the one almost nobody checks for. Performance Max and Advantage+ Shopping are built to chase any conversion signal available — including branded search, retargeting-warm traffic, and direct visits that would have converted anyway. The platform reports these as new wins. They're often just relabeled sales.
When five competitors in the same category all prompt ChatGPT with "write a Facebook ad for a skincare brand," you get five ads with suspiciously similar structure — a hook question, three benefit bullets, an urgency close. I can spot AI-generated, unedited ad copy in a scroll now, and so can your audience, even if they can't articulate why it feels flat.
The accounts I've seen actually decline in CTR over six months are the ones running AI copy straight out of the tool with no rewrite pass.
Ask Google's support team exactly why Performance Max shifted 30% of budget from Asset Group A to Asset Group B this week. You won't get a real answer. The system optimized toward its own signal, and that signal isn't fully visible to you.
This is fine when performance is good. It's a real problem when performance drops and you have no lever to pull except "wait and see" or "turn it off and rebuild."
I audited a startup spending ₹1.2 lakh a month, running full Advantage+ campaigns because "that's what Meta recommends by default now." They had 9 conversions a week. That's nowhere near enough for the algorithm to learn anything — it was essentially spending randomly and calling it optimization.
We switched to manual campaign structure with tighter, smaller audiences until they hit consistent weekly conversion volume. CAC dropped 45% in the first month, simply because we stopped paying the "still learning" tax on a system that never had enough signal to leave learning phase in the first place.
The Rule I Use Now
AI tools are good at generating volume and finding patterns in large, clean datasets. They're bad at judgment calls in ambiguous, low-data situations, and they're bad at telling you when their own optimization is actually just moving credit around instead of creating it.
Concretely: let AI write first drafts, summarize data, and handle bidding once you have real volume. Keep a human checking incrementality, auditing where automated campaigns pull their "wins" from, and doing the final edit pass on anything that goes in front of a customer.
✅ Branded search performance checked before and after launching any automated campaign
✅ New Customer Acquisition goal enabled and reviewed separately from blended ROAS
✅ No AI ad copy running without a human edit pass
✅ At least 15–20% of budget in a manually controlled, auditable campaign
✅ Automated bidding only turned on above a real weekly conversion threshold, not by default
✅ Weekly reporting anomalies verified manually before acting on them
Pull your PMax or Advantage+ numbers this week and check them against branded search from the same period. Most people are shocked by what they find.
Next up, I want to get into retention and lifecycle marketing properly — the emails, flows, and win-back sequences that actually keep the customers you've already paid to acquire, instead of losing them right after the first purchase.
See you there.
— Suraj
Frequently Asked Questions
Often, yes. PMax pulls inventory across Search, Display, YouTube, and Shopping using a shared signal, and it will happily claim credit for conversions that would have happened anyway through branded search or direct traffic. Excluding brand terms and comparing branded search performance before and after launch is the only reliable way to check.
As a rough floor, I want to see at least 30–50 conversions a week per campaign before trusting fully automated bidding. Below that, the algorithm doesn't have enough signal and you're often paying a premium for a system that's still guessing.
As a first draft, yes. As final copy, no. Unedited AI copy tends to converge on the same structural patterns across brands and categories, which readers register as generic even when they can't say exactly why. Always add back a specific, brand-true detail before publishing.
No — when there's enough data volume, they genuinely outperform manual campaigns on efficiency. The fix isn't turning them off, it's auditing what they're actually doing: checking incrementality, keeping a manual campaign running in parallel as a baseline, and not handing them 100% of budget by default.
Comments
Post a Comment