AI in media buying: what it actually does, where it fails, and one test where human beat the machine

Andrei Rulko has been buying traffic for over eight years, first on search, then programmatic, now across push, native, Facebook, and Performance Max. He manages campaigns for affiliate teams and direct advertisers. We talked for an hour about AI tools, a split test he ran against a machine’s recommendation, and the campaign where the algorithm produced a perfect dashboard and zero actual sales.

Spoiler: the algorithm was technically right the whole time. That’s exactly the problem.

Let’s get the basics out of the way. When you say “AI in media buying,” what does that mean on a Tuesday afternoon when you’re actually running campaigns?

Good framing. Because the way it gets talked about in conference decks and the way it actually shows up in your account are pretty different things. 

In practice it’s automated systems making real-time decisions about where to bid, how much to pay per impression, which creative to show to which person, and when to pull budget away from something that isn’t working. Not magic. Not a sentient robot. A system trained on historical data that’s trying to predict the next outcome.

You feel it most in three places:

• Bidding. The platform calculates a conversion probability for every impression and bids accordingly. You set the target, the machine handles the math.
• Targeting. AI clusters behavioral signals, builds lookalike audiences from your converters, and expands or contracts reach based on what’s actually working.
• Creative rotation. Your variants go in, performance gets tracked per audience segment, and underperformers get quietly deprioritized. Sometimes before you’ve even noticed they weren’t pulling.

For affiliates specifically, there’s another layer worth understanding. Push networks, native platforms, Facebook, they all run their own AI underneath whatever campaign structure you’ve built. You’re not just buying traffic anymore. You’re configuring inputs for an algorithm, and the quality of those inputs is now most of the job.

When did it become impossible to ignore?

Around 2022 for me. Before that you could still build a real edge through manual work — tighter negative keywords, sharper bid adjustments by hour, better audience exclusions. I genuinely believed I was outsmarting the platform on some accounts.

Then I started noticing something uncomfortable. The accounts where I’d loosened my grip and let automation run were producing better CPAs than the ones I was actively managing. That took a while to sit with. But it was consistent enough that I had to accept it.

The platforms got too good at base-level execution. The advantage shifted. You’re not paid anymore to do what the algorithm does better. You’re paid to do what it can’t.

Walk me through a campaign. Where does the AI actually touch the process?

Basically everywhere now. See the honest picture:

Stage	What the AI is actually doing
Audience segmentation	Clusters behavioral data, builds lookalikes from your converter list
Bidding	Scores each impression in real time, bids based on predicted conversion probability
Creative rotation	Tests variants, shifts spend toward top performers per audience segment
Budget pacing	Reallocates budget toward better-performing placements, often several times a day
Anomaly detection	Flags unusual performance changes in reporting

By the time you check in the next morning the campaign has usually already made several budget decisions without you. Which is mostly fine. The weak spot is still interpretation, the numbers surface fast, but figuring out why something changed is still yours to do.

📖 Read also: Nutra offer types explained, if you’re running nutra on paid traffic, this covers the offer structures you’ll actually encounter

What has AI made genuinely faster or cheaper?

Three things where I’ve actually felt the difference:

• Reporting. What used to take a few hours of pulling and formatting data is now minutes. That’s real time back in the week, across every account.
• Audience testing. Old approach: run one variant, wait two weeks, check results, launch the next one. Now you run multi-variant tests and the algorithm allocates spend toward winners in real time. The learning cycle is genuinely shorter.
• Bid management at scale. I don’t manually touch individual bids on most campaigns anymore. The platform is better at it than I am when you’re across hundreds of ad sets. I’d rather spend that time on strategy.

Where I’d push back is on the big round numbers you sometimes see quoted, “40% reduction in cost per lead” and so on. Those figures usually assume a baseline of pretty mediocre manual management. If you were already running tight accounts, the improvement is real but smaller. The biggest gains go to people who were doing the most repetitive, least-optimized work before.

You ran a split test where you put your own judgment against the machine’s recommendation

This is the one I get asked about most.

I’d been running one nutra offer across multiple GEOs for about six months. Real data from the tracker, from the auto-traffic platform, with conversions, with creative performance broken down by placement and audience. I collected all of it and fed it into an AI tool with one ask: give me a recommendation. Which placements to keep, which to cut, how to prioritize creatives, what the best GEO-offer combination looked like.

The machine thought about it and came back with a clean answer.

I almost just clicked OK. But something felt off – the AI was recommending I cut some placements that seemed wrong to me given what I knew about that specific offer’s context. So instead of following the recommendation, I ran a split. Three campaigns built exactly as the AI suggested. Three campaigns built the way I’d have done it with my own read of the same data.

My campaigns won. And after optimization, the gap grew.

The AI had picked the statistically safe path – the combinations that looked strongest on average across the full dataset. But it had deprioritized some placements that were underperforming in aggregate. What it couldn’t see was that those specific placements were the only ones converting on a particular sub-offer variant I knew had value for reasons that don’t live in a spreadsheet. The AI didn’t have that context. I had it because I’d been working that offer and those GEOs for months.

The takeaway is that AI is a very fast, very capable analyst who has never actually worked your account. It’ll process 1,000 combinations in a second. But which of those combinations fits your specific offer, your specific GEO quirks, your advertiser relationship, that’s still a human call.

Have you seen when AI fails completely on a campaign?

Oh, this one. Yeah.

Performance Max, goal set to leads, reasonable budget. I let it run. The algorithm did exactly what it was built to do: it found inventory, it found clicks, it found a very cheap-lead environment.

What it found was an MFA site. Made for advertising. Pages stuffed with AI-generated text about something vaguely financial, 15 ad units arranged around every article, built specifically to attract cheap programmatic spend. Cost per click: 3 cents. Leads started coming in. The dashboard looked beautiful. Low CPA, high volume, the kind of screenshot you’d normally want to send to a client.

Then I opened the CRM.

Zero. Not a few stragglers. Zero conversions. Not one of those “leads” turned into a sale. The people who had filled out the form had accidentally tapped a banner while trying to navigate a site they weren’t actually reading. They had no intent. They were collateral damage from cheap inventory.

And here’s the thing that still annoys me about this story: the AI wasn’t wrong. It had been told to find cheap leads. It found cheap leads. Task completed. It had no concept of what a lead is worth after it leaves the platform. It optimized exactly for the metric it was given and ignored the one that actually mattered.

The gap between “completed the platform KPI” and “served the actual business goal,” that gap is the job. That’s what you’re paid to manage. An algorithm will never care about your CRM. That’s not a limitation of this particular tool. It’s structural.

📖 Read also: Affiliate marketing in Kenya, one of INB’s newer GEOs, and a good example of what “emerging market” actually looks like when you break down the numbers

Where does human judgment still win outright?

Five buckets where I haven’t seen AI come close:

• Translating vague client goals into actual strategy. “We want more leads” can mean a dozen different things. Figuring out which one the client actually needs, and whether leads are even the right metric, is not a data question.
• Cultural and brand context. AI doesn’t know that a specific creative angle lands badly in one market, or that the tone of copy feels wrong to a real human in that context even though CTR looks fine.
• Intelligence that doesn’t live in a dashboard. When an affiliate manager tells you something off the record about what’s converting well this month, that doesn’t go into any model.
• Knowing when the machine is being misled. A tracking issue, a bot traffic problem, a seasonality effect outside the model’s training window, the AI optimizes enthusiastically toward the wrong thing and you only catch it if you’re close enough to notice something feels off.
• Accountability. When a campaign goes sideways and a client wants answers, they don’t call the algorithm.

What are the risks people underestimate most?

The MFA problem specifically isn’t rare or a one-off. It’s structural. Cheap programmatic inventory exists because publishers built it to exploit exactly the incentives that platform AI optimizes for. You need governance layers to stop it from going there, it won’t stop itself.

Risk	What it actually looks like in practice
Black-box opacity	Performance Max and Advantage+ produce results but hide the logic. When performance drops, there’s almost nothing concrete to debug.
Data quality	Bad tracking or misconfigured conversion events get optimized at machine speed. AI scales mistakes as fast as wins.
MFA and garbage inventory	Platforms optimize for cheap clicks. Some of the cheapest inventory in existence is designed specifically to game that incentive.
Over-automation	If you hand everything to the machine, you own the consequences without understanding the cause.
Misaligned KPIs	AI optimizes exactly for what you tell it. If that metric doesn’t reflect the real business goal, it succeeds at the wrong thing.

How do you actually maintain control without slowing every campaign down?

Governance runs alongside automation, not instead of it. In practice:

• Pre-bid controls: brand safety lists, inventory exclusions, content category blockers. These run before the algorithm makes any decision, so they don’t interfere with the optimization cycle at all.
• Regular placement audits, even when they’re buried in the interface. For Performance Max you have to push for them, they’re not surfaced prominently.
• Clear escalation thresholds. If CPA rises more than a defined percentage week over week without an obvious external explanation, I look manually before letting the campaign continue spending.
• Strategy stays off the machine entirely. What offer, which GEO, what creative direction, what the actual success metric is, those stay with me. The moment you let AI define its own objective, you’ve handed over the part that matters most.

The goal is to define the boundaries well enough that when the machine optimizes, it’s optimizing toward something real.

If someone is running an affiliate campaign with INB.bio, should they be using AI tools more, less, or differently than they are now?

More. Definitely more. But differently than most people use them.

The instinct a lot of buyers have is to either resist the automation (“I know my account better than a machine”) or hand everything over and check in once a week. Both lose.

The right approach is to use AI tools like you’d use a very fast, very hardworking analyst who just joined your team. Feed them good data. Give them a real objective — not just “find cheap leads” but something connected to what actually happens downstream. Check the CRM, not just the dashboard. And keep one part of the process where your judgment is what drives the outcome.

Not because the machine can’t handle it. Because that’s how you stay sharp enough to know when it’s wrong.

The buyers who will still be valuable in three years are the ones who understand both sides, who can read what an algorithm is doing, trust it when it’s right, and override it when it’s been given the wrong objective. That’s a skill. It’s also the one that doesn’t get automated away.