How AI is quietly rewriting the economics of media buying

For two decades, the binding constraint in performance marketing was the same: making the ad. Generative AI has quietly demolished that constraint — a media buyer can now spin up a hundred angles before lunch. The uncomfortable part is what the cheap-creative era reveals. When everyone can produce infinite variations, production stops being the moat. Judgment becomes the whole game.

The bottleneck moved, and most buyers haven't noticed

Think about where the money and the time used to go. Briefing a designer. Waiting on a video editor. Paying for stock, for a shoot, for a round of revisions that came back wrong. The cost of a single tested creative was measured in days and dollars, which is exactly why most accounts ran three to five variations and prayed. Creative volume was rationed because creative was expensive.

That economics is gone. Meta has rolled generative tools across Advantage+ — including video-generation features that turn a handful of product images into multi-scene video with text and music — and has been shipping AI-driven ad features at pace. Google has built the same machine into its Asset Studio: Imagen-powered image generation, lifestyle compositions from a single product shot, batch edits across many images at once, and video generated from static inputs. The platforms will now generate text, image, and video assets from your website URL in seconds.

The marginal cost of one more creative has collapsed toward zero. That single fact rewrites everything downstream. When the bottleneck moves, the source of advantage moves with it — and a lot of operators are still optimizing the part of the funnel that stopped mattering.

Why creative was always the real lever anyway

This shift lands harder because of where the auctions already were. Meta and Google spent years absorbing the decisions buyers used to make by hand. Targeting, placement, bid management, audience construction — the algorithm took the wheel. Broad targeting paired with a large creative library now routinely beats the old craft of stacked interest sets and surgical ad-set structures.

What was left for humans to control? Creative. Meta's own framing puts expected value as roughly bid × estimated action rate × ad quality, and creative is the input that most moves both action rate and quality — which is to say, it's a bigger lever than your bid. Industry analyses citing platform data science consistently land in the same neighborhood: a large share of auction outcomes trace back to the creative itself. Reach, brand, and targeting split the rest.

So the picture is this: creative was already the dominant lever, and AI just made the raw material of that lever nearly free. That's not a small efficiency gain. It's a change in what skill even means.

What got cheap, what got expensive

It helps to be blunt about the ledger.

• Got cheap: the act of producing a creative. Concepting visuals, generating variants, localizing into ten languages, re-cutting a winner into nine aspect ratios. The things that used to fill a production calendar.
• Stayed expensive — or got more expensive: knowing which creative to make in the first place, recognizing a winner before the data is statistically clean, and resisting the urge to keep feeding a loser because you like it. Selection. Hypotheses. The discipline to kill.

Here's the trap. Infinite cheap creative doesn't reduce the workload — it inverts it. You're no longer rate-limited by production, so the constraint becomes your ability to evaluate. A team that can generate 200 variations but can only confidently judge five of them has not gained 40x leverage. It has gained a backlog and a budget-burning machine. The expensive thing is now attention, and attention does not scale the way a video model does.

The new job description: feed the machine, then judge it

The clearest way to see the shift is in how the human role has changed. The old job was to steer the buy — pick sites, times, placements, match creative to audience. The new job, as the platforms themselves now frame it, is to provide better fuel: stronger signal and a richer suite of multimodal assets, then let the system drive. The buyer moves from operator to portfolio manager.

That reframes the daily work around four things AI can't do for you:

1. Hypotheses. AI generates executions, not insights. "What if the angle is fear of overpaying, not desire to save?" is a human bet. The machine will render that bet beautifully in fifty formats — but it won't have the bet. Your edge is the quality and originality of the hypotheses you feed in.
2. Selection under uncertainty. A hook is showing a few points better CTR on day two with thin volume. Promising winner, or noise? Calling that early — before the data is clean enough to be safe — is the skill that now separates accounts.
3. Reading signal vs. attribution theater. When a meaningful share of affiliate-attributed conversions may not be truly incremental, and coupon and cashback placements often capture demand that was already going to convert, the buyer who measures lift instead of last-click is buying real growth while everyone else buys credit for it.
4. Knowing when to stop. Creative fatigue at real spend tends to set in within one to two weeks. The cheap-creative era doesn't eliminate fatigue — it accelerates the cadence of refresh-and-replace. Discipline about killing tired winners becomes a core competency, not an afterthought.

Commoditization: what stops being a moat

The hard truth for anyone who built a business on production capacity: the ability to make good-looking ads fast is no longer defensible. Everyone has the same image models, the same video models, the same in-platform generators. More capable tools don't just lower the floor — they raise the average, which makes the median ad better and the median advantage smaller.

There's a deeper convergence risk too. These tools are trained on similar historical performance data drawn from the same market. Point similar models at similar inputs and they drift toward the same messaging, the same formats, the same "winning" template. You end up with technically proficient creative that is indistinguishable from your three closest competitors. Taste, as the cynics note, also commoditizes — once everyone clears the bar, clearing it no longer makes you special.

So if production isn't the moat, what is?

What becomes a moat instead

The defensible assets in a cheap-creative world are the ones AI consumes but cannot manufacture:

• Proprietary data and signal. The platforms optimize on what you feed them. A clean first-party event stream, well-defined conversion signals, and accurate value-back data is fuel no competitor can copy. As the models commoditize, your data is what makes your instance of the same model perform differently.
• Measurement and incrementality. If you can prove lift while competitors argue about last-click, you can scale spend with a confidence they don't have. Incrementality stops being a finance-team nicety and becomes a buying weapon.
• The offer and the post-click. AI makes the ad cheap; it does not make your offer better than the next operator's, or your landing page convert. When creative is commoditized, the funnel below the click — offer strength, page experience, pricing, the post-click promise — becomes a larger share of the edge.
• Speed of iteration as a system. Not raw output speed — everyone has that now — but the velocity of the full loop: hypothesis → generate → launch → read → kill → re-hypothesize. The operator who closes that loop in a day beats the one who closes it in a week, regardless of who has the prettier ads.
• Judgment as institutional memory. The pattern library in a senior buyer's head — which angles work for which offers, what fatigue looks like before the metrics confirm it — is the asset that compounds. It's the one thing on this list that gets more valuable as creative gets cheaper.

The buyers who win the next cycle

It would be easy to read all this as a threat to media buyers. It's the opposite — but only for a specific kind of buyer. The era that rewarded whoever could produce the most creative the fastest is over; the platforms gave that superpower to everyone for free. The era that's beginning rewards whoever can decide the best, fastest, on the thinnest evidence, with the cleanest signal.

That's a harder skill to fake and a harder one to commoditize. Production was a cost you could outsource. Judgment is the one part of the job that was always the actual job — AI just stripped away everything that used to disguise it. The operators who internalize that, and rebuild their workflow around selection and measurement instead of output, won't be replaced by the machine. They'll be the ones holding the wheel while it drives.