Production got cheap.
Reviewing it did not.

The adoption figure is not in dispute. Generative AI accounts for roughly one-third of ad assets in 2026, up from one-quarter in 2025, and is forecast to reach 43% by 2027.¹ Nearly two-thirds of video ad buyers now use it.¹

Figure 01

GenAI share of ad assets

Proportion of advertising assets produced with generative tooling.

Source: industry forecasts for generative AI in advertising production, 2026. 2027 is a forecast, not a measurement. The 2025 and 2026 figures are reported as approximate shares - a quarter and a third - and we render them as 25% and 33% accordingly rather than implying single-point precision.

Around this curve sits a set of economic claims that arrive with striking confidence: $2.65 per asset. Seventeen to twenty-six minutes per finished asset. An 80% reduction in production cost. A tenfold increase in output. A 400% increase in content volume.²

We tried to source them.

The adoption curve is evidence. The cost multipliers are advertising for the thing they describe.

Production was not the constraint

Grant the vendors their numbers for a moment. Assume production really is ten times faster and eighty per cent cheaper. The question that follows is the one nobody in that sales conversation asks: was production the thing stopping you?

For a small number of organisations, yes. For most, the binding constraints on creative were somewhere else entirely - in approval cycles, in legal and brand review, in the number of stakeholders who must see a thing before it ships, and in whether there is enough traffic to learn anything from what did ship.

Relieving a constraint that was not binding does not increase throughput. It moves the queue.

The arithmetic of reading a test

The last item in the right-hand column is the one that can be calculated rather than argued, so let us calculate it.

To detect a 10% relative improvement on a 2% baseline conversion rate, at 95% confidence and 80% power, a standard sample-size approximation gives roughly 78,400 impressions per variant.

Now fix the media budget. Take a campaign delivering one million impressions in a quarter - a realistic figure for a mid-sized programme. Divide it.

Figure 04

Impressions available per variant, at fixed traffic

One million impressions, divided across an increasing number of creative variants.

This is our arithmetic, with parameters stated. Sample size per variant uses the approximation n = 16·p(1−p)/δ² with p = 2% baseline conversion and δ = a 10% relative lift, giving 78,400. The one-million-impression budget is an illustrative figure, not a measured benchmark. Change any assumption and the ceiling moves - a higher baseline rate or a larger detectable effect lowers the requirement substantially. The shape of the curve does not change.

At one million impressions, the number of variants you can evaluate to that standard is approximately thirteen.

Not thirteen that you can make. Thirteen that you can read. The hundredth variant is not a hundredth of a learning - it is a rounding error with a filename.

This is not an argument for making thirteen assets - volume has real uses, set out below.

It is an argument against the sentence that follows most volume claims - that more assets means more learning. More assets on the same budget means less learning per asset, and past a threshold you can calculate, none.

The steelman for volume

There is a serious argument for producing far more than thirteen, and it does not depend on any vendor claim. It deserves stating properly rather than being waved at.

Feature 02 of this edition documents the shift to automated buying: Google's automated shopping product now takes two-thirds of Shopping budgets. These systems are not campaigns in the old sense. They are allocation engines, and they consume creative as raw material. Supply them four assets and they have almost nothing to allocate between; supply them sixty and they have a combinatorial space to search. The system's own optimisation is doing a kind of selection that no human test protocol replicates.

Three other uses are equally legitimate and equally indifferent to statistical significance:

• Coverage. Placements, aspect ratios and formats multiply. One concept correctly built for every slot is not thirteen experiments; it is one experiment, correctly delivered.
• Localisation. Markets, languages and cultural variants. A version for Tamil-speaking users is not competing with the Hindi version for a learning; both are serving different people.
• Fatigue. Refresh slows decay in performance over a flight. Rotation is a maintenance activity, not a test.

Each of these is a real reason to make a hundred assets, and generative tooling serves all of them well. What none of them produces is a finding.

Volume for coverage, localisation, fatigue and machine allocation is sound practice. Volume described as learning is the error, and it is usually the same volume.

The failure is not overproduction. It is the report at the end of the quarter that lists ninety-four assets and presents the best-performing three as insight, when the difference between them sits well inside the noise their impression counts can support.

What the market is paying for this

Money is arriving well ahead of the evidence. Generative AI in the creative industries is put at $4.06bn in 2025, rising to $5.38bn in 2026 - a compound growth rate of 32.3% - and forecast to reach $14.03bn by 2030.³

Set that against what Figure 02 established. The category is compounding above thirty per cent a year on an evidence base in which the central economic claims - cost per asset, time per asset, output multiple - could not be traced to a single study with a stated method.

That gap is not evidence of a bubble, and we are not claiming one. Plenty of technologies are adopted for reasons that are obvious to practitioners long before anyone publishes a controlled study. But it does mean the spending decision is currently being made on vendor arithmetic, and a buyer at these growth rates should expect to generate their own numbers rather than wait for the category to produce them.

The pricing model manufactures the volume

One published price ladder is worth looking at directly, because it reveals the commercial logic: $3.90 per ad on a $39 monthly plan producing ten videos; $1.98 per ad on a $99 plan producing fifty.²

Figure 05

Unit price falls only if consumption rises

Published subscription pricing, cost per asset.

Source: published vendor subscription pricing, 2026. These are list prices, verifiable as such. The observation drawn from them is ours: the per-asset figure only improves if the buyer commits to five times the output, whether or not five times the output is useful.

The cheaper per-unit price is real. It is also conditional on producing five times as much. A buyer optimising the headline number is committing to volume before establishing that volume helps - and the total bill went up, from $39 to $99, while the number that improved was the one nobody is billed on.

This is the same structure Feature 01 of this edition identifies in media buying: a falling unit cost inside a rising total is not efficiency, it is a mix change. The unit got cheaper because you bought more of a cheaper thing.

Two costs that scale with volume

Two exposures grow with asset count and neither is priced into the per-asset figure.

Brand consistency. Every generated variant is a chance to drift - a wrong tint, an off-scale logo, type that is nearly but not quite the brand face. At ten assets a human catches it. At a hundred, the review that catches it costs more than the assets did, and the review that does not catch it ships the drift into market.

Accessibility exposure. Feature 15 of this edition sets out the position: the European Accessibility Act has been enforceable since 28 June 2025, first legal notices arrived nine days later, and breaches are observable from outside the organisation. Every additional public asset is an additional observable artefact. Automated layout optimises for visual density and brand impression, not for contrast thresholds or meaningful alternative text.

Both costs land in review - the stage the tooling does not touch.

What to do about it

Calculate your own ceiling before you scale output. Take your actual quarterly impressions, your actual baseline conversion rate and the smallest lift worth acting on. The formula is in the Method box. The answer tells you how many variants you can learn from - and everything beyond it is production for coverage, not for evidence.

Separate the two kinds of volume in your plan. Assets made to be tested and assets made to fill placements are different products with different success conditions. Conflating them is what produces a hundred variants and no findings.

Price the review, not the asset. If generation costs $2.65 and review costs a brand manager an hour, your unit economics are set by the second number. A tenfold increase in the cheap input against a fixed expensive one does not reduce cost per shipped asset by anything like the headline.

Move brand and accessibility rules into the component, not the checklist. Constraints enforced in the template hold at any volume. Constraints enforced by a reviewer hold until the reviewer is overwhelmed, which now happens sooner.

Do not accept a cost claim without a baseline. "80% cheaper" than what, measured over what period, including which stages. If the answer is unavailable, the number is marketing.