AI creative volume lowers the value of volume

For about three years the advantage in mobile game creative went to whoever could produce the most variants. Volume found winners faster, fed the delivery models more signal, and kept fatigue at bay by rotation. Generative tooling has now made that volume available to everyone, and an advantage available to everyone is not an advantage.

The claim here is that the marginal value of a creative variant is falling toward the cost of reviewing it, while the marginal value of a genuinely distinct concept is rising. Teams that respond to cheaper production by producing more are optimising a variable the market has already neutralised. Many production leads will disagree, because their pipelines still exist to deliver volume and their targets still count it.

The auction cancels a shared advantage

The mechanism is the auction itself. For every impression, a network picks the ad most likely to produce the outcome it is optimising for. An advertiser with a hundred variants of a concept has a better chance that one of them is the best fit for a given viewer than an advertiser with five. That was the volume edge: more tickets in the selection lottery.

The edge only exists relative to competitors. If every bidder in the auction arrives with a hundred variants, the selection advantage cancels. What remains is the distribution of concepts across bidders: who is offering the viewer something the others are not. Variant count becomes a cost of entry, and the differentiator moves up a level.

AppsFlyer's State of Gaming for Marketers 2026, published in January, put top advertisers at 2,400 to 2,600 variations a quarter, which is up by a quarter or more year on year. UA Ledger read that report the day it came out in "2,500 creatives a quarter: reading AppsFlyer's State of Gaming 2026". The number I take from it a month later is the growth rate: if the top tier is adding a quarter more variants each year and everyone else is catching up with the same tools, the volume gap between advertisers is closing, not widening.

Where the cost went

Cheaper production did not make creative cheap. It moved the cost.

Every variant still has to pass through a review and a trafficking step, then a learning phase, then a decision about whether to keep it; none of those got faster because generation did. A team that doubles its variant output without doubling its review capacity ends up with reviewers skimming, learning budgets spread thinner across more ads, and a rising share of variants that never exit learning at all. The exploration tax I described in the context of portfolio diversity applies with more force here: a hundred variants of one concept can cost more to learn than the concept is worth.

There is a market-level version of the same problem. Generative tools reproduce their training data, and that training data is whatever already works. Feed an industry's winners into the same handful of models and the output converges: the same pacing, the same failure-state framing, the same colour grade. When every advertiser in a genre is serving close cousins of the same ad, fatigue stops being a per-account phenomenon and becomes a shared one. The viewer tires of the category's look, and no individual account can rotate its way out.

Unity's Q4 results earlier this month, with Unity Vector now most of its Grow revenue, and AppLovin's 66 percent revenue growth, both point in the same direction: delivery is increasingly a model finding the right viewer for an ad. Models are very good at ranking near-identical variants. They cannot manufacture a distinct concept that does not exist in the account.

A decision rule for the variant cap

The practical response is to cap variants per concept at the point where one more variant's expected lift falls below the cost of testing it, and to move the freed budget into concept discovery.

A simple procedure estimates the cap.

• Rank a concept's variants by spend-weighted performance after they have all exited learning.
• Measure the gap between the best variant and the median.
• Read the gap. A large one means variation is still finding real differences; a small one means the variants are interchangeable.
• Compare the incremental lift of the best variant over the concept's first-round winner against the fully loaded cost of testing one more variant, including review time and learning-phase spend.

When the gap between best and median has narrowed to within noise for two consecutive rounds, stop producing variants of that concept. Any further volume is buying rank-ordering the network would have done anyway.

A worked illustration

Take a concept that has been through three rounds of variation, twelve variants per round, with illustrative results. Round one's best variant beat the original by 18 percent on cost per retained user. Round two's best beat round one's best by 6 percent. Round three's best beat round two's by 1 percent, and the median variant in round three sat within 2 percent of the best.

Each round cost, say, 9,000 across production and review plus learning-phase spend. Round one earned that back many times over. Round two roughly broke even at the concept's spend level. Round three lost money with near certainty, and the team would have been better off spending the 9,000 on a single new concept in an emotional register the account does not yet cover.

The pattern repeats across accounts I have seen: the second round of variation usually pays, the third rarely does, and pipelines built for volume keep producing a fourth because a fourth is what the pipeline's target counts.

Reprice the pipeline

If the marginal variant is worth less, the production team's key metric should stop being variants shipped and start being concepts validated. That's a harder number to hit, and a much more useful one to answer for.

The tooling can still help. Point it at the empty spaces in the portfolio rather than at the winners, ask for the concept the account lacks instead of more of the concept it has, and let the network do the ranking. Ranking is now free. Distinctiveness is not.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• When the model buys and makes the creative
• What a good creative brief looks like in 2026