When the model buys and makes the creative

Creative used to be the one part of the buy the network could not do for you. That is no longer a safe assumption. Unity said in August that Vector had passed a one billion dollar annual run rate, AppLovin has opened Axon Ads Manager to self-serve advertisers, and both companies describe systems that assemble and then cut and re-cut variants inside the same loop that places the bid. The direction is clear even where the labels stay cautious.

My claim is that when the same model both makes and buys the creative, creative stops being an input you supply and becomes a search space the network explores on your behalf. So the studio's job moves from producing ads to defining the boundaries of the search. Plenty of experienced buyers will disagree, and will point to the humans still writing briefs. The briefs will survive. What they describe will not.

The loop closes, and the incentive shifts

Think about what a network's model is optimising when it cuts a variant. Not your ninety day payer value. It is optimising the conversion event it can see, inside the window it can see it, at the bid it just placed. If a variant that misrepresents the game converts cheaply and the bounce happens after the model's horizon, the model has done its job, and the cost lands in your retention curve rather than in its loss function.

This is the underlying mechanism, and there is nothing sinister about it: a closed loop rewards whatever is legible inside the loop. Human creative teams had the same incentive with a much slower feedback cycle, which acted as a brake. Remove the brake and add generation, and the drift towards whatever converts at the margin becomes fast and continuous.

AppsFlyer's State of Gaming for Marketers report, published in January with Unity and Newzoo, put top advertisers at roughly 2,400 to 2,600 creative variations a quarter, up by a quarter or more on the year before. At that volume nobody reviews each cut; approval already happens by exception, and model-made creative simply makes the exception rate honest.

The second-order effect nobody prices in

The obvious worry is misrepresentation. The less obvious one is convergence.

When one model generates variants for thousands of advertisers on the same inventory, it learns one grammar of what works on that inventory. Your ads and your competitor's ads come out of the same distribution, tuned by the same objective. In March this site argued that AI can generate the first seventy percent of a playable and that the final thirty percent decides performance. Inside a network's loop, that final thirty percent increasingly belongs to the network rather than to you, and it is the same thirty percent everyone else gets.

Two consequences follow. First, differentiation moves upstream to the assets the model can only get from you: the game's own mechanics, its characters, its art, its progression. Second, learnings stop being portable, because a concept that wins on one network's model now tells you less about a second network than it used to, the win having been partly the model's rendering of your concept rather than the concept itself.

There is a trade-off here rather than a pure loss. The convergent grammar is convergent because it converts, so a studio that refuses it entirely pays a CPI premium for its principles. The question is where to draw the line, and that requires deciding in advance rather than variant by variant.

Three fences a studio can still hold

If the network owns the search, the studio owns the fences. In practice there are three that hold up under automation.

The asset fence. Decide which source material the model may work from, and version it. A curated asset kit of approved gameplay captures, characters, UI states, music is a different thing from handing over a build and a brand deck. The model can only recombine what it has. Deliberate scarcity in the kit is a control lever, not a production shortcut.

The claim fence. Write the short list of things you will let an ad show the game doing, and the shorter list you will not. Fail states that do not exist in the game, mechanics lifted from a different genre, difficulty misrepresentation. That list belongs in whatever review or filter you run on generated variants, as a hard exclusion rather than a note in the brief.

The measurement fence. Set the horizon you judge creative on, and make it longer than the network's. If the model optimises on a day-one event, your creative review should look at day seven or day fourteen cohort quality per concept family. Where the two disagree, the longer horizon wins the budget decision, even if the network's dashboard looks worse.

A decision rule for what to hand over

An illustrative split that has worked in review meetings: classify every creative decision on two axes, how reversible it is and how much it can move the game's perceived identity.

Low identity impact and reversible, such as hook order, caption phrasing, colour grading, aspect ratio? Hand it to the model entirely and never look at individual variants.

Low identity impact but slow to reverse, such as end-card structure that trains the store page conversion, is the middle case: let the model explore, then review winners weekly before they scale.

High identity impact is the one you keep. Which mechanic the ad shows, which character leads, what the game claims to be: humans decide, and the model executes within the claim fence.

Worked through with illustrative numbers, a studio running 400 live variants might find that around 300 sit in the first bucket, 80 in the second and 20 in the third. The creative team's whole week belongs to those 20 and to the fences. That is a smaller surface than it used to manage, and a far more consequential one.

What to watch for in your own account

The early signs that the loop has drifted show up before retention does. Variant families you never briefed start appearing, the share of spend on your top concept rises while that concept's day seven retention falls, and store page conversion drops because the ad and the page no longer describe the same game.

When those appear, the fix is not to reject variants one at a time. Tighten the asset kit or the claim list and let the search run again inside the smaller space. The model will find the best thing in whatever box you give it. Building the box is the creative strategist's job now, and the teams that treat it as a specification rather than a brief will be the ones whose games still look like themselves in the ads.

Related archive reading

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

• AI creative volume lowers the value of volume
• What a UA team looks like in 2030