The State of AI in UA Creative

TL;DR: AI made ad generation cheap and fast, but it didn't make winning any easier. Finding a winner is a loop - decide, build, read, repeat, and AI only cheapened the building. Knowing what to try and what won is still human judgment. You run the loop until it turns one up.

----------------------------------------------------------------

If you run user acquisition for a mobile game, you have heard that AI changed everything about creative. That’s true, but not in the way the slogan means. The difference decides where you spend the next year of effort.

Here is what actually happened. Making an ad got cheaper and faster, and it is still getting cheaper and faster every month. A strategist who used to ship a handful of concepts a day can ship dozens now, and more than half of the top-grossing mobile games already run AI somewhere in their creative pipeline. The cost of another variation is decreasing.

What did not change is how creative winners are made. The number of games launched each month roughly doubled, and the number of hits that found an audience barely moved. We got much better at making content and no better at finding the ones that win. That gap is the most important fact in UA right now, and everything below is where it comes from and what to do about it.

What AI actually does now

Across the pipeline, AI now does the work that used to eat a creative team’s week with many creative tools that make it easier or faster to produce creatives - be it assets, optimization, editing, and more.

It is also able to spin up variations of the same concept at a lower and lower cost - similar to templates.

But creative tools have the same ceiling. They run on a concept you already have.

---

Keep exploring with Sett:
* How the UA manager became a scientist, and why the creative is now the targeting
* The two tracks every creative team runs: exploit the winners, explore for the next.
* What UA work becomes when your team is a set of AI agents you direct.

---

LLMs out of the box are interpolation engines. Hand it an idea and it returns a thousand nearby versions, fast and cheap. What it cannot do is invent the idea. It cannot tell you that the unrun angle on your game is the boss fight players keep screenshotting, because it has never played your game and has no idea why your last winner won.

Or what is fun for humans to play.

And it cannot tell you which creative will convert, because that answer is in the market, not in the model.

So the honest state of AI in UA creative is this. It has nearly solved making more. It has barely touched making winners.

Why this lands on creative

It is worth understanding why all of this pressure fell on creative and not somewhere else, because the reason is also the reason it won’t reverse.

For a decade the hard part of UA was finding the right person, and the networks did it for you. Facebook and the rest held a deep, personal picture of who everyone was, and put your ad in front of the people most likely to install and pay. ATT cut off the data that fed that picture, and it decayed. The networks could no longer target the way they had, so they pulled what was left, the bidding and budgeting and audience-matching, inside their own models.

That took the last dials out of your hands. The only input you still control is the creative. So the pressure that used to go into targeting now goes into making ads, and the move that lever rewards is volume. AI arrived exactly when the demand for volume became bottomless, which is why production is the part of the stack that got automated first.

Variations are cheap, concepts aren’t

The cheapest thing to make is a variation of something already winning. It is also the thing that, while having value (extending the life of a winning concept and battling creative fatigue), is not moving you closer to finding the next winner.

That lever that opens new ground is a new concept, and that’s the one AI left untouched. A studio that pours its new, cheap capacity into more variations is optimizing the part that was never the constraint. The constraint moved to concepts, and concepts are still made by judgment.

The job is a loop

Which means the word “production” has been hiding the real job. Making an ad was never the job. Finding a winner is, and finding is not a step you do once. It is a loop you run.

The loop has three parts. First, deciding what to try: which audience, which hook, which mechanic, drawn from competitors, neighboring genres, your own data, and the things only true about your game. Second, building it. Third, reading what happened when it ran, and why. The third part rewrites the first, and the loop comes around again, sharper each time.

AI is making the second part, building, cheaper and faster every month. That’s real, and it’s the easiest of the three. The other two are judgment, which is the part the model can’t do: knowing what to try, and knowing what worked. You can make a thousand ads a day and be no closer to knowing which one to make next.

So a winner is found, not chosen. You start from your most grounded read of what might work, build it, run it, and let the result point you at the next thing to try. Most attempts fail, on the order of ninety-seven percent in our experience, which is why a handful of clever ideas can’t carry you and a thousand copies of one can’t either. You keep the loop turning, and you let the results, not your attachment, decide what survives.

Every tool is a bet on the loop

Once you see the loop, every product in the market sorts the same way, by how much of it the tool owns and what it leaves to you. Where your stack sits tells you what to fix.

Point generation tools are the fastest-growing category and the easiest to adopt: Poolday, Creatify, MakeUGC, and dozens more. You give them a brief or a winning ad, and they generate assets, UGC-style videos, statics, sometimes playables. They are genuinely good at the building step and they will cut your production cost hard. But they own only the middle of the loop. They don’t decide what to make and they don’t read what worked, so the two hard parts stay on your desk. And because they all draw on similar models, whatever they can make for you they can make for your competitor. The edge is temporary by construction.

Building it in-house is the full loop, owned, running on your own data. It is the right answer in principle, and almost no one pulls it off. It takes machine-learning, creative, and AI engineering talent in the same room, building tooling that keeps pace with foundation models that turn over every quarter. A few of the largest studios manage a version of it. Everyone else starts it, understaffs one of the three parts, and ends up with a loop that is open exactly where it matters.

Agencies and creative studios sell human taste, which is still the best concept engine there is, and which the models can’t match at the top end. What you buy is a senior creative who has shipped winners in your genre, handing you a few high-conviction concepts. What you can’t buy is speed. People don’t produce at machine tempo, and there is no systematic way for last week’s results to reshape next week’s brief. You get craft, delivered slowly, with the feedback step missing. It is the opposite failure of the point tools: judgment without velocity.

Creative analytics platforms like Motion, Foreplay, and Segwise score your live creative and tell you which attributes drove performance: the hook in the first three seconds, the UGC framing, the color that held attention. Reading is the part of the loop everyone underinvests in, so this looks like the answer. But it only owns the reading. It hands you a verdict and leaves the next move to you. A dashboard that tells you what worked last week, which you can’t turn into next week’s tests fast enough, is a post-mortem.

Full-stack orchestration is a single system that runs all three parts and keeps them turning: it decides what to try, builds it, reads the result, and feeds that back into the next round. This is the only shape where the parts compound instead of resetting every batch of creatives, because the reading sharpens the next idea automatically. It is also the hardest to build, and worth less if any part is weak, the reading most of all, since that is what makes the loop learn. Most platforms claiming this are a bundle of point tools with a dashboard bolted on, and a bundle of mediocre parts with a loop drawn around it is worse than one excellent part.

Why the networks won’t do it for you

The obvious objection is that the networks already run a loop bigger than yours will ever be. More data, more compute, more impressions. Your loop should not be able to beat Axon’s.

It can, because it is a different loop. The network’s job is distribution: take what you give it and put it in front of the right people. It is extraordinary at that and getting better. What it cannot do is decide what to give it, because that decision runs on things the network has no access to and no reason to chase: the specific truths of your game, the angle only you can credibly run, the read on why your last winner actually won. The network optimizes the ad you hand it. It can’t explore the ad you haven’t made. It is waiting for you to feed it something it couldn’t have produced itself, and it suppresses you when you feed it more of the same.

That is also why this edge lasts. Once everyone has cheap generation, making more stops being an advantage, because the auction competes it away. How well you explore your own game doesn’t, because no one else can run that search.

What to do about it

Most teams are pointed at the part that no longer matters. Three changes fix that.

Stop measuring your creative team solely by output. Volume is the part that got cheap; rewarding it now rewards the wrong thing. Measure it by two things instead: how fast it gets from a live result to the next test, and how many of its new concepts came from reading the last batch rather than copying a winner.

Spend time on creating as structured a context layer as possible for any AI you work with to teach it and guide it to what is true for your game when it comes to UA strategy, creative strategy, experiments you ran, winning concepts and losing concepts.

And whatever you buy, ask the single question that sorts the market: does it close the loop, and does it close it on my game? A tool that only generates will save you money and win you nothing that lasts. A system that explores, builds, and learns, pointed at your specific game, is the only thing that compounds.

That last description is what we built Sett to be. Not a generator. A loop: agents that decide what to try, build the playables and videos to test it, read every result, and feed it into the next round, faster and wider than a team can by hand. The judgment stays human. The search runs at machine scale.

You can’t predict a hit. You don’t have to. You run the loop until it turns one up.