The UA Creative Scientist

TL;DR: Production got cheap, discovery got hard. The networks are recommendation engines now, so your creative is the targeting. You don't predict hits anymore, you search for them, faster than anyone else. Meet the Creative Scientist.

97% of ideas fail, and the best teams plan for it. Read the full piece.

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

Something changed in mobile games over the last two years, and most of the people inside the industry still talk as if it didn't.

We got very good at making things. The tools collapsed the cost of building a game, and then collapsed the cost of the ads that sell it. You would expect that to produce more hits. It produced the opposite.

The number of games launched each month roughly doubled, while the number that found a real audience barely moved. When you make building cheap and discovery stays as hard as it always was, discovery quietly becomes the whole game.

That is where we are now. The bottleneck didn't disappear when production got easy. It just moved, and almost nobody moved with it.

The networks became recommendation engines

To see why, look at where the decisions are actually being made. A UA manager in 2016 was a buyer. You chose audiences, engineered the signal to send to the network, set bids, picked placements, and the network did roughly what you told it. That person barely exists anymore.

The networks that matter now, AppLovin's Axon, Meta's Andromeda, Unity's Vector and the rest, are not tools you operate. They are recommendation engines. You hand them a piece of creative and they decide who sees it, matching each ad to the person most likely to act on it, and become a high LTV user, across billions of impressions a day, at a precision no human buyer could touch.

AppLovin spent years building Axon and only just opened it to everyone; the company describes it as built to work directly with AI agents, and it threw off $1.84 billion in a single quarter running this way. Liftoff's CEO says their model now runs on twenty times the data it had a couple of years ago.

These are not ad managers. They are some of the largest content-recommendation systems ever pointed at human attention, and they happen to be aimed at selling games.

These systems also have opinions about what you feed them. They reward concepts that are genuinely different and quietly bury the ones that are only cosmetic variations of a current winner. The old reflex, find a hit and clone it twenty ways, now works against you. The machine rewards range and punishes repetition.

Once you accept that, the job rearranges itself underneath you. If the machine decides who sees what, then targeting is no longer a thing you do. The one input still entirely in your hands is the creative itself, which means the creative has become the targeting. The concept you choose is what tells the system which humans to go find.

That sounds like a small shift in responsibility. It is closer to a full inversion of the craft. For a decade the creative was the thing you wrapped around a media-buying strategy. The media buying is automatic now, and the creative is the strategy.

Ideation became a search

The part most people get wrong is thinking they can shortcut this with talent. You cannot reason your way to the right creative. If you could, the smartest team in the room would simply think very hard and produce the winner, and everyone who has done this work knows that is not how it goes.

The best creative directors alive are wrong most of the time, because taste is a kind of prediction, and you cannot predict a hit. What you can do is find one. That gap, between predicting and finding, is the entire difference between how this job used to work and how it works now.

Finding is empirical. You form a guess, you put it in front of the machine, you read what comes back, and you let the result decide where you look next. There is already a word for working this way. It's science.

So it helps to picture the work as a search through a tree. At the root is your game and a  sense of who might want it, and why. Every concept you run is a branch, a small hypothesis of the shape "this kind of person will care about this game for this reason."

Most branches are wrong, and the network tells you so fast and cheap. A few are right, and those become the trunk you grow the next round of branches from.

You are not trying to guess the winning ad on the first move, because you can't. You are trying to search the tree well: push hard where the signal is strong, cut what's dead without mourning it, and cover more of the tree than your competitors can before the window closes.

Ninety-seven percent (97%!) of ideas fail. A process that refuses to build that in, and keep moving anyway, is just an expensive way to feel disappointed.

When the search rewrites the game

The good branches have to come from somewhere, and they come from reading signals the rest of the market hasn't read yet. Now and then a branch pays off so strangely that it stops being about the ad at all.

Match Villains, doing north of $70 million in in-app purchases, that started running ads for a police-escape mechanic which appeared nowhere in the actual game. The ad worked. It worked so well that the studio behind it, Good Job Games, went and built a whole separate game to deliver what the ad had promised.

The ad came first. The game came second.

Sit with how backwards that sounds for a second and it turns into the most rational thing in the world. If a creative is the cheapest test you have of whether people want a certain kind of game, then a winning creative is not just a winning ad. It is the market telling you what to build. The search ran in reverse and rewrote the product.

The six places to look

Most searches are less dramatic than that, and they all roughly draw from six wells. The first is the competitive landscape: what is already winning around you, and just as usefully, what everyone has started copying, since that tells you which ground is crowded and which is still open.

The second is your own genre, the proven loops and hooks it reliably converts on, which is home base and the safest place to dig.

The third, and the richest, is the genres next door, because most fresh winners are not invented so much as crossed together, one genre's loop wearing another genre's skin. You mine your neighbors for mechanics, for types of fun, for themes and art and audiences you would never have thought to borrow.

The fourth well is the universal layer, the formats and hooks working across every genre at once, where you want the pattern underneath the meme and not the meme itself.

The fifth is your own performance data, which is less a source of ideas than a map of where you have already been, a running record of which branches are explored, which are exhausted, and which you have never touched.

The sixth is the game itself, its real and specific truths, the things that are only true about your game and nobody else's. That last well is where the arbitrage hides, because an angle no competitor can run is an angle no competitor can bid up, at least until the market notices and fills it in.

Speed is the whole edge there. The studios that explore fastest don't only save money, they reach the open pockets before everyone else crowds in.

Every angle is a door

Underneath all six wells you are really doing one thing, which is stretching the funnel. Every angle you find is another door into the same game, a different reason for a different kind of person to tap install. The more doors you open that actually lead somewhere, the more pockets of high-value players you reach, and the lower your blended cost to acquire them falls.

In many cases, these doors are simply ways to lower CPIs through marketing a sliver of your game as a very casualized version of the core loop. Sometimes almost as if you're marketing FTUEs and not the actual game.

This is not a refinement you do once you can afford it. CPIs climb around 15 percent a year, and games more than four years old already hold roughly 60 percent of all the revenue in the market. A new studio cannot outspend that, and cannot buy users more cheaply than the incumbents who got in when users were cheap.

The one variable left fully in your control is how much of the space you can explore, and how fast. You don't out-budget the giants. You out-search them.

Let the search begin

This is a real job now, and it deserves a real name. Call it the creative scientist: the person who stops treating creative as an act of taste and starts treating it as a search, who designs the experiments, reads the signals, maps the space, and keeps turning up the pockets everyone else walked past.

It is a UA job and a creative job fused into one, which is exactly why neither of the old titles fits it cleanly anymore.

The trouble is that the job, done properly, is now too big for a person. The space you have to cover is enormous, the networks burn through creative in weeks where they used to take a quarter, and one human with a spreadsheet cannot branch fast enough to keep the tree alive.

That is the real reason we built Sett the way we did. Not to replace the scientist, but to hand them a lab. Agents that explore the space, build the playables and the videos to test each hypothesis, and feed every result back into the next round of branching, at a speed no team could reach by hand.

The judgment stays human. The search becomes machine-scale.

None of this makes hits predictable. That was never on the table, and anyone who tells you otherwise is dishonest. What has actually changed is that you no longer have to predict, because you can explore at a scale and speed that turns "we got lucky" into something much closer to a method.

You still cannot name the winner before you have seen it. But search the tree well enough, with the right machine doing the digging, and you can be almost sure the winner is somewhere in there, and that you will be the one who finds it.

That is a system for winning.