Churn prediction is the easy part
Most players decide to leave in their first sitting, so a churn score that fires after a week of silence reports a decision already made. Three games this year showed us how to find that moment, and how to prove a fix worked.
By Patrick McGrath ·
Your churn tile is red, and the list behind it is long. At one studio we worked with this year, a player turned "at risk" after seven days without a session and "churned" after fourteen. The list ran to tens of thousands of names. The decision behind every one of them was a week old.
We saw the same timing in all three games we worked on this year: a strategy game, a portfolio of puzzle titles and a builder game. On the biggest puzzle title, half of new players were gone before level six, about 25 minutes of play. In the builder game, two thirds of a typical new cohort left before the end of the first day. In the strategy game, about 82% of installs had gone quiet by 72 hours, the earliest point its planned lapse messages would fire.
One of those engagements started from a cliff. The builder game's retention seemed to fall away after the first week, and our own proposal was sized on fixing it. The curve had been read at day one, seven and thirty, and three points give a curve no shape. When we added day three and day fourteen, the cliff went away. After the first day, every checkpoint lost about the same share of the players still there. The first day lost more than two thirds. Day seven was where a bad first day became visible.
Source: Higher Agency engagement data. One mobile builder game, its eight most recent fully observed install cohorts, 2026. Excludes a small self-selected segment that retains far better.
When most players leave, predicting who will leave is easy and close to worthless. In the strategy game, 60% of installs played a single session on install day, and 94% of them were gone by day seven. Flagging every install would have been right 89% of the time. The flag was right 94% of the time. And a third of the players who stayed had looked exactly like churners on install day.
The strongest predictors misled us more. Builder-game players who reached the trading loop on day one were still playing on day seven at 77%, against 44% for players who only leveled up. Strategy-game players with a hundred quests or more on record retained 17 times better than players with none. Both numbers describe who stayed, not why. The traders were 2% of the cohort and picked themselves, and a quest count keeps climbing for anyone who keeps playing. Turn either into a lever and you build a feature that rewards players who were staying anyway.
The score keeps getting cheaper. Google Analytics computes a churn probability for free. Google shut down Firebase Predictions and now points developers to a feature that picks which experience each player gets. This summer SAP bought the company behind TabPFN, a model that predicts from a table with no training at all. The score was never the scarce part. In 2014 Wooga published a churn model for its most valuable players, then tested it. Free currency sent to the players it flagged moved neither churn nor revenue against a control group that got nothing. The paper's lead author later co-authored a randomized trial that worked: at-risk players in a puzzle game got easier puzzles, and they played about a fifth more than a holdout and spent more too. The gift hadn't moved anyone. A change to the game did.
What found the moment, every time, was reading the game. On the biggest puzzle title the obvious suspect was ads, and a headline correlation convicted them. Controlled for level, the case fell apart. The answer was in the design sheet, which labeled level three easy. Level three's board had twelve times the pieces of level one's, and winning it took twelve times as long. In the builder game, players broke off at about 90 seconds, before any gate or price. Our first read blamed the power meter, until the config showed those first actions cost no power at all. The data told us where players left. Finding out why took recorded sessions and a designer who knew the build.
The same symptom had different causes in sibling games. On one puzzle title, the players who left were losing more than the players who stayed. On its sibling, from level three on, the players who left were winning more. One churn model across the portfolio would have prescribed the same easing for both, and on the second it would have done nothing.
Then comes the part nobody budgets for: proving a fix worked. The puzzle publisher had eased difficulty twice and reverted both times. One revert rested on a 91% revenue drop at p = 0.84, which is noise. The test that would settle the question needs about 160,000 players an arm and 45 days. The strategy game's rescue plan assumed 50,000 installs a day. Real volume was 25 to 30 times lower, so its three-day test needed months, and it depended on a push message the game had no server to send. By the end of our work, three puzzle teams had shipped new openings, two of them to every player with no holdout. Nobody can now say what those changes did to churn.
The fix also fights the business. Easing difficulty on one title cost 9.4% of purchases, because players buy when they fail. About a fifth of a cohort's 90-day purchase revenue has arrived by day seven, and the scorecards read day seven. The product manager who resists making the game easier is reading the only number they're measured on. On that number, harder pays, and the players it loses never show up.
So we build in a different order from the one vendors sell. Start by instrumenting the first session, down to the step the player was on and what was on screen when they left, and read the game before you model it. Then design the test before the fix, with a holdout and a horizon fixed in advance, sized to the traffic the game actually has. The score comes last, and its job is narrow. In two field experiments outside games, the customers most likely to leave and the customers an offer actually moved overlapped about as much as chance. A treatment you've proven belongs with the players it moves, and those are rarely the ones at the top of the risk list.
Finding the moment is a month of work with an operator. We sell it as the Product & Growth Audit, and it ends with build-ready specs and a test plan your traffic can run. When the moment is worth watching every day, the AI Build Program automates the watching. Live ops after AI covers what changed for one team once it landed.