7.5× the installs, better retention: the post-mortem
A racing title that began life hyper-casual and grew a progression system wanted scale without the classic scale-rot. Fourteen weeks later it bought 30,000 installs a day at a $0.38 Android CPI, with day-7 retention up from 6% to 9% inside a genre band of 5 to 10%. Questions by Micha this time, answers by Ozan and Tereza.
4k → 30k
installs per day, week 1 → week 14
$0.38
Android CPI at peak, vs ~$0.40 hyper-casual median
6% → 9%
D7 retention, inside the 5–10% genre band
31
creatives live at peak, refreshed weekly
Q: What breaks first when a casual game scales 7x?
The creative supply line, always. At 4,000 installs a day, three good ads cover you; at 30,000 the same ads fatigue in a fortnight because you're buying deeper into every audience. The engagement was, at its core, a manufacturing problem: keep 25 to 35 healthy creatives live while the weekly rotation retires the sick ones. The market's biggest spenders ship thousands of creatives a quarter for exactly this reason; we matched the discipline at boutique volume.
Q: Where did the fail-state hooks come in?
They were the engine. The racing genre's most reliable hook family remains a run that goes visibly, fixably wrong, and our variants of it beat every polished alternative on installs per mille for fourteen straight weeks. The craft is in the wrongness: too stupid and you recruit rubberneckers who churn by day two; too subtle and nobody's thumb stops. The sweet spot, a mistake the viewer diagnoses in under a second, held IPM between 4 and 5 all engagement.
Q: How do you scale spend without buying garbage?
Gates, again, but tuned for volume. Every geo-network cell had a D7 floor and a CPI ceiling; a cell that broke either for two consecutive reads got its budget moved, not argued with. The interesting discovery was that retention rose under scale, from 6% to 9% at day 7, mostly because the fail-state hooks recruit players who arrive already wanting to correct the run, and the progression system gives them a reason to stay. Creative is targeting; the ad decides who shows up.
Q: Which geo nearly ruined everything?
A tier-2 market we won't name delivered installs 30% under the blended CPI and looked heroic for a fortnight, until the cohort curves came in: day-7 retention under 3% and ad revenue per user at roughly half the model. One geo, briefly 20% of spend, dragging the blended payback line underwater. The fix was undramatic, a cell-level kill like any creative funeral, but it earned the engagement's only standing rule change: no geo exceeds 15% of daily spend before its first cohort read.
Q: What did fraud filtering catch at this scale?
Enough to matter. Industry-wide, roughly 22% of non-organic installs carry fraud signals, and cheap Android volume is where it concentrates. The MMP's fraud rules plus network-level blocklists rejected a mid-single-digit percentage of installs at peak, clustered in exactly the inventory that looked most flattering on CPI. Every rejected install is a bid you stop making; the blended CPI fell measurably the week the strictest rules went live.
Asked after this story
The bid math. Hyper-casual lives on razor margins: Android CPIs around a $0.40 median, monetized almost entirely by ads, day-30 retention of 1 to 3%. Hybrid-casual adds progression and in-app purchases, which lifts tolerable CPI and makes retention worth defending. You scale a hyper-casual until arbitrage dies; you grow a hybrid-casual like a game.
The SDK networks did the heavy lifting at peak: AppLovin's AXON and Mintegral bought the bulk of Android scale, with Google App Campaigns steady underneath and TikTok supplying bursts when a hook family was fresh. That mix mirrors the market's current shape, where AppLovin leads tier-1 gaming and Mintegral ranks second on Android volume.
It was the chosen stop, not the ceiling. Past it, the next tier of volume priced above the LTV model's comfort in two of the five main geos, and the studio preferred margin to headlines. Scale is a dial, not a trophy; the D7 gate would have permitted more spend at worse economics, and nobody wanted to own that decision.
Scaling and nervous about it?
Correct instinct. Bring your cohort curves to a working session and we'll find the geo that's lying to you.
Start a run