There is a UA team somewhere right now celebrating a record-low CPI. And their app is quietly losing money.

The two facts are not in tension. They are the same fact, seen from the wrong metric.

CPI, cost per install, is the most-optimised and most-misleading number in mobile user acquisition. It is easy to measure, easy to compare, easy to move, and it tells you almost nothing about whether the users you bought were worth buying. ROAS, return on ad spend, is harder in every way and closer to the truth.

This is not a trick question with an obvious answer, though. Both metrics have a job. The mistake is letting the easy one drive.

Why CPI is so seductive

CPI is available immediately. Spend divided by installs. You can read it an hour into a campaign, break it out by source and creative, and optimise against it in real time. It feels like control.

And that immediacy is exactly the trap. Because the thing CPI measures, the cost of getting an app onto a device, is not the thing you actually care about. You care about revenue. CPI is a proxy for revenue that stops measuring at the precise moment the user becomes interesting.

Two sources can deliver identical CPIs and completely different businesses. Source A at a $3 CPI whose users retain and spend. Source B at a $3 CPI whose users open the app once and vanish. On a CPI dashboard, these are the same. In your P&L, one is growth and the other is a slow leak.

Optimising to CPI systematically rewards Source B, because the cheapest installs are very often the lowest-intent ones.

Why ROAS is right, and why it is hard

ROAS asks the only question that matters: for every dollar in, how many dollars came back?

That is the correct question. The difficulty is entirely in the timing.

ROAS is a lagging metric. On day zero, when you are making budget decisions, you do not know your D30 ROAS. You know your D0 ROAS, which is almost always deeply negative, because acquisition cost is front-loaded and revenue trickles in over weeks. So you are forced to predict: to look at early signals and estimate where the cohort will land.

The teams that win at ROAS are the ones that get good at that prediction. Early proxies like D1 and D7 retention, D7 ROAS, and early monetisation events, mapped against historical cohorts to forecast the D30 or D90 number you actually care about. It is real work, and it is the difference between UA as a growth engine and UA as an expensive install counter.

2026 made this harder, and more important

Two forces collided.

Installs got more expensive. US mobile game CPI reached roughly $4.90 in Q1 2026. iOS runs 30 to 50 percent above Android. When installs were cheap, a sloppy CPI strategy was survivable, because you could afford to waste some. At today’s prices, every low-quality install is a bigger hole.

And measurement got fuzzier. With ATT opt-in stalled around 27 percent, more than 70 percent of iOS attribution decisions in 2026 rest on probabilistic or SKAdNetwork data rather than deterministic matching, with estimated variance of 12 to 18 percent against ground truth. On Android, Privacy Sandbox enforcement is reshaping signal availability.

Here is the uncomfortable implication. In a probabilistic measurement world, CPI looks more precise than it is. You are reading it to two decimal places off attribution that carries double-digit percentage uncertainty. The metric feels exact and is not. ROAS, measured across a cohort over time, is more robust to that noise precisely because it aggregates real revenue rather than trusting a single attributed install event.

The harder measurement environment does not make CPI safer to lean on. It makes it more dangerous.

So which one drives?

ROAS drives the strategy. CPI informs the tactics. They operate at different altitudes and the error is collapsing them into one.

Use ROAS to decide where the money goes. Which sources, which geos, which campaigns get more budget is a ROAS question, judged on predicted cohort return, not on install cost.

Use CPI as a real-time diagnostic inside those decisions. Once a source is proven on ROAS, CPI is a useful early-warning signal. A sudden CPI spike on a good source is worth investigating today, before the ROAS data matures. A CPI that looks too good to be true on a new source usually is, and it is a flag to check retention before scaling.

Never let CPI make an allocation decision on its own. The moment “this source is cheaper” becomes “so move budget to it,” without a ROAS check, you are optimising toward the leak.

The practical setup

1. Instrument for cohort ROAS from day one. You cannot optimise to a number you are not measuring, and retrofitting cohort tracking mid-campaign is painful.
2. Build a prediction model, even a simple one. Map D1 and D7 signals to eventual D30 and D90 ROAS from your own history. Precision matters less than having a consistent forecast to act on.
3. Set ROAS targets by cohort, not blended. A blended target hides the bad sources inside the good ones.
4. Keep CPI on the dashboard as a diagnostic, clearly labelled as a leading indicator, never as a goal.
5. Reconcile against real revenue monthly. Where predicted ROAS and actual revenue diverge is where your model needs work.

The honest caveat

CPI is not useless, and pure ROAS purism has its own failure mode. Very early in a campaign, or at very low spend, ROAS data is too thin to act on, and CPI plus retention is a reasonable interim signal. New markets with no cohort history force you to fly partly on CPI until the data accumulates. The point is not to ban CPI. It is to keep it in its lane.

The teams that struggle are the ones using CPI because it is available, not because it is right. Available and right are different properties, and in UA the gap between them is where budgets quietly die.

This is why SpinX runs on CPA with MMP-verified attribution. A cost-per-action model aligns the entire relationship to outcomes rather than install counts, and our bidder tunes toward ROAS and retention goals, not toward whatever produces the cheapest install this hour. When your partner is paid for results, the incentive to chase a flattering CPI disappears from the structure.

Pick your driving metric deliberately. The easy one is easy for a reason, and the reason is not that it is telling you the truth.

See how SpinX helps performance teams optimise toward real ROAS at spinx.io.