The CPI obsession is a category error

Cost per install is a price, not a result. The industry treats it as a result, and that mistake is the root of a surprising share of bad UA decisions. A CPI target does not tell you whether you bought anything worth having. It tells you what you paid at the door, which is the same information a shop would get from knowing its average purchase price without knowing what it sold.

The claim here is stronger than the familiar "look at ROAS too". It is that CPI belongs in a different logical category from the metrics it is usually listed beside, and that putting it on the same dashboard row as ROAS or payback does active harm, because people optimise the number they can move fastest.

What kind of thing CPI actually is

Every UA metric is one of three things: an input, an intermediate, an outcome. Spend is an input. Impressions, clicks, installs are intermediates, the things that happen on the way to the result. And the result is revenue, or retained players against cost.

CPI is the ratio of an input to an intermediate. It is a unit price for a unit of intermediate output. That makes it useful for exactly one purpose: understanding the marginal cost of pushing more volume through a given channel and creative. It is genuinely valuable for that. Marginal CPI rising as you scale a campaign is the clearest signal available that you are exhausting an audience.

What CPI cannot do is tell you whether the unit you bought has value. An install is not a homogeneous good. The distribution of value across installs in a typical free-to-play game is so skewed that the average is nearly meaningless; a small fraction of players produce most revenue. Two campaigns at the same CPI can differ in value by an order of magnitude, and a campaign with double the CPI can be the cheaper one per unit of revenue. UA Ledger made a related point in March in "Why lower CPI can hide a weaker player cohort", but the argument there was about cohort quality. The argument here is about the metric's category.

The mechanism: why the number gets optimised anyway

If everyone knows this, the interesting question is why CPI still dominates so many weekly reviews.

The first reason is latency. CPI resolves in hours. ROAS at day 30 resolves in a month, and true payback for a mid-core title may not be knowable for a year. Humans and algorithms both prefer feedback that arrives. The metric with the shortest loop wins the attention contest regardless of its relevance.

The second is attribution certainty. On iOS under SKAN and AAK, the install is one of the few events you can still count with confidence at campaign level. Downstream events arrive coarse, delayed, often null. CPI feels solid because the platform mechanics made it the least degraded number, not because it is the most important.

The third is incentive structure. Agencies and internal teams are often judged on efficiency metrics that resolve within the quarter. A CPI target is auditable and defensible; a payback target that nobody can confirm for nine months is a political liability, so the organisation quietly promotes the price to the status of a result.

The second-order damage

The obvious cost is buying cheap installs that never pay back. The less obvious costs are what a CPI culture does to creative and to bidding.

Creative teams optimising for install cost learn to make ads that maximise click-to-install conversion. In practice that means misleading gameplay, fake fail mechanics, store pages tuned to a promise the game does not keep. The ad wins on CPI. The cohort churns by day three, the creative team hears that the ad worked, and it takes months for the retention curve to expose what the loop has been training all along.

On bidding, a CPI target pushes buyers toward the cheapest inventory: low-intent placements, incentivised traffic, geographies with low prices and low purchasing power. Automated bidders given a CPI cap will find those pockets with impressive speed. Every install the algorithm buys is a real install; the target was simply pointed at the wrong quantity.

There is a trade-off in the opposite direction that gets less attention. Teams that overcorrect to ROAS-only bidding lose the marginal-cost signal that CPI provides so well. They scale a campaign whose ROAS looks fine on average while the marginal install costs three times the average, because the average blends early cheap volume with late expensive volume. The right answer is not to discard CPI but to put it back in its category.

A decision rule that keeps CPI in its place

A practical framework is to give each metric a single job and refuse to let it do another.

• CPI answers one question: is the marginal cost of volume in this campaign rising, flat or falling? It is read as a trend, never as a target.
• Early value proxies (day one retention, a tutorial completion rate, first purchase within 48 hours) answer a different question: is the cohort I bought behaving like cohorts that paid back before?
• Those proxies are the daily steering signals.
• ROAS and payback at fixed horizons answer: did the spend produce a return? These are the targets, set per channel and per geography, and reviewed on their own timeline.

The rule for weekly meetings is simple. Nobody may propose a bid change on the basis of CPI alone. A CPI movement is a prompt to look at the early value proxies for the same cohort. If CPI fell and the proxies held, scale. If CPI fell and the proxies fell, you found cheap traffic, not efficiency.

A worked illustration

Consider an illustrative 4X title with two Android campaigns. Campaign A has a CPI of $4 and day-one retention of 38%. Campaign B has a CPI of $7 and day-one retention of 47%, with a first-purchase rate roughly double A's.

A CPI review would cut B's budget. A category-correct review starts elsewhere: B's cost per retained player at day one is about $14.90 against A's $10.50, so on that proxy alone A still looks cheaper. Then the first-purchase rate tips it. If B produces twice the early payers per install at 1.75 times the price, B's cost per early payer is lower. The buyer scales B and watches its marginal CPI as the signal for when to stop, which is the only job CPI was ever qualified to do.

That last step matters. The purpose of demoting CPI is not to ignore it. It is to make sure the number that arrives fastest stops answering questions it was never asked.

Related archive reading

These articles provide related context and remain subject to their stated review status.

• UA metrics explained: CPI, ROAS, LTV and payback
• The auction, not the audience, sets your CPI