CAC is the metric that decides whether a subscription app has a business. CPI and cost per trial are the metrics you actually steer on, because CAC arrives weeks late. CPA is not a metric at all until somebody says which event it counts. The practical skill is not picking a favourite number, it is knowing which of them is currently lying to you: CPI is honest about cost and silent about value, cost per trial is fast and mostly trustworthy, and CAC is the truth arriving too late to act on. Steer on the furthest number down the funnel that has stopped moving by the time you have to decide.
These four numbers describe the same spend and routinely sit a factor of thirty apart. That is not a rounding problem. Each one answers a different question, and most disagreements about app performance are two people quoting different rows at each other in good faith. What follows sets out what each number can decide, what it cannot, and how long you have to wait before it means anything. If you want the mechanics of calculating CAC itself, including cohort matching and what belongs in the numerator, that is covered separately in our guide to subscription app CAC. This article is about choosing between the numbers.
The five numbers, and what each one counts
Cost per install (CPI) is acquisition spend divided by installs. It is the fastest number in app marketing and the least informative, because it describes the price of a download and a download is not a customer.
Cost per acquisition (CPA) is spend divided by whichever event was designated as the acquisition. This is the one that causes the most damage, because the word hides its own denominator. A cost per registration, a cost per trial start and a cost per paid subscription are all called CPA, and on the same campaign they can differ by fifty times.
Cost per trial is spend divided by trial starts. For a subscription app this is usually the most useful single number, for reasons of timing covered below.
Subscriber acquisition cost is media spend divided by paying subscribers produced by that spend. It is CAC with a narrow numerator.
Customer acquisition cost (CAC) is the whole acquisition cost of one paying subscriber: media, creative production, tooling and the people or agency running it. Fully loaded CAC is the version that decides whether the company works rather than whether the channel works.
What each number can decide, and what it cannot
The two columns worth reading first are the last two. Most metric arguments are really disagreements about decision speed, and the final column is the one nobody puts in the table.
| Metric | What it measures | Decisions it supports | Decisions it cannot support | Trustworthy after |
|---|---|---|---|---|
| CPI | Cost of one install | Creative ranking within a market, bid and budget moves inside the week, spotting a broken ad set | Anything about whether an install was worth having | Hours |
| CPA | Cost of one unnamed event | Nothing until the event is named, then it behaves like whichever row it actually is | Comparison with anyone else's CPA, including an agency's case study | Depends entirely on the event |
| Cost per trial | Cost of one trial start | Day to day steering, creative decisions, market comparison, the working proxy for CAC | Telling you whether those trials convert, which is where apps differ most | One to two days |
| Subscriber acquisition cost | Media cost of one paying subscriber | Channel level go or no go, market entry and exit, scaling decisions | Company level profitability, because it excludes everything that is not media | Four to six weeks |
| Fully loaded CAC | All acquisition cost per paying subscriber | Payback, pricing, hiring and agency decisions, board reporting | Anything fast enough to act on inside a campaign | A quarter |
The real difference between these metrics is latency
Every step down the funnel buys truth and costs time. That trade is the whole subject, and it is measurable rather than a matter of taste.
Trials resolve almost immediately. Adapty's State of In-App Subscriptions 2026, drawn from roughly 16,000 apps, reports that 89.4% of trial starts happen on Day 0, the install day itself. RevenueCat's 2026 report finds the same pattern with category variation, from about 78% on Day 0 in Productivity to 89.9% in Business. These are two separate datasets rather than one confirming the other, but they agree on the shape: users who do not start a trial in the first session mostly never do.
Paid conversions behave completely differently. RevenueCat reports that about half of all paid conversions occur on Day 0, while 19.2% arrive in week six or later. Read that denominator carefully: it is the share of conversions that eventually happen, not the share of installs. The consequence for measurement is severe. A cohort's subscriber count is still climbing six weeks after the money was spent, so its CAC is still falling. Judge at day seven and you have seen roughly half the subscribers you will eventually get, which reports a CAC close to twice the real one.
The same lag shows up in revenue. RevenueCat puts median revenue per install at $0.23 at day 14 and $0.34 at day 60, so the identical cohort looks about 48% more valuable simply for being measured later. Nothing changed except the date of the query.
This produces the only rule in this article worth memorising. Steer on the furthest number down the funnel that has stopped moving by the time you need to decide. For a daily creative call that is CPI or cost per trial. For a monthly scaling call it is cohort subscriber acquisition cost. Using a slow metric for a fast decision means acting on half-formed data; using a fast metric for a slow decision means scaling something that was never working. Both failures are common and they look nothing alike.
One cohort, one month, three numbers thirty-eight times apart
The figures below are hypothetical and chosen to sit near published medians, not taken from any client account. Conversion rates are close to Adapty's 2026 global averages of 10.9% install-to-trial and 25.6% trial-to-paid, rounded for arithmetic that is easy to follow.
| Funnel step | Rate | Volume | Cost per unit |
|---|---|---|---|
| Meta spend, September | n/a | £20,000 | n/a |
| Installs | n/a | 12,500 | £1.60 CPI |
| Trial starts | 11% of installs | 1,375 | £14.55 per trial |
| Paying subscribers | 24% of trials | 330 | £60.61 subscriber acquisition cost |
| Plus creative and management | n/a | £23,500 total | £71.21 fully loaded CAC |
The last row adds £2,000 of creative production and £1,500 of campaign management to the £20,000 of media, which is the difference between asking what the channel cost and asking what the subscriber cost.
A £1.60 CPI is a respectable number in most consumer categories and would pass without comment in a weekly report. The same spend produced a £71.21 fully loaded CAC. Both numbers are correct, neither is being manipulated, and they support opposite conclusions about whether to increase the budget. The multiplier between them is fixed by the funnel: one divided by 11% times 24% is 37.9, so every install must be multiplied by roughly 38 to reach a subscriber, and every penny of CPI is 38 pence of CAC.
That multiplier is also where the leverage sits, and it points away from where most teams push. Cutting CPI by 10% to £1.44 saves £6.06 of CAC. Lifting trial-to-paid from 24% to 27%, which is a paywall and onboarding job rather than a media one, saves £6.74 and does not require the auction to cooperate. This is the arithmetic behind the familiar advice that the cheapest acquisition improvements usually happen after the install, and it is why lowering CPI should be a later move than it usually is. The full version of this calculation, including net revenue after store commission, sits in the CAC calculation guide.
How each number misleads, specifically
CPI hides mix shift
A falling CPI is often a geography change rather than an efficiency gain. RevenueCat's 2026 data puts median revenue per install at day 60 at $0.55 in North America against $0.11 in India and South East Asia, a five times gap, while media in those markets is very much cheaper. Shift delivery towards cheap inventory and CPI improves immediately while revenue per install falls further than cost does. The account looks like it is winning for as long as nobody looks past the first row. Platform mix does the same thing on a smaller scale, which is the subject of iOS versus Android CPI. The defence is to read CPI by market and platform rather than in aggregate, and never to celebrate a blended CPI move without checking what moved underneath it.
CPA hides its own event
CPA is the only metric here that can be quoted honestly and still mislead completely, because the reader supplies a denominator the writer never stated. When an agency reports a 50% CPA reduction, that claim means one thing if the event is a registration and something far more valuable if it is a subscription. Our own Napper case study publishes a 50% CPA reduction alongside doubled revenue and does not name the event on the page, which is exactly the thing this section says you should ask about. Ask us, and ask everyone else. A CPA figure without its event and attribution window is not a result, it is a shape.
Cost per trial hides trial quality
Cost per trial is the best fast proxy available, and it fails in one specific way: it assumes every trial is worth the same. Trial-to-paid varies enormously by category, with Adapty reporting 35.0% in Health and Fitness against 19.1% in Entertainment, and it varies by creative within a single app. A hook that overpromises will produce cheap trials that cancel on the first day. RevenueCat's data shows how early that decision gets made: on three day trials, 55% of all cancellations happen on Day 0. If you rank creatives on cost per trial alone, the most misleading ad in the account will usually win.
CAC hides when it was measured
Beyond the cohort matching problems covered in the CAC guide, the simple failure is reading the number too early and treating the result as a trend. Because conversions keep arriving for six weeks and beyond, a CAC measured at day seven is structurally worse than the same cohort measured at day sixty. Teams that check at inconsistent intervals produce charts that appear to show performance improving or degrading when all that changed was the age of the cohort at the moment of the query. Fix the window, then compare.
ROAS and LTV hide which half moved
Return on ad spend combines cost and value into one figure, which is what makes it good for reporting and poor for diagnosis. A ROAS decline can mean acquisition got more expensive or subscribers got less valuable, and those need opposite responses. LTV has the additional problem that for most apps it is a projection rather than a measurement, and projections made on young cohorts carry false precision. Use them as outcomes, and use the cost metrics above to explain them. The timing question specifically, including why a day seven number is not a small version of a day sixty number, is worked through in our piece on Day 7 ROAS.
Which number belongs on which dashboard
The mistake is not using the wrong metric, it is using one metric for every decision. Match the number to the cadence of the decision it informs.
- Daily, creative and budget: CPI and cost per trial, read by market and platform. Fast, directional, never used to justify a strategy.
- Weekly, allocation: cost per trial plus early paid conversions, with the explicit understanding that the paid number is incomplete.
- Monthly, scale or stop: cohort subscriber acquisition cost at a fixed window, compared against the previous cohort measured at the same age.
- Quarterly, business: fully loaded CAC against realised revenue and payback. This is the number for pricing, hiring and whether the channel deserves the budget at all.
If a single number has to go on the wall for a subscription app in growth mode, make it cost per trial by market. It is the furthest down the funnel that still resolves inside a day, which is precisely what the Day 0 trial data makes possible.
What this changes inside a Meta account
The choice of metric is also a choice about how much of that 38 times multiplier you hand to the platform. Meta documents app event optimisation as reaching "people most likely to take the action you specify, not just the ones most likely to install your app", and both Start Trial and Subscribe are standard app events. Optimising for installs makes CPI the thing the auction minimises, which is exactly the metric that is silent about value. Optimising deeper moves the platform's objective closer to the number you actually care about, at the cost of a sparser signal.
Two cautions worth stating plainly. Meta publishes no explicit minimum event volume for optimisation, so treat any specific threshold you are quoted as an opinion rather than a documented rule. And value optimisation is not generally available: Meta's own documentation describes it as available on a limited basis to advertisers on an allow list, so it should not appear in a plan as though it were a switch anyone can flip. How this interacts with campaign structure is covered in our write-up on Advantage+ app campaigns, and the wider funnel view sits in the subscription app marketing strategy guide.
Benchmarks are useful for sanity checking a CPI, and useless for deciding whether yours is good, because a good CPI is entirely defined by what happens after the install. If you want the current ranges by category and platform, they are in the 2026 mobile app CPI benchmarks and, for one vertical in detail, the health and fitness CPI benchmarks.
Frequently asked questions
What is the difference between CPI, CPA and CAC?
Cost per install is acquisition spend divided by installs. Cost per acquisition is spend divided by whatever event the ad account was told to count, which might be an install, a registration, a trial start or a purchase. Customer acquisition cost is the full acquisition cost of one paying customer, which for a subscription app means one paying subscriber. CPI and CPA are media efficiency measures, so they tell you how cheaply the platform bought something. CAC is a business measure, so it tells you whether that something was worth buying. The gap between them is the funnel: at an 11% install-to-trial rate and a 24% trial-to-paid rate, one subscriber costs roughly 38 installs.
Is CPI or CAC more important for a subscription app?
CAC decides whether the business works, so it is the more important number. CPI is the more useful number day to day, because you can read it within hours and CAC takes weeks. Treating that as a contradiction is what causes trouble. Use CPI and cost per trial to steer creative and budget inside the week, then use cohort CAC to decide whether the whole activity should continue at all. A team that only watches CPI scales a loss efficiently. A team that only watches CAC makes about one decision a month.
What does CPA actually mean for a mobile app?
Nothing, until somebody names the event. CPA is spend divided by a chosen action, and the action changes the number by an order of magnitude. A cost per registration, a cost per trial start and a cost per subscription are all reported as CPA in different dashboards, and they can differ by fifty times on the same campaign. Whenever a CPA figure appears in a report, a deck or an agency case study, the first question is which event it counts and over what attribution window. If that cannot be answered, the number is decoration.
How long should I wait before judging a cohort's CAC?
Longer than most teams do, and the reason is in the data. RevenueCat's 2026 report finds that about half of all paid conversions happen on the day of install, while 19.2% arrive in week six or later. That means a cohort's subscriber count keeps rising, and its CAC keeps falling, for well over a month after the spend. Judging at day seven captures roughly half the eventual subscribers and therefore reports a CAC close to double the true figure. Pick a fixed window, apply it to every cohort, and compare like with like rather than waiting for a number that never fully settles.
Can I use cost per trial as a proxy for CAC?
For most subscription apps it is the best available working proxy, because trials resolve almost immediately. Adapty's 2026 dataset puts 89.4% of trial starts on the install day itself, so cost per trial is knowable within a day or two while subscriber CAC is not. The condition is that trial-to-paid stays stable between the segments you are comparing. If one creative or one market brings in trials that convert at half the rate of another, cost per trial will rank them the wrong way round, which is why the proxy needs a monthly check against realised cohort CAC rather than blind trust.
Why do Meta's reported cost per purchase and my CAC never match?
Because they are different measurements, and no amount of reconciliation work will make them agree. Meta reports on its own attribution window and includes modelled conversions, particularly on iOS. Your subscription platform reports what was actually billed, net of refunds, on the date it happened. The two use different denominators, different timestamps and different definitions of a conversion. Use Meta's number for relative judgements inside the ad account, such as which ad set is cheaper than another, and use your own revenue data for absolute judgements about whether acquisition pays back.
The short version
CPI tells you what the auction charged. Cost per trial tells you whether the promise landed. Subscriber acquisition cost tells you whether the product delivered on it. Fully loaded CAC tells you whether any of it was worth doing. Cheap installs are not automatically good installs, and a subscription app only benefits from a lower CPI if trial and paid conversion hold up behind it. Pick the number that matches the decision in front of you, be explicit about how old the data is, and never let one metric carry a decision it was not built for.
A good share of the apps that come to us arrive with a healthy CPI and an unhealthy CAC, which is usually a creative and message problem rather than a media buying one. If you want to see what that work looks like in practice, the Steps & Beasts case study covers a fitness app where creative testing and onboarding changes moved revenue 145% and active subscriptions 118%, and the consumer apps page explains how we approach subscription acquisition on Meta. If you are trying to work out how much creative your spend actually needs, the creative refresh calculator is free and takes a minute, and our pricing is published if your app is already spending and you want to know whether this is worth a conversation.