Where iOS apps lose users: a funnel from download to paid, with benchmarks

Download, onboarding, paywall, trial, paid — what published research says a typical step converts at, which numbers Apple can tell you, and which ones need events from inside the app.

The funnel, and who can see each step

A subscription app's funnel runs roughly: downloaded, opened, finished onboarding, saw the paywall, started a trial, paid. Apple can see three of those six — the download, the trial start and the first charge — because they are store transactions. It cannot see the three in the middle, which is where most apps lose most people.

That split matters more than it sounds. It means the steps you can measure without writing any code are the ones you have the least control over, and the steps you can actually change require instrumentation. Both halves are needed to say anything useful: store data tells you the size of the loss, events tell you where it is.

Download to trial: single digits is normal

RevenueCat's State of Subscription Apps 2025 puts the median download-to-trial rate between 4.3% for low-priced apps and 9.8% for high-priced ones. Higher-priced apps converting a larger share of downloads into trials is counterintuitive until you notice it is partly a selection effect: an expensive app attracts fewer, more deliberate downloads.

The practical reading is that a single-digit download-to-trial rate is not a crisis to be fixed by paywall tweaks. It is the normal shape of the market, and the interesting question is what happens to the ninety-plus percent who never get that far.

Onboarding is where the silent loss happens

Design onboarding around the first useful outcome, defer nonessential setup and measure your own cohorts. Earlier numerical onboarding claims on this page were retired because their primary study could not be verified. See the reusable app system research for current sources and experiment guidance.

None of these are things Apple can tell you about. They are also the steps most apps never instrument, which is why 'our conversion is bad' so often turns out to mean 'two thirds of people never reached the paywall'.

Trial to paid, and the length of the trial

Trial-to-paid medians vary widely by category: RevenueCat reports 48.7% for travel and 43.8% for media and entertainment. Comparing your app to a cross-category average is how a perfectly healthy funnel gets declared broken.

Trial length is one of the few structural levers with a clean published answer: trials of 17 to 32 days converted best, at a 45.7% median. Longer is not better, and neither is the reflexive three-day trial.

One caveat that applies to your own numbers rather than the benchmarks: a trial started near the end of your reporting window has not had time to convert. Trial-to-paid over a short window always reads low, and reading that as a decline is a mistake the report itself should warn you about.

Hard paywall or freemium: a six-fold difference

By day 35, RevenueCat's medians for download-to-paid are 12.1% for apps with a hard paywall and 2.2% for freemium apps. That is not a small preference; it is the largest single structural difference in the dataset, and it means the two models cannot be compared on the same benchmark at all.

It is also not a free win. The same report puts refund rates at 3.4% for freemium and 5.8% for hard-paywall apps, with health and education apps around 4.7–4.9%. And roughly 30% of annual subscriptions are cancelled within the first month — a number that should temper any celebration of a day-one conversion spike.

So the honest version of the advice is conditional: a hard paywall converts more downloads and refunds more of them, and which is better depends on whether your app can demonstrate its value before it asks.

How to use a benchmark without fooling yourself

Match the segment before comparing. Price band, category and paywall model each move these medians by more than most product changes will. A comparison against the wrong segment is worse than no comparison, because it comes with a number attached.

Prefer medians to averages, and treat a small sample as a small sample: below about fifty people at the first step of a funnel, a ten-point difference is one person, and no amount of chart drawing changes that.