Cohort Analysis for Mobile Apps: How to Read Retention and Revenue Curves That Move Budget

Cohort analysis for mobile apps is the practice of grouping users by a shared starting event, usually the install date, the first subscription date or the campaign that acquired them, and then tracking how each group retains, converts and pays over time. A cohort is a group of users who did the same thing in the same time window. A cohort curve is the line that shows what share of that group is still active, or how much revenue it has produced, on day 1, day 7, day 30 and beyond.

Cohort analysis matters because almost every blended app metric hides the thing a growth team needs to know. Total daily active users can rise while every new cohort retains worse than the last. Blended ROAS can hold steady while one channel quietly fills the funnel with users who never renew. In Admiral Media’s work managing more than €500M in mobile ad spend across 150+ mobile brands, the cohort view is where budget decisions get made: which campaign to scale, which creative to cut, which optimization event to bid on, and how long to wait before judging a new market.

This guide explains how to build cohorts that answer real questions, which time windows to read for which decision, how to connect cohort data back to bidding, and what Admiral Media has learned from case studies where changing the optimization signal changed the economics of an app.

What cohort analysis actually tells you that a dashboard does not

Cohort analysis tells you whether the users you are buying today are better or worse than the users you bought last month, measured at the same age. That age-matching is the whole point. A dashboard that reports “revenue this week” mixes users who installed yesterday with users who installed a year ago, so improvements in acquisition quality and declines in retention cancel each other out and nobody sees either.

Three questions are only answerable with cohorts:

• Is acquisition quality improving? Compare the day 7 retention or day 7 revenue per install of this week’s cohort with the same cell for cohorts from four, eight and twelve weeks ago.
• When does a cohort pay back? Plot cumulative revenue per install against the cost per install for that cohort and read the day on which the curve crosses the cost line. Admiral Media’s guide to CAC payback period for mobile apps covers the payback math in detail.
• Which acquisition source produces durable users? Split cohorts by channel, campaign, creative concept or country and compare the shape of each curve, not only its starting point.

The mechanism behind all three is the same. Ad platforms optimize toward whatever signal you feed them in a short window after install. Cohort analysis is how you check whether that short-window signal still predicts the long-window value your business depends on. When the two drift apart, the algorithm keeps finding users who look good on day 1 and disappoint by month 3, and only a cohort view shows the drift.

The four cohort types every app team should run

Most app teams need four cohort types: install cohorts, subscription or first-purchase cohorts, acquisition-source cohorts and behavioral cohorts. Each one anchors on a different event and answers a different budget question, so running only one of them leaves blind spots.

Analytics platforms already support these anchors. Google Analytics 4 lets you set a cohort inclusion condition, such as first touch, any transaction or a specific event, and a separate return criteria condition that a user must meet to stay counted, as described in the GA4 cohort exploration documentation. Apple’s App Store Connect Analytics offers both download cohorts and subscription cohorts, with metrics such as Download to Paid, Proceeds per Download and Subscription Retention, according to Apple’s cohorts help page.

Behavioral cohorts are the most underused of the four. They are how you discover which early action actually predicts a paying user, which is exactly the input you need when choosing a proxy event for campaign optimization. Admiral Media’s guide to app event optimization explains how to turn that finding into a bidding signal.

Why the cohort window matters more than the cohort itself

The window you read a cohort at decides what you conclude about it. Read too early and you reward channels that convert impulsive users; read too late and you scale nothing because you are always waiting for data. The right window depends on how long your users take to pay, and that varies far more by business model than most teams assume.

RevenueCat’s State of Subscription Apps 2026 report shows how wide that spread is. Across subscription apps, 50.6% of conversions happen on Day 0, yet 19.2% happen in Week 6 or later. For freemium apps, Week 6+ conversion accounts for 22.9% of conversions, compared with 15.3% for hard paywall apps. In the Travel category, the Week 6+ share reaches 28.2%.

The practical consequence is simple. If a freemium app judges campaigns purely on day 7 revenue, it is systematically undervaluing the sources that bring patient users, because a meaningful share of those users have not paid yet. A hard paywall app can lean harder on early windows, because more of its value is visible early. Neither approach is universally right. The window has to match how your own cohorts mature, and you can only learn that by tracking full curves for several months before you lock in a short-window proxy.

Early-window behavior still matters, just for a different reason. The same RevenueCat report finds that 55% of all 3-day trial cancellations happen on Day 0. That makes the first 24 hours a strong early-warning signal for trial quality, even when it is a weak signal for final revenue.

The Admiral Media Cohort Readout Framework

The Admiral Media Cohort Readout Framework is a seven-step method for turning cohort tables into spend decisions. It exists because most cohort reports get built, admired once and ignored, since nobody defined in advance which cell answers which question. The framework forces that definition before any data is pulled.

The Admiral Media Cohort Readout Framework

1. Name the decision first. Write down the budget or product decision the cohort must inform, such as “shift spend from campaign A to campaign B” or “change the trial length”. A cohort without a decision attached is reporting, not analysis.
2. Pick one anchor event. Choose the inclusion event that matches the decision: install for acquisition questions, trial start or first subscription for monetization questions, a product action for event-selection questions. Do not mix anchors in one table.
3. Define value and return precisely. Decide whether a cell means retained users, paying users or cumulative revenue per user, and whether revenue is gross or net of store fees and refunds. Ambiguity here is the most common reason two teams read the same cohort differently.
4. Match granularity to decision speed. Use daily cohorts to catch broken campaigns, weekly cohorts for channel and creative decisions, and monthly cohorts for pricing and LTV. Weekly cohorts in GA4 run Sunday to Saturday rather than as a rolling seven days, so align your reporting week accordingly.
5. Split by the levers you control. Break cohorts down by channel, campaign, creative concept, country and platform. A split by a dimension you cannot act on, such as device model, rarely changes a decision.
6. Compare only mature, equal-age cells. Never compare a cohort’s day 30 value with another cohort that is only 12 days old. Mask or grey out immature cells so nobody reads a partial number as a final one.
7. Close the loop into bidding. Use what the long-window cohorts reveal to recalibrate the short-window signal the ad platforms see: the optimization event, the value rules, the ROAS goal window or the SKAdNetwork conversion value schema.

Step 7 is where cohort analysis earns its cost. Admiral Media’s teams treat the long-window cohort as the source of truth and the platform’s short-window optimization as an approximation that needs regular correction. When they diverge, the fix is almost always a change in what the algorithm is told to find, not a change in budget alone.

How Admiral Media turns cohort insight into better bidding

The most valuable output of cohort analysis is a better optimization signal. Ad platforms such as Google App campaigns and Meta app campaigns learn from the events and values you send them, so changing which event you send, or attaching a value to it, changes the users the algorithm goes looking for. Four Admiral Media case studies show the pattern.

ChatPDF: adding an LTV signal to bidding

Admiral Media managed ChatPDF’s paid acquisition on Google Ads and Meta, consolidated overlapping ad sets, tested target CPA against target ROAS bidding, added value rules and introduced an LTV signal into bidding, alongside weekly creative concept tests. On the index baseline comparing Year 1 with Year 2 year to date, the results by channel were:

• +320% ROAS on Google Ads: year-over-year ROAS growth, with subscriptions up 142% and CAC down 38%.
• +280% ROAS on Meta: year-over-year ROAS growth, with subscriptions up 171% and CAC down 45%.
• +156% subscriptions and -42% CAC: the headline results reported on the case study page.

The cohort lesson is that a platform bidding on a flat conversion treats every subscriber as equal. Value rules and an LTV signal tell it that some subscribers are worth more than others, which is only possible once you know from cohort data which users produce long-term value.

FET: rebuilding the event combination and iOS conversion setup

Admiral Media identified a combination of optimization events for the FET dating app that aligned campaigns with the actions that drive monetization, rebuilt the iOS SKAdNetwork conversion setup from scratch, and produced more than 60 ad iterations across 9 creative concepts. The FET case study reports:

• -66% cost per subscription: the headline result of the FET engagement.
• +162% subscriptions: growth in paying users reported on the case study page.
• +181% conversion rate: the reported uplift in conversion rate.

On iOS, where privacy rules limit user-level cohorts, the conversion value schema is effectively the cohort definition the ad networks see. Admiral Media’s guide to SKAdNetwork conversion values explains how to design that schema so the early signal tracks the revenue your cohorts actually produce.

Himalaya: moving the optimization event as data matured

For the audio-learning app Himalaya, Admiral Media audited the marketing tech stack, improved event tracking and defined proxy events for optimization. Campaigns started with signup optimization and moved to trial optimization once there was enough data, while the team tested by OS, country and language across the US, Latin America and Chinese-speaking audiences. The Facebook Ads results reported in the Himalaya case study were:

• +875% installs: increase in the number of installs.
• +676% users acquired: increase in the number of users acquired.
• -79% CAC and -83% CPI: reductions in customer acquisition cost and cost per install.

This staged approach is how cohort maturity and bidding should interact. A shallow event like signup gives the algorithm enough volume to learn; once cohorts show which users go on to start trials, and there is enough trial volume, the signal moves deeper toward value.

PURE: judging channels on an agreed early ROAS window

Admiral Media ran user acquisition for the PURE dating app on US Android with the goal of competitive CPI and high D7 ROAS, and tested a programmatic DSP against an established self-attributing network with separate budgets. The PURE case study reports:

• -74% CPI: the headline CPI reduction.
• $2.44 versus $9.43 CPI: the programmatic DSP’s CPI compared with the self-attributing network in the same US Android test.
• D7 ROAS goals exceeded: the case study states that the D7 ROAS goals were exceeded, which led to expanded market launches.

The key detail is that both channels were judged against the same pre-agreed cohort window. Comparing a new channel’s day 3 numbers with an incumbent’s day 30 numbers is one of the fastest ways to make a wrong channel decision.

How to build your first useful cohort table

A useful first cohort table is small: weekly install cohorts as rows, days since install as columns, and one clearly defined value in every cell. Teams that start with a sprawling dashboard of every possible split usually end up trusting none of it, so the Admiral Media team recommends building one table well before adding more.

The build sequence that works in most accounts looks like this:

• Export the raw events, not the summary. Pull install date, acquisition source, platform, country and every revenue or subscription event with its timestamp from your product analytics or data warehouse. Summary exports from ad platforms already apply attribution windows and cannot be re-cut later.
• Calculate user age for every event. For each revenue or retention event, compute the number of days between the install and the event. This single column is what turns an event log into a cohort table.
• Pivot into a triangle. Put install week on the rows and user age on the columns, then fill each cell with cumulative revenue per install or the share of users still active. The result is the familiar triangle shape, with the newest cohorts showing the fewest filled columns.
• Add the cost column. Place cost per install for each cohort beside the triangle. Cumulative revenue per install divided by cost per install gives cohort ROAS at every age, and the column where it crosses 100% is the payback point.
• Add one split at a time. Once the blended triangle is trusted, repeat it for the largest channel, then the largest country, then the top creative concepts. Each new split should be justified by a decision it will inform.

Two quality checks save a lot of confusion later. First, reconcile the total revenue in the cohort table against store or finance reports for the same period; a large gap usually means missing refunds, currency handling problems or events that fire twice. Second, check that every cohort row includes users from the full week. Partial weeks at the start or end of a date range produce rows that look like outliers but are simply incomplete.

Reading cohort curves: the shapes that should change your plan

The shape of a cohort curve is more informative than any single cell. Two cohorts can share the same day 7 retention and still be worth very different amounts, because one flattens into a stable base of habitual users and the other keeps decaying toward zero.

Admiral Media’s teams look for four patterns when reviewing cohort tables:

• Flattening curve: retention drops in the first days and then levels off. This is the healthy pattern, and the level where it flattens is the base your LTV rests on. The goal of acquisition is to raise that floor, not just the day 1 peak.
• Steady decay with no floor: every period loses a similar share of what remains. This usually points to a product or value problem that more spend will amplify rather than fix. Admiral Media’s guide to retention-first user acquisition covers how to change targeting and creative when this pattern appears.
• Strong start, weak middle: a cohort retains or converts well early and then falls away faster than older cohorts. This is the classic signature of an optimization signal that has drifted: the algorithm has learned to find users who complete the early event but not the later behavior.
• Slow riser: cumulative revenue starts low and keeps climbing for weeks. This is common in freemium apps, and it is exactly the pattern a short ROAS window punishes. If a channel consistently produces slow risers, give it a longer evaluation window before cutting it.

Admiral Media also watches for newer cohorts consistently underperforming older ones at the same age. In most accounts we manage, that pattern traces back to one of three causes: creative fatigue pushing delivery toward cheaper and less intent-rich inventory, a broadening of targeting during scaling, or a change in pricing or onboarding that affected all cohorts from a specific date. Annotating cohort tables with the dates of creative launches, bid changes and product releases makes the cause far easier to find.

Cohort analysis on iOS and under privacy constraints

Cohort analysis on iOS is harder because user-level attribution is limited when users do not consent to tracking, so cohorts by campaign or creative have to be assembled from aggregated and delayed signals. The core principles still apply, but the data sources change.

Three practical adjustments help. First, run install and subscription cohorts from first-party and store data, such as App Store Connect’s download and subscription cohorts, which do not depend on ad attribution. Second, design the SKAdNetwork or AdAttributionKit conversion value schema so that it encodes the early behaviors your full cohorts show are predictive of revenue; Admiral Media’s AdAttributionKit measurement playbook covers the current framework. Third, validate channel-level conclusions with incrementality tests rather than relying on attributed cohorts alone, because attributed cohorts can overstate the contribution of channels that capture demand rather than create it.

Common cohort analysis mistakes and how to avoid them

The most common cohort mistakes come from comparing cells that are not truly comparable. Each one can produce a confident and wrong budget decision.

• Reading immature cells. A cohort that is 10 days old has no day 30 value. Tools sometimes show a partial number in that cell, and teams read it as a collapse. Mask immature cells.
• Mixing gross and net revenue. Store fees, taxes and refunds change cohort revenue materially. Pick one definition and use it everywhere, including in the ROAS targets you set on the platforms.
• Using the wrong calculation type. GA4 offers standard, rolling and cumulative calculations. Standard counts users active in a single period, rolling requires activity in every period up to that point, and cumulative shows a running total. Rolling retention will always look lower than standard retention, and comparing the two is meaningless.
• Splitting too thin. A cohort of a few dozen users per campaign per week produces noise, not signal. GA4 also applies thresholding to small demographic segments. Aggregate to weekly or to the concept level until each cell holds enough users to be stable.
• Ignoring seasonality. A cohort acquired during a holiday peak or a big promotion may behave differently from one acquired in an ordinary week. Compare like with like, or annotate the seasonal cohorts so they are not used as the baseline.
• Treating cohorts as a monthly report. Cohorts are most useful as a weekly operating view tied to specific decisions. Admiral Media’s overview of the app marketing metrics that matter shows where cohort metrics sit in a wider reporting stack.

From cohort curves to LTV and budget targets

Cohort curves are the raw material for LTV prediction, payback targets and ROAS goals. Once you have several months of mature cohorts, you can see what share of 12-month revenue a typical cohort has produced by day 7 or day 30, and use that ratio to set short-window targets the ad platforms can actually optimize toward.

That ratio is specific to your app, business model and pricing, so it has to come from your own data rather than an industry rule of thumb. The RevenueCat finding that freemium apps see a larger share of conversions in Week 6 or later than hard paywall apps is a reminder that two apps with identical day 7 numbers can have very different long-term value. Admiral Media’s guide to mobile app LTV walks through how to turn cohort curves into an LTV model, and the guide to predictive LTV bidding covers how to feed predicted value back to the platforms.

The loop that results is the core of a scalable acquisition program. Long-window cohorts calibrate the short-window target. The short-window target guides the algorithm. The algorithm changes the users you acquire. New cohorts show whether the change worked. Admiral Media runs this loop continuously across its client portfolio because each turn of it compounds: a slightly better signal finds slightly better users, which produce cleaner data for the next calibration.

How Admiral Media runs cohort analysis for client apps

Admiral Media builds cohort analysis into the operating rhythm of every account rather than treating it as a one-off audit. In practice, that means a weekly acquisition-source cohort review tied to budget moves, a monthly subscription cohort review tied to pricing and offer decisions, and a quarterly recalibration of optimization events, value rules and conversion value schemas against mature cohorts.

The Admiral Media team pairs this analysis with creative production, because cohort splits by creative concept regularly show that some concepts attract users who retain and pay while others attract cheap installs that fade. The Fastic case study shows what is possible when creative volume and performance data are connected: over Admiral Media’s work together with the fasting app, the case study reports +639% installs, +1,655% purchases, +439% revenue and +952% monthly active users, with a -70% cost per result through the AI Creative Factory, according to the Fastic case study.

If your blended numbers look stable but you are not sure whether the users you are buying today are as valuable as the ones you bought last quarter, a cohort review is the fastest way to find out. Admiral Media’s team can audit your cohort setup, your optimization events and your measurement on iOS and Android, and show where the short-window signal has drifted from the value your business depends on.

Frequently Asked Questions

What is cohort analysis for mobile apps?

Cohort analysis for mobile apps groups users by a shared starting event, such as the day they installed, the day they started a subscription or the campaign that acquired them, and tracks how each group behaves over time. It shows retention, conversion and revenue at the same age for every group, so you can compare new users with older users fairly. This reveals whether acquisition quality is improving or declining, which blended metrics like daily active users or total revenue cannot show. Most app teams use it to decide where to spend budget and which optimization events to use.

What is the difference between retention rate and cohort retention?

A plain retention rate is often calculated across all users at once, which mixes new and old users together. Cohort retention calculates the share of a specific group, for example everyone who installed in one week, that is still active on day 1, day 7 or day 30 after their own install date. Because every user is measured at the same age, cohort retention lets you compare one week’s users with another week’s users fairly. That makes it the right metric for judging campaign and channel quality.

Which cohort window should I use to judge a user acquisition campaign?

The right window depends on how quickly your users pay. RevenueCat’s State of Subscription Apps 2026 report finds that 50.6% of subscription app conversions happen on Day 0, but 19.2% happen in Week 6 or later, and the late share is higher for freemium apps than for hard paywall apps. Fast-converting apps can lean on day 7 results, while slow-converting or freemium apps need longer windows to avoid undervaluing good channels. The safest approach is to track full curves for several months and then choose the earliest window that reliably predicts long-term value for your app.

How many users does a cohort need to be reliable?

There is no universal minimum, but very small cohorts produce noisy results that swing from week to week for random reasons. If a campaign or creative split leaves only a few dozen users per cell, aggregate to a weekly or concept-level view, or combine similar campaigns, until the numbers stabilize. Analytics tools may also hide small segments; GA4, for example, applies thresholding to demographic dimensions. Stable cohorts are more useful than granular ones that nobody can trust.

Can I run cohort analysis on iOS without IDFA?

Yes, but the data sources change. Install and subscription cohorts can come from first-party product analytics and from App Store Connect, which offers download cohorts and subscription cohorts without relying on ad attribution. Campaign-level cohorts on iOS depend on SKAdNetwork or AdAttributionKit conversion values, so the conversion value schema should encode the early behaviors that your full cohorts show predict revenue. Incrementality tests help confirm channel conclusions that aggregated data cannot prove on its own.

How does cohort analysis improve ROAS?

Cohort analysis improves ROAS by showing which users actually produce long-term revenue, so you can change the signal the ad platforms optimize toward. Examples include bidding on a deeper event, adding value rules or introducing an LTV signal. In Admiral Media’s work with ChatPDF, testing target CPA against target ROAS, adding value rules and introducing an LTV signal contributed to a 320% year-over-year ROAS increase on Google Ads on an index baseline. The cohort data is what tells you which signal to send.

How often should I review app cohorts?

Review acquisition-source cohorts weekly, because that is the pace at which budget and creative decisions are made. Review subscription cohorts monthly, since renewals and plan changes happen on a monthly or annual cycle. Recalibrate optimization events, value rules and conversion value schemas quarterly against mature cohorts. Tying each review to a specific decision keeps cohort analysis from turning into a report that nobody acts on.