A small daily-math game on iPhone has accumulated 20,052 installs over roughly eighteen months. Today, the registered base is 9,625 ever-players, with about 470 daily active users, 970 weekly active, and 1,800 monthly active. Of those daily players, only 75 play every single day. That number -- 75 -- is the most important figure in this report.
This study takes one app's full back-end data, anonymizes it completely, and uses it to map something most product teams never see published: the actual paths users take from download to daily habit, the points at which they drop off, the gaps after which they return, and the mechanics that appear to be doing the work of bringing them back.
The findings, compressed:
Each of those numbers tells a story. The signup wall costs nearly half your installs at the door. One acquisition spike built two-thirds of the registered base, and the trickle that followed has not replaced it. Geographic concentration is producing real network effects domestically while leaving the app fragile globally. The Iron Core of 116 users is the only group where the product has visibly succeeded, and it is the proof point on which everything else depends. The Week 2 to Week 3 retention cliff is the design battleground that most product investment underweights.
What follows is a section-by-section walk through the data -- funnels, avenues, timelines, cohorts, comparisons -- followed by ten transferable lessons for any founder or product manager building toward daily habit.
Every install collapses into a smaller number at every stage. Here is the collapse, exactly.
The cliff sits between "ever played" and "active in the last 30 days." 7,825 registered users have played the app at some point but did not play in the most recent month. That cohort, the lapsed but registered -- is what the rest of this whitepaper is really about.
Two more numbers worth holding next to each other: 94% of users who created an account played at least one game (an unusually high activation rate, well above the typical 60–80% for casual mobile games), and 0.37% of installs play every single day. The product is excellent at converting registration into a first session, and very narrow at converting first sessions into daily habit. Those are two different problems with two different solutions.
Every install ends up on exactly one of seven paths. Here are the volumes on each.
Aggregate funnel numbers hide the fact that users follow distinct routes through the product. Some are wins. Some are clean churns. Some loop back. Mapping which path each install takes is more useful for decision-making than the totals.
9.0% of installs. Created an account, played, and are still active in the last 30 days. This is the cohort the product is built for. Approximately 970 of them played in the last 7 days; 75 play every single day.
39.0% of installs. They registered, they played, they stopped. The app is still on their phone. This is the largest single addressable segment in the dataset and the entire focus of any re-engagement programme.
Approximately 1.9% of installs and roughly 4% of registered ever-players. Lapsed for at least a week, came back, and played again in the last 30 days. The largest documented gap-and-return is 21 days. This avenue is the proof that re-engagement works at meaningful scale.
3.1% of installs and 6% of accounts. Made the leap to register and then never started a single game. The friction sits between account creation and first-game render -- likely a permissions ask, an onboarding screen, a notification prompt, or a time-of-day that didn't suit the user. A high-leverage place to instrument and fix.
9.5% of installs. Started in guest mode, played, and at some later point converted to a registered account. The strongest argument for guest tiers in the entire dataset -- these users would have been blocked at the door by a hard signup wall.
39.4% of installs. Started in guest mode and either (a) are still playing as guests, invisible to per-user analytics, or (b) silently churned. Roughly half of the daily active number (~225 of ~470) lives here. The single largest blind spot in the whole product.
24.1% of installs. Opened the app, decided no, removed it. Most were probably wrong-fit users delivered by acquisition; some were right-fit users who hit a friction (permissions, onboarding, first word too hard, performance) that knocked them out before the product could prove itself. This figure overlaps with the account/guest split -- some of these users uninstalled after choosing.
Two patterns are worth noting. First, more installs end up on Avenue 6 (39.4%) than on the dream path (9.0%). The guest-mode cohort is the single largest population in the product, and the one with the least visibility. Second, Avenues 2, 3, and 5 -- every "lapsed and recoverable" or "delayed activator" path -- together account for roughly 50% of all installs. Half of every install delivered is, in some form, recoverable.
When users drop off matters more than that they drop off. Here is the timeline.
Cohort retention by week, for the six most recent install weeks where data is available:
| Cohort week | New users | Week 1 | Week 2 | Week 3 | Week 4 |
|---|---|---|---|---|---|
| Mar 15–21 | 49 | 24% | 18% | 14% | 8% |
| Mar 22–28 | 33 | 30% | 30% | 21% | 18% |
| Mar 29–Apr 4 | 29 | 34% | 31% | 24% | -- |
| Apr 5–11 | 22 | 45% | 32% | -- | -- |
| Apr 12–18 | 22 | 27% | -- | -- | -- |
| Apr 19–25 | 46 | -- | -- | -- | -- |
Three things stand out.
First, the spread is wide. Week 1 retention runs from 24% to 45%, a near-doubling between cohorts only weeks apart. Something is materially different about the Apr 5–11 cohort. The most plausible explanations are channel mix (a different acquisition source delivered higher-intent users), a featured placement, or word-of-mouth from an existing power user. Knowing which is the difference between repeating it and not.
Second, the cliff is between Week 1 and Week 2. Across cohorts, average drop is ~30% of returning users lost in the second week. By Week 4, retention has compressed to single digits in the worst cohort and high teens in the best. This is the design battleground.
Third, the absolute weekly cohort sizes (22–49 new registrations per week) are tiny relative to the May 2025 spike (6,450 in one month). Organic acquisition is a trickle. The retention work is the right work, but it is being done on a small inflow.
Lapsed users come back at meaningful scale. Here is when, and what shape it takes.
The dataset documents 378 comeback users -- registered players who lapsed for at least 7 consecutive days and returned to play in the most recent 30-day window. Several patterns from this cohort are worth surfacing.
The average comeback user has played approximately 130 of the last 180 days. These are not casual one-touch users -- they are committed players who slipped, then recovered. The biggest documented gap-and-return arc in the dataset is 21 days: a player who went silent for three weeks and is now back and active.
The shape of comebacks suggests a power-law distribution. Most returns happen after a short gap (1–3 weeks); a long tail of returns happens after gaps of a month or more, but at much smaller volume. The implication for re-engagement strategy is precise: the highest-leverage moment for a re-engagement push is between days 7 and 14 of inactivity, when the user has lapsed enough to need a nudge but not so long that the habit muscle has atrophied.
What this implies for product investment: a re-engagement push notification series timed at days 7, 10, and 14 of inactivity, calibrated separately by streak length at lapse, would intercept the largest portion of the comeback population. A second-tier "we miss you" push at day 21 is worth running, but with the understanding that it is recovering single-digit percentages of the lapsed pool, not double digits.
One month built two-thirds of the registered base. Everything before and after has been a trickle.
In May 2025, this app added 6,450 new registered players in 30 days. That single month accounts for 67% of the entire current registered base. The months before and after run between 30 and 300 new registrations each, with a long-tail decline back toward 30–80 per month in the post-spike period.
The most plausible explanation for May 2025 is calendrical. The app is a daily-math product, concentrated in Ireland (see Section 08), and May is the run-up to the Irish Leaving Certificate, the national university entrance exam. A school-channel campaign or a teacher-driven moment of word-of-mouth at exam time fits the data better than any other hypothesis.
Two implications. First, single-spike acquisition is both miracle and risk. A founder reading the dashboard in June 2025 might reasonably have concluded the app had achieved escape velocity. By July, the trickle had returned. The May cohort was a gift, not a flywheel, and a year later, the per-month numbers suggest the work to build a repeatable acquisition system has not yet succeeded.
Second, a spike-cohort decays differently than a steady-inflow cohort. Without per-cohort retention curves on the May 2025 group specifically, this cannot be proven from the data exposed. But the size of the comeback population (378) and the shape of the dipped-in tier (966 users with 1–6 days of activity in 180 days) both strongly suggest that a meaningful portion of the original 6,450 are now in the lapsed-but-recoverable bucket -- having touched the app during exam season, drifted, and returned occasionally.
The strategic question for any app facing this situation is not whether to celebrate the spike. It is whether to optimise the post-spike cohort for re-engagement, or to put the same energy into engineering a second spike. Both are legitimate. Doing neither is the only wrong answer.
Users sort into five clear tiers by how often they play. Roughly 1 in 20 active users carries the entire load.
Looking at the last 180 days of registered-user activity, the population sorts into a clean pyramid:
The Iron Core (116 users · 1.2% of registered) is the only group where the product has visibly succeeded. They have integrated the app into their daily life -- 150+ days played out of 180, with an average longest gap of just 3 days. They are the proof point that, for the right user, the product builds genuine habit. They are also the canary: if these 116 users show signs of softening, that's a P0 signal that something material has broken in the product loop.
The Power tier (111 users) are the highest-leverage cohort to study. They play 3–4 days a week. Their average longest gap is 7 days. They almost made it to daily but didn't. Identifying what slipped, a Tuesday meeting, a phone-down weekend, a streak that broke and felt unworth restarting -- is, on a per-user basis, the highest-leverage product question on the board. Promoting Power into Iron Core is a doable move; it doubles the daily-habit base.
The Regular and Casual tiers (combined ~786 users) are the broad middle. They are healthy users -- they play more than they don't, but they are not yet the product's primary audience. Many of them have a particular reason for using the app: a study habit, a morning ritual, a streak they want to protect. Most product investment should optimise for moving them up to Power, not for converting them to daily users overnight.
The Dipped In tier (966 users · 49% of 180-day-active) is the largest single cohort. They played the app 1–6 times in the last six months and stopped. For most, the product was probably wrong, or arrived at the wrong moment. Re-engagement campaigns work on a fraction of this group, but the deeper lesson is upstream: are the channels delivering Dipped In users the same channels delivering Iron Core, or different? If they are the same, acquisition needs to filter harder. If they are different, scale the Iron Core channels and starve the Dipped In ones.
Your average user does not exist. 1.2% of the base plays nearly every day. 49% touched the product and left. The arithmetic mean of those two populations is a fiction. They are different people with different relationships to the product, and the work of growth is to design for each separately.
82% of active users sit in one country. Ireland is the moat and the single point of failure.
| Country | Active users | Registered | Est. guests | Share |
|---|---|---|---|---|
| Ireland | 14,987 | 5,098 | ~9,889 | 82% |
| United States | 1,900 | 1,954 | -- | 10% |
| United Kingdom | 1,400 | -- | ~1,400 | 8% |
| Spain | 633 | -- | ~633 | 3% |
| Portugal | 244 | -- | ~244 | 1% |
| France | 207 | -- | ~207 | 1% |
| Australia | 192 | 23 | ~169 | 1% |
This is concentration most apps would kill for in a single market. Network effects compound -- friends, classmates, family members all use the same app, share the same scores, talk about the same daily word. The viral coefficient inside Ireland is doing real work that no paid acquisition could replicate at the price.
It is also a fragility. A single algorithm change at the App Store, a single Irish education-system shift, a single competitor launching in Dublin -- any of these is a percentage of usage at risk that would be a footnote in a more diversified base. Concentration is a strategy until the day it is a vulnerability, and the data does not signal which day that will be.
The Ireland number also conceals the largest single growth opportunity in the entire dataset. Of the 14,987 active Irish users, approximately 9,889 are guests. If the conversion rate from guest to registered (currently 19.4% across the full base) were applied uniformly to the Irish guest pool, that is approximately ~1,920 net new tracked, push-able, retainable users available without spending a euro on acquisition.
The non-Irish markets tell a different story. The United States is 10% of active users with 1,954 registered users -- close to a 1:1 ratio of registered to active. Outside Ireland and the US, almost every user is a guest. The product has reach in the UK, Iberia, France, and Australia, but no relationship there. Closing that gap is a question of localisation, payment readiness, and whether the daily-word format ports cleanly across language and curriculum.
Roughly 8 in 10 registered users have engaged with the leaderboard or group features. The other 2 in 10 play alone, and they look very different.
The dataset shows 103,300 leaderboard views generated by 7,420 users. That averages 14 leaderboard views per leaderboard-using player, a heavy social usage pattern. Of the 9,625 registered ever-players, that means roughly 77% have engaged with groups or leaderboards, and roughly 23% (~2,205 users) have not.
The data exposed in the dashboard does not split the comeback, retention, and streak metrics neatly across these two segments -- that level of segmentation requires querying the underlying Firestore directly. What follows is a comparison framework based on the patterns that are visible, and the strong inference each enables.
| Metric | In-group / leaderboard users | Solo players |
|---|---|---|
| Population | ~7,420 (77%) | ~2,205 (23%) |
| Average leaderboard views per user | 14 views | 0 views |
| Likely share of Iron Core | Very high. Iron Core users typically benchmark daily | Low -- solo Iron Core users exist but are the exception |
| Likely share of Comeback users | Higher -- group activity creates external pull | Lower, no external trigger to return |
| Likely D7 / D30 retention | Materially higher | Materially lower |
| Push notification efficacy | High -- group-driven nudges resonate | Lower, only product nudges exist |
| Churn risk on streak break | Lower -- group accountability acts as a brake | Higher, no external accountability |
Even without per-segment retention numbers, the structural argument is strong: group-attached users have more reasons to return after a gap than solo users do. A solo user who lapses receives only product-driven re-engagement signals (push notifications, email, the icon on their phone). A group-attached user receives those plus social signals, a friend's score appearing on the leaderboard, a group-chat message about today's word, an explicit prompt from a family member. The data shows 103,300 leaderboard views; that is 103,300 individual moments where a user's social environment reminded them the app exists.
The corollary is sharper: solo players are over-represented in churn and under-represented in comebacks. The 2,205 solo registered users are likely sitting heavily in the Dipped In and lapsed-not-recovered cohorts. They are also the population where the product has the least leverage -- there is no friend-of-a-friend to bring them back, and there is no sense of being missed.
Three actions follow naturally:
First, treat group joining as a mid-funnel activation event, not a feature. A user who joins a group is materially more likely to retain. Move the group-creation prompt earlier in the user journey, particularly during the first week, and measure the lift in D7 and D30 retention for users who joined in their first session vs. users who never did.
Second, design a re-engagement programme specific to solo players. Push copy that works for group users ("your friends just played") will not land for someone with no friends in the app. Solo-player re-engagement needs intrinsic hooks, a streak the product itself remembers, a personal best to beat, a content drop ("today's word is hard").
Third, use the group-vs-solo split as a leading indicator of acquisition channel quality. If a particular acquisition channel delivers users with a higher group-join rate in their first 7 days, that channel is delivering higher-LTV users. This is a more useful per-channel quality metric than D1 retention alone.
A small population of long-streak users carries an outsized share of the daily activity, and is also the highest-churn-risk cohort on any single day.
Currently active streaks (users who played today or yesterday with a 1+ day current streak) total 315. Of those:
The 11 users with 100+ day streaks are the elite of the elite -- players who have shown up nearly every day for at least three months. They overlap heavily with the Iron Core and almost certainly with the in-group cohort. They are also, paradoxically, the highest-churn-risk users on any given Tuesday: a single missed day can trigger a streak break that the user perceives as too costly to restart, leading to a complete walk-away.
This is where streak-shield mechanics earn their keep. A shield that auto-saves a streak after one missed day, paired with a notification that explicitly tells the user the shield was used, converts a potential churn moment into a re-engagement moment. Without per-event data on shield usage, this cannot be proven from the dashboard. But it is the single mechanic most consistent with the comeback pattern documented in Section 05.
378 users went silent for at least a week and came back. Without them the Iron Core would be smaller, the daily active number would slip, and the trickle of new monthly registrations would not be enough to maintain the base.
The single most operationally interesting cohort in the entire dataset is the comeback population. 378 users -- roughly 4% of registered ever-players, roughly 21% of the monthly active number -- had a gap of 7+ days and returned to play in the most recent 30 days. They are not new users acquired from a channel. They are recovered users that the product already had.
Three things make this cohort important.
First, the cost to recover them is materially lower than the cost to acquire a new user. A push notification costs effectively nothing per send. A deep-linked share message costs nothing. The user already knows the product; the friction is internal, not external.
Second, comeback users have higher lifetime value than fresh installs. The data shows comeback users average ~130 days played in the last 180 -- they are committed players who slipped, not casual players who tried once. Their retention curve, returned to its pre-lapse shape, is meaningfully higher than that of a fresh install.
Third, the comeback population is a free signal on what the re-engagement system is doing. Tracking comeback rate over time, by gap length, is the cleanest single number for measuring whether re-engagement investment is working. It is more honest than D7 retention (which depends heavily on cohort mix) and more leveraged than DAU (which depends on acquisition).
The data lets us frame, not prove. Each hypothesis below is testable with one query against the underlying export.
Many word and math apps run a shield mechanic that auto-saves a streak after one missed day. When the shield consumes, users typically receive a notification ("Your shield saved your 47-day streak -- don't lose it again"). This creates a strong psychological pull to return the next day. Test: correlate shield-consumed events with app-opened events within 48 hours.
The data shows 103,300 leaderboard views across 7,420 users, an average of 14 views per leaderboard-using player. This is enormous social-comparison volume. When a friend posts a score, plays today's word, or appears on the leaderboard, lapsed users in their network are likely pulled back. Test: check whether comeback users belong to multi-member groups at higher rates than the dipped-in cohort.
The simplest, most boring, and probably the most operationally important. A consistent daily push at the user's habitual play time will catch a percentage of the lapsed every single day. Test: segment comeback events by app-open source equal to push-notification.
The dataset documents 9,500 share events. A friend who texts "did you do today's one. I got it in 3" is one of the most powerful re-engagement nudges available, and it is happening at scale here. Test: correlate comeback events with shared-result link clicks within the same day.
Math and education apps see exam-season returns. The dataset's biggest cohort came in May 2025 (Irish exam run-up). Some of the comeback users are likely students who left after May 2025 exams and returned the following spring as the same exam season approached again. Test: plot comeback events against academic calendar dates and look for correlation.
The honest version of all five hypotheses: with the data accessible to me, I can frame each but cannot prove them. Each is testable in roughly thirty minutes against the right export. The recommendation is to run them in sequence and assign re-engagement budget proportionally to whichever hypothesis the data validates.
For other founders and product managers reading this -- here is what one real, mid-scale education app looks like at each milestone.
| Milestone | This app | Typical industry range |
|---|---|---|
| Install → account creation | 51.1% | 30–50% |
| Install → guest play | 48.9% | n/a (most apps don't offer guest) |
| Guest → registered (eventual) | 19.4% | Underreported, no clean benchmark |
| Account → ever played | 94% | 60–80% (this app is strong) |
| Account → active last 30 days | 17.6% | 10–25% |
| Install → uninstall | 24.1% | 25–40% (this app is below average -- good) |
| Week 1 retention (range) | 24–45% | 20–30% for casual games |
| Week 4 retention (range) | 8–18% | 5–10% for casual games |
| Top country share of active | 82% | <30% for global apps |
| Daily-habit core (% of registered) | 1.2% | 0.5–2% |
| Group / social engagement (% of registered) | 77% | Highly variable -- depends on whether social is core |
| Comeback rate (lapsed → returned in 30d) | ~4% of registered | 3–8% with active re-engagement |
The retention numbers are strong for a casual education app. The activation rate (94%) is excellent. The geographic concentration is unusual. The acquisition spike-then-trickle pattern is common but rarely admitted in public.
Ten transferable lessons. None depend on the specifics of this app.
In the spirit of honesty.
This dataset has limits. The dashboard exposes aggregate cohort views and summary statistics; per-user event-level data is partially obscured and would require direct Firestore access for some of the deeper segmentations.
Specifically, the data does not expose per-user acquisition source -- there is no way to prove from this analysis which channel drove the May 2025 spike, or which channel sustains the post-spike trickle. Channel-quality conclusions in this report are inferred, not measured.
The data does not expose monetization information. If the app has in-app purchases, advertising revenue, or a subscription, none of it is visible here. This is purely an engagement analysis, not a unit-economics one.
The data does not expose session length, time-of-day patterns at user level, or device-level performance characteristics. Each would change some of the lessons above.
Finally, this is one app. The whole point of this report is to translate one app's truth into transferable lessons, but transfer always loses fidelity. A founder reading this should not assume their numbers will look like these. They should assume the questions apply, even if the answers do not.
116 users played a small math app on their phone 150 of the last 180 days. They didn't have to. There was no contest, no prize, no pay. Whatever the app gave them, they wanted enough to come back almost every day for half a year.
That is the only number on this whitepaper that ultimately matters. Everything else, the funnel, the avenues, the cohorts, the spikes, the geography, the comebacks -- is in service of understanding how to build, support, and grow that 1.2% of users who quietly demonstrate the work is real.
The other 98.8% are the road to get there.