Retention Analysis: How to Improve User Retention Rate

Summary: Whether retention is healthy is judged by whether the curve flattens, not by a single percentage. This article gives the three definitions of retention and their formulas, the four curve segments (cliff, long slope, plateau, tail) with their root causes and remedies, the benchmarking baseline and the principle of "compare against your own previous cohort," and eight measurement disciplines — including defining active as a value action rather than a login, waiting for the cohort to mature before concluding, segmenting before fixing, and noting that systematic retention work shows initial results in 60 to 90 days.

Retention is the most easily misread metric. One team stares at a D30 retention of 18% and tries everything to lift it; another team with the same 18% is growing fast — the difference is curve shape: the first keeps sliding, the second flattens by week three. A single number lies; the curve doesn't.

The three numbers of retention are also constantly conflated: installed but not deleted, opened, actually in use — these are three different things, mapping to three completely different kinds of churn. If you can't tell them apart, you can't talk about improvement.

Key Takeaways

Three types of churn must be separated: install churn, engagement churn, subscription churn; mixing them yields wrong conclusions.
The only test of health is a flattening retention curve, and each new cohort at least no worse than the last.
Benchmarks are only an interval reference; what you should really compare is your own historical cohort.
The four curve segments each map to a different root cause: cliff is an activation problem, long slope is a habit problem, low plateau is a product-market-fit problem, no uptick needs a revival.
Define active as a value action, not a login; align the time scale with the product's natural usage frequency.
Systematic retention work shows initial results in 60 to 90 days; full effects take 6 to 9 months.

I. First, Separate the Three Definitions

The basic retention formula is: retention rate = (ending users − new users during the period) ÷ starting users × 100. For example, starting with 2,000 users, adding 400 during the period, ending with 2,200, the retention rate is 90%, meaning 10% churn.

But the real error-prone part is "what counts as churn." Three types of churn must be counted separately:

Install churn: deleted after install or never opened again, usually measured in D1, D7, D30 windows.

Engagement churn: opened, but produced no valid behavior. A user opens the fitness app on day one, logs nothing, and never returns by day seven — by install retention he counts as "retained," by engagement retention he has already churned. This type is the easiest to miss precisely because it looks like retention.

Subscription churn: cancels or fails to renew after paying; the most expensive churn for subscription businesses.

How you define "active" determines the quality of every number. Treating login as active makes the curve look much better, but you can diagnose nothing. Active should be a value action: completing a task, producing content, finishing a transaction — not just opening the app once.

II. The Curve Matters More Than the Number: The Only Health Test Is Flattening

A retention curve is the share of the same group, grouped by signup date, still active at each time point. It starts at 100% and goes down; the shape comes in three kinds:

Slides all the way to near zero — the worst case. Means the product is a leaky bucket with no stable core formed; user lifetime value has a ceiling. In this case, spending more on acquisition just accelerates the leak.

Drops to a position then flattens — healthy. Where it flattens is your stable core, the group that keeps delivering value. This is the only reliable test of retention health.

Drops, flattens, then rises — a smile curve; churned users come back because the product got better. Rare and very strong, but don't count on it.

A good curve also has comparable traits: consumer apps' D1 sits between 40% and 60%; the drop from D1 to D7 is less than half of D1; D7 to D30 stabilizes; D30 to D90 is basically flat. If the curve is still sliding after D30, you haven't found the people who truly need you.

III. How to Use Benchmarks: Reference Bands, Self-Comparison First

Industry benchmarks can be referenced by category. The median D30 install retention is roughly: productivity tools 15% to 25%, social 10% to 20%, casual games 3% to 8%, health and fitness 10% to 20%, photo and video 5% to 12%. Subscription monthly churn: consumer subscription 5% to 12%, B2B 2% to 5%, streaming content 5% to 10%; annualized, consumer subscription 30% to 50%, B2B 10% to 30%.

There are also reference bands on the composite benchmark: consumer subscription "good" is about 40%, "excellent" about 70%; SMB and mid-market SaaS about 60% and 80%; enterprise about 70% and 90%.

But the biggest use of benchmarks is to rule out order-of-magnitude errors, not to set targets. The real question to ask is: did this cohort retain better than the last one? Each new cohort at least no worse than the last means improving; better than your own last quarter means improving. A good retention for a daily messaging tool is a disaster on a quarterly tax-filing product — the product's natural usage frequency sets the reasonable band, and comparing numbers across categories is meaningless.

One practical correction in recent years: the measurement start need not be the signup month. If the product has a free tier or low-cost trial, early cohorts contain many one-time-triers; retention anchored on the signup month mostly measures "how many tourists you attracted," not product quality. The safer move is to rebase at month 3 and look at month 12 retention relative to month 3 — the curve usually begins to flatten around month 3, and the flattening position predicts long-term retention. Accordingly, acquisition cost should also be counted as "cost per user still retained at month 3," not signup cost.

IV. Four-Segment Diagnosis: Where the Problem Sits

Where the curve falls points directly at the root cause; handle the four segments separately:

Low at D1 — activation problem. Onboarding too long or confusing; promise vs. experience mismatch; crashes on first launch. Compare D1 against the category median to see if it's low, and check crash rate by OS and device model.

Steep drop from D1 to D7 — the first session had no "aha" moment; the user doesn't know why they need it; or the paywall came too early, hitting before they wanted to pay; or there's no reason to return — no notifications set, no goal set, no habit formed.

Continuous slide from D7 to D30 — content or service depth insufficient; users exhausted the obvious value and see no more; poor notification quality (none, or so aggressive people unsubscribe); missing growth and progression signals, users hit a ceiling.

Slow decay after D30 — subscription billing surprise; engagement loop broken (content cadence, social hooks, new challenges all gone); or pulled away by a competitor.

One discipline that must be emphasized: segment first, then fix. Deterioration in the mixed curve is almost always concentrated in one segment or channel. Split by channel, plan, user profile, platform before acting, or you'll usually fix the wrong block.

V. Intervention Prescriptions: Act by Curve Region

Cliff (early plunge) → fix activation. Shorten the path from signup to first value: focused product onboarding, a 3-to-5-item new-user checklist, and fewer steps before the "aha." Keep onboarding to 30 to 90 seconds, at most 2 minutes; getting users to first value within 60 seconds can lift D1 by 30% to 50% above the median. Each extra onboarding step drops conversion 5% to 10%; cut redundant steps ruthlessly.

Long slope (continuous slow decline) → build habit. Use in-app announcements and contextual nudges to surface features users haven't discovered, and attach the product to a recurring workflow rather than a one-time task. Adding 2 to 3 personalized questions (goal, level, preference) typically lifts D7 by 5% to 15%.

Low plateau (stable core too small) → listen to feedback. Those who stay know why; those who leave told you on the way out. Run satisfaction surveys, read detractors' verbatim, and put high-frequency themes into the roadmap. Also note: request notification permission after first value, not before — opt-in can rise from about 25% to over 50%, and users with notifications on retain 30% to 60% better.

No uptick (churned, never returns) → run revival. Dormant users aren't lost. Telling them about relevant new features via in-app notification on their next visit is the cheapest re-acquisition channel. Targeted win-back when inactive on day 3 typically recovers 5% to 15% of the churned-in-users.

The most restraint is needed with push. It is the single strongest retention channel — users who received at least one push within 90 days retain 3x those who didn't; but 6 or more per week raises uninstall risk 3.4x. The premise is always segmentation and frequency discipline.

One more supplementary metric worth watching: stickiness (DAU ÷ MAU). Healthy values: social above 50%, productivity tools 20% to 30%, subscription 15% to 25%. It is finer-grained than D30 — the curve may look fine while stickiness reveals a user who opens only a few times a month.

VI. Measurement Discipline and the Cadence of Results

Eight disciplines, in order of importance:

Build cohorts by signup date, not a blended full-population curve; define active as a value action, not a login; match the time scale to the product's natural rhythm (consumer watches D1/D7/D30, B2B watches weekly or monthly cohorts); overlay old vs. new cohorts to see if the curve flattens at a higher position; segment by channel, plan, profile, platform; wait for the cohort to mature before concluding (a two-week-old curve shows no shape); study the plateau group — they define your best-fit audience; pair the curve with qualitative feedback.

Conversely, these are wrong: using login as active to flatter the curve; applying daily retention to a weekly-frequency product; directly comparing benchmarks built on different methodologies; increasing acquisition on a still-declining curve; staring at the D30 single point while ignoring shape deterioration; fixing without segmenting.

Set expectations on cadence: systematic retention work usually shows initial results in 60 to 90 days and full effects in 6 to 9 months. First 30 days build measurement and baseline; second 30 days build interventions (push plan, lifecycle sequence, onboarding revamp); third month enters the optimization loop. D1 and D7 can rise 2 to 5 points in the first month; D30 and D90 need 4 to 6 months, because new cohorts need time to mature.

FAQ

Q: What retention rate counts as good?

A: No universal number — it depends on the product's natural usage frequency. You can reference the category D30 medians (productivity tools 15–25%, social 10–20%, casual games 3–8%) and composite bands (consumer subscription good ~40%/excellent ~70%, enterprise 70%/90%), but benchmarks only rule out order-of-magnitude errors. What you should really compare is your own historical cohort — whether the new cohort is at least no worse than the last; better than last quarter means improving. The more reliable test is whether the curve flattens.

Q: Why is the business worsening when the retention rate hasn't changed?

A: Because a single point masks shape deterioration. With D30 holding at 18%, the curve may have shifted from "flattening by week three" to "continuously sliding" — the stable core is shrinking, just not yet reflected at that point. Likewise, using login as active makes the number look good but kills the diagnosis. So you should "report the retention rate, but do the work on the curve," and watch it together with stickiness (DAU/MAU).

Q: New users drop sharply in the first three months — is the product broken?

A: Not necessarily; see whether it flattens after the drop. If the product has a free tier or low-cost trial, early cohorts contain many one-time-triers; this steep drop is the price of a "wide entrance." The method is to rebase at month 3 and look at month 12 retention relative to month 3 — the curve usually begins to flatten around month 3, and the flattening position reveals the product's true retention ability. Acquisition cost should also be counted per user still retained at month 3.

Q: Which time window should I look at first?

A: Look segment by segment, not pick one window. Low D1 is an activation problem (onboarding, promise consistency, first-launch crash); D1–D7 steep drop is no "aha" in the first session or paywall too early; D7–D30 slide is insufficient depth, notification quality, and growth signals; slow decay after D30 is billing surprise, broken engagement loop, or competitor pull. Locate by segment first, then act — fixes on the blended curve often hit the wrong spot.

Q: How many push notifications should I send?

A: It is the single strongest retention channel and the easiest to overuse. Users who received at least one push within 90 days retain 3x those who didn't; but 6 or more per week raises uninstall risk 3.4x. The principle is behavior-based segmentation, strict frequency control, specific content (with the user's real behavioral data), and placing the permission request after the user's first value, not at launch.

Q: How long until retention work shows results?

A: Initial results in 60 to 90 days; full effects take 6 to 9 months. First 30 days build measurement, baseline, and segmentation; second 30 days build interventions (push plan, lifecycle sequence, onboarding revamp); third month enters the optimization loop. D1 and D7 can rise 2 to 5 points in the first month; D30 and D90 need 4 to 6 months, because new cohorts must mature before the curve shape can be read.

About Tengence

The hard part of retention analysis is not the formula, but consistent definitions and segmentation discipline: how active is defined, how cohorts are grouped, when conclusions may be drawn — these decide whether every later action is valid. Tengence (TENGENCE) focuses on the growth technology of enterprise public-domain traffic acquisition and private-domain conversion, and the user segmentation, behavioral data, and retention measurement covered here are a direction it continuously follows.

Data Sources

Mobile App Retention: Benchmarks, Curve & 7 Tactics——Digipeak: onboarding 30–90 seconds / at most 2 minutes; first value within 60 seconds lifts D1 by 30–50%; each extra step drops conversion 5–10%; Airship study of 63 million users (received push 3x retention, 6+ per week 3.4x uninstall risk); stickiness benchmark; results cycle 60–90 days / 6–9 months.
Mobile App Churn and Retention: A Working Playbook (2026)——ASOhack: three churn definitions; D30 category medians and subscription monthly/annual churn; four traits of a good curve; D1 / D1–D7 / D7–D30 / D30+ four-step diagnosis; four intervention quantifications (personalization +5–15% D7, opt-in 25%→50%+, soft/hard paywall, day-3 win-back 5–15%).
User retention: how to measure and improve it——Bettermode: retention formula and worked example; curve should flatten, new cohort no worse than the last; six improvement levers.
How to Improve User Retention for Your Startup——DEV Community: flattening is the only test; smile curve (a16z September 2025 ChatGPT monthly retention case); Lenny Rachitsky benchmark band; rebase at M3 to view M12/M3, "AI tourists" and cost counted on M3 retention.
Retention Curves: How to Read, Benchmark & Bend Them (2026)——Kompassify: four-segment curve prescriptions (cliff / long slope / low plateau / no uptick); eight Do's and eight Don'ts; curve vs. retention rate difference; acquisition cost 5–25x retention, Bain +5% retention brings +25–95% profit.