Why Averages Lie
Imagine your product had thirty percent retention last month. Is that good? Is it getting better or worse? Is it the same for every user, or are some users sticking around while others churn quickly? The single average number answers none of these questions. It might be hiding everything important.
Cohort analysis is the discipline of breaking aggregate metrics into groups of users who share a common characteristic (usually when they joined) and tracking each group over time. Where averages produce one number, cohorts produce a story: how each batch of users behaved, whether new batches behave better or worse than old ones, where users drop off and when.
This article is about cohort analysis as a PM practice. What cohorts are, how to read retention curves, what different curve shapes mean, and the common mistakes that make cohort analyses misleading. We will not turn you into a data analyst. We will give you enough understanding to read cohort data and act on it.
What a Cohort Is
A cohort is a group of users who share something in common, usually when they joined or when they first did something. The most common cohorts are signup cohorts: all users who signed up in January 2024, all users who signed up in February, and so on. Other cohorts are also useful: users who first paid in a given month, users who first hit a milestone, users in a specific segment.
Why Cohorts Matter
Users who join at different times have different experiences. The product they signed up for in 2022 is not the product available in 2024. The acquisition channel, the onboarding flow, the feature set, the competition all changed. Comparing aggregate retention across these different experiences is confused; comparing cohorts to each other reveals what changed.
A Simple Example
Suppose your product has fifty percent month-over-month retention overall. That sounds fine. Now break it into cohorts. The January cohort has eighty percent retention. The February cohort has sixty percent. The March cohort has thirty percent. The aggregate is the average, hiding the trend. The cohorts reveal that retention has been collapsing for new signups for three months. The product has a serious problem that the aggregate hides.
Reading a Retention Cohort Table
The most common cohort view is a retention table. Each row is a signup cohort. Each column is the number of months (or weeks or days) since signup. Each cell shows the percentage of that cohort still active in that period. Modern analytics tools (Amplitude, Mixpanel, Heap, others) produce these tables with one or two clicks.
A Simplified Example
Cohort Month 0 Month 1 Month 2 Month 3
Jan 2024 100% 60% 45% 40%
Feb 2024 100% 62% 48% 42%
Mar 2024 100% 64% 50% —
Apr 2024 100% 68% — —
May 2024 100% — — —
Reading this table tells a story. Month 0 is always 100% (everyone in the cohort is, by definition, active at signup). The numbers in subsequent months show how many are still active. Empty cells mean we don't have data yet because that month hasn't happened for that cohort.
What This Table Tells You
Looking down a column tells you whether new cohorts are retaining better than old ones. In Month 1, retention went from 60% to 68% across the cohorts. That's improvement; each new batch of users sticks better than the previous batch. Something the team has done (better onboarding, better acquisition channels, better activation) is producing results.
Looking across a row tells you the shape of one cohort's retention curve. The January cohort dropped sharply in Month 1 (40% loss) but the loss slowed in Months 2 and 3. The curve is flattening, which is the most important shape to look for.
What Curve Shapes Mean
- Flattening curves are healthy. After initial churn, the remaining users stick. The flat part of the curve is your loyal user base.
- Curves that decline to zero are concerning. Every cohort eventually loses everyone. No loyal base forms. The product is not retaining anyone long-term.
- Curves that improve over cohorts (newer cohorts retain better) suggest the product is improving. This is the strongest positive signal.
- Curves that get worse over cohorts (newer cohorts retain worse) suggest the product is deteriorating. This is one of the earliest warning signs and is often missed because aggregate retention lags.
Why Retention Compounds
Retention is not just one metric. It compounds. Small improvements in retention have outsized effects on long-term outcomes because retention is multiplied month after month.
A Simple Math
Imagine two products. Product A has 90% monthly retention. Product B has 95%. The difference seems small. Over twelve months, Product A retains 28% of users; Product B retains 54%. Over twenty-four months, Product A retains 8%; Product B retains 29%. The five-point difference compounds to nearly four times the long-term user base.
Why This Matters
Most teams spend more effort on acquisition than on retention. The math suggests they should reverse the priority. A small retention improvement is worth a much larger acquisition improvement, because retention multiplies. The strategic insight from cohort analysis is often: the team should be spending more time on retaining the users it has rather than acquiring more.
Segmenting Cohorts
Beyond signup-date cohorts, you can segment cohorts by other characteristics. This often reveals where retention varies dramatically across user types, which is more actionable than aggregate retention.
By Acquisition Channel
Users from different channels often retain very differently. Organic users typically retain better than paid ones. Users from referrals often retain best. Users from a particular campaign or partner may retain extremely badly. Without segmenting, these patterns are invisible.
By Activation Path
Users who completed onboarding versus those who didn't. Users who tried a specific feature in their first session versus those who didn't. These segments often have wildly different retention curves, which tells you which early-experience moments matter for long-term retention.
By Plan or Pricing Tier
In subscription products, paid users typically retain better than free, and higher tiers better than lower. Segmenting reveals whether your free-to-paid funnel is working and whether different price tiers have different user dynamics.
By Geography or Device
Some markets retain very differently from others. Mobile users may retain differently from desktop users. These segments often reveal where the product is working well and where it isn't, which directs investment geographically or platform-wise.
By Behaviour
Users who did X in their first week versus those who didn't. Power users versus light users. These behavioural cohorts often produce the most actionable insight: users who completed action Y retain at 70% versus 30% for those who didn't; getting more users to action Y should be a priority.
Common Patterns You Will See
The Smile Curve (Rare and Good)
Some products have curves that decline initially, then rise as engaged users invite others or expand usage. This is rare but the strongest possible signal. It usually happens in network-effect products where each user becomes more valuable over time.
The Cliff (Common and Worrying)
Steep drop-off in the first week or month, then continued decline. Users tried the product, didn't see value, and left. The fix is in activation: helping more users reach the moment of value in their first session.
The Slow Bleed (Common and Insidious)
Gradual decline over many months. No dramatic drop, just steady loss. Often missed because each month looks like the last. Cohort analysis surfaces the bleed when aggregate metrics hide it.
The Plateau (Healthy)
Initial drop, then flat retention thereafter. The loyal users have formed; they aren't going anywhere. The flat level represents the product's sustainable user base. This is the most common shape for healthy products.
The Reactivation Spike
Sometimes a re-engagement campaign or new feature produces a temporary spike in retention for a specific cohort. The spike is informative; if it lasts, the intervention worked. If it reverses, the campaign produced temporary movement without changing fundamental behaviour.
Common Mistakes in Cohort Analysis
Mistake One: Looking Only at Aggregate Retention
Aggregate retention can mask radical changes happening at the cohort level. By the time aggregate moves, the cohort pattern has been clear for months. PMs who don't look at cohorts miss early warning signs.
Mistake Two: Comparing Cohorts of Different Lengths
An older cohort has more data than a recent one. Comparing their average lifetime or total revenue is misleading because the older cohort had more time to accumulate. Compare cohorts at the same age (month 3 of each, week 6 of each), not at the same calendar date.
Mistake Three: Reading Noise as Signal
Small cohorts produce noisy curves. A cohort of fifty users may have wild swings that don't mean anything. Be careful about drawing conclusions from cohorts below several hundred users.
Mistake Four: Looking at Too Many Cohorts
Some teams produce cohort tables with fifty rows and twenty columns. The eye can't process this. Five to ten cohorts at a time, spaced meaningfully (monthly or quarterly), is usually enough to see patterns.
Mistake Five: Confusing Cohorts With A/B Tests
Cohorts are observational, not experimental. Differences between cohorts could be caused by anything that changed over time, not just product changes. Don't draw strong causal conclusions from cohort differences; use them as hypotheses to test more rigorously.
Mistake Six: Defining "Active" Loosely
If active means logged in , you'll see inflated retention. Users who log in once and bounce count as retained. Tighten the definition: active means they performed a meaningful action. The tighter definition produces lower retention numbers that more accurately reflect product value.
Using Cohorts to Drive Decisions
Cohort analysis is most valuable when it changes decisions. A few patterns that drive meaningful action.
Diagnosing a Slowdown
Aggregate growth has stalled. The team is confused; acquisition looks fine. Cohort analysis reveals that recent cohorts retain dramatically worse than older ones. The slowdown isn't in acquisition; it's in retention. The team can now investigate what changed about new users' experience.
Finding the Activation Threshold
Segment cohorts by whether users completed certain actions in their first week. The cohort that completed action X retains at 65%; the cohort that didn't retains at 20%. Action X looks like a critical activation milestone. The team focuses on getting more new users to action X.
Identifying Bad Acquisition Channels
By-channel cohort analysis reveals that one channel produces users with terrible retention. Even though the channel looks cheap on a cost-per-acquired-user basis, the lifetime value is so low that the channel is net-negative. Pause the channel and reallocate.
Measuring Real Impact of Major Changes
After a major redesign, the team can compare cohorts from before and after the change. If post-redesign cohorts retain meaningfully better than pre-redesign ones, the change worked. This is a slower but more reliable signal than short-term A/B tests.
Validating Pricing Changes
After a pricing change, cohorts of new signups can be compared. Are users on the new pricing retaining differently? Are they expanding? Pricing changes have long latency, and cohort analysis is one of the few ways to measure their real impact.
How Often to Look
Cohort analysis is not a daily metric. The patterns develop over weeks and months. A reasonable cadence:
- Monthly. Review the main retention cohort table. Note any new patterns or surprises. Ten to twenty minutes.
- Quarterly. Deeper review. Segment by channel, by plan, by behaviour. Compare quarter-over-quarter trends. Look for unexpected patterns.
- When something changes. After major launches, pricing changes, redesigns, or acquisition shifts, cohort analysis at three and six months tells you whether the change is producing real effects.
- When metrics surprise you. Whenever aggregate metrics move unexpectedly, look at cohorts to understand what's actually happening underneath.
A Note on Tooling
Cohort analysis is well-supported by modern product analytics tools. Amplitude, Mixpanel, Heap, Pendo, PostHog, and others all produce cohort tables and retention curves with little setup. For simple cases, you can also produce them in spreadsheets from raw event data, though this becomes tedious at scale.
The tool matters less than the practice. PMs who look at cohorts monthly, regardless of tool, build a better understanding of their product than PMs who have sophisticated tools but rarely look. Build the habit first; the tooling is secondary.
A Final Word
Cohort analysis is one of the most valuable practices in product analytics. It reveals what aggregate metrics hide. It surfaces problems before they become severe. It tells stories that summaries can't. PMs who use it well develop a richer understanding of their products and make better decisions as a result.
If you take one practice from this article, take this: look at your product's retention cohort table this week. If you have never seen one, generate it now. Look at the last six cohorts. Read across each row. Compare down each column. What do you see? What surprised you? What does the shape of the curve look like? The first time you do this, you will see things you didn't know. That is the start of a practice that will inform every product decision you make for the rest of your career.
Key Takeaways
- A cohort is a group of users sharing a characteristic (usually signup date). Cohort analysis tracks each group over time, revealing what averages hide.
- Healthy retention curves flatten; unhealthy ones decline toward zero. Improving cohorts (newer ones retain better) are the strongest positive signal.
- Retention compounds. Small improvements have outsized long-term effects. Most teams under-invest in retention relative to its leverage.
- Segment cohorts by channel, behaviour, plan, geography, or device. Segments often reveal where retention varies dramatically and where to focus.
- Cohort analysis is most valuable when it changes decisions: diagnosing slowdowns, finding activation thresholds, identifying bad channels, measuring change impact.