Cohort analysis: the retention view that catches what aggregate metrics hide.
Cohort analysis is the practice of grouping customers by a shared starting characteristic (usually the month or week they joined) and tracking behavior across the cohort over time. The most common output is a retention curve. The curve's shape tells you whether the business compounds. Aggregate retention can look healthy while channel-level cohorts hide a money-losing acquisition source underneath. The slice reveals what the average conceals. Cohort analysis has been standard SaaS practice since the early 2010s and is now table stakes for any growth team running on data.
Key takeaways
- Cohort analysis groups customers by a shared starting characteristic and tracks behavior over time. The retention curve shape tells you whether the business compounds.
- Three common curve shapes: falling to zero (one-and-done), falling to a flat line (subscription), and the smile curve (network effects).
- Aggregate retention hides channel-level differences. The single highest-leverage move is slicing cohort curves by acquisition channel.
- Newer cohorts retaining better than older cohorts is the healthiest pattern. Newer cohorts retaining worse is the warning sign.
- Monthly cohorts are the standard for most businesses. Weekly works for high-volume; quarterly works for long-cycle B2B.
- A channel with high CAC and high retention can be more valuable than a channel with low CAC and low retention. The blended view does not show this; the cohort view does.
What cohort analysis actually is
Cohort analysis is the practice of grouping customers by a shared starting characteristic (usually the month or week they joined) and tracking behavior across the cohort over time. The most common output is a retention curve: what percentage of each cohort is still active at 30, 60, 90, 180, 365 days after joining. The shape of the curve tells you almost everything you need to know about whether your business compounds.
The work matters because aggregate numbers hide everything important. A business reporting "70 percent retention" might have flat 70 percent retention across all cohorts (healthy), or might have 90 percent retention for the oldest cohorts and 40 percent for the newest (broken; new acquisition is bringing in worse customers). The cohort view separates these two stories. The aggregate view conceals them.
Cohort analysis became standard practice in SaaS analytics through the 2010s. Tomasz Tunguz, Jason Lemkin at SaaStr, and David Skok at Matrix Partners all wrote influential essays on cohort retention curves. Mixpanel and Amplitude built cohort tools into their core products. By 2020, no growth team should ship a retention analysis that does not include a cohort view.
The three retention curve shapes
Three shapes capture almost every real business. Each shape implies a different growth model, a different unit-economics structure, and a different operating discipline. Knowing which shape your business has determines whether scaling acquisition makes sense or whether the team should fix retention first.
| Shape | What it looks like | Implication |
|---|---|---|
| Falling to zero | Curve declines steadily over time, eventually reaching near zero | One-and-done categories (furniture, big appliances). Growth requires perpetual acquisition replacement. |
| Falling to a flat line | Curve drops then stabilizes at a meaningful percentage | Subscription products with stable retention. The flat-line height determines long-term revenue. |
| Smile curve | Retention drops then climbs back up | Network-effect products. Surviving customers become more engaged. Rare and extremely valuable. |
Aggregate retention can hide all three shapes. A business with a falling-to-zero curve and one growing-cohort smile shape will look flat in the aggregate but be either dying or thriving depending on which cohort weights more in the period. Looking only at cohort curves catches what aggregate hides.
How to run a cohort analysis
Six steps for running a cohort analysis that produces decisions, not just charts. The work takes one day for a clean SaaS warehouse, longer if customer event data is scattered across analytics tools. Mixpanel, Amplitude, and any modern data warehouse all support the analysis natively.
- Define the cohort start event.The most common is signup date. Other options: first purchase, first paid month, account activation. Pick the event that matters for the question you are asking.
- Pick the retention metric.D7, D30, D60, D180, D365 are typical milestones. The retention metric is the behavior you care about (returns, purchases, sessions). Logged-in users is a vanity metric; behavior matters.
- Group customers into cohorts.Monthly cohorts are the standard. Weekly works for high-volume businesses. Quarterly works for low-volume B2B with long sales cycles.
- Compute retention rates per cohort.Percent of each cohort still active at each time milestone. Plot as a heatmap (cohorts as rows, time periods as columns) or as overlapping line charts.
- Look for trend lines.Newer cohorts retaining better than older cohorts is the healthiest pattern. Newer cohorts retaining worse is the warning sign that requires immediate investigation.
- Slice by acquisition channel and persona.The same business often has very different cohort curves by source. A channel with cheap CAC and bad retention is destroying value; the aggregate hides this.
Slicing cohorts by acquisition channel
The single highest-leverage move in cohort analysis is slicing by acquisition channel. Aggregate retention can look healthy while one channel is silently bringing in customers who churn fast. Channel-level cohort curves reveal which channels are sustainable and which are subsidized by the better channels.
The pattern shows up everywhere in audit. A DTC brand with 65 percent month-3 retention has Meta-acquired customers at 75 percent and Google PMax-acquired customers at 35 percent. The Meta channel is healthy; the PMax channel is destroying value. The aggregate looks fine. The slice reveals the problem.
The fix is to bring CAC and retention together at the channel level. A channel with $50 CAC and 75 percent retention may be more valuable than a channel with $20 CAC and 30 percent retention, despite the lower CAC. The blended view does not show this. The channel-level cohort view does.
Quick answers
- What is cohort analysis in plain English?
- Group customers by the month they joined. Track how each group behaves over time. The shape of each group's retention curve tells you whether the business is getting healthier or worse over time.
- What are the three retention curve shapes?
- Falling to zero (one-and-done categories), falling to a flat line (subscription with stable retention), and the smile curve (network-effect products where surviving users get more engaged).
- How is cohort analysis different from regular retention?
- Aggregate retention is one number for the whole customer base. Cohort retention shows the number broken down by when customers joined. The cohort view reveals trends the aggregate hides.
- How often should I run cohort analysis?
- Monthly at minimum. The best practice is to look at cohort retention every time the team reviews acquisition or product changes, to catch whether the recent work moved the curve.
- Which tool runs cohort analysis?
- Mixpanel, Amplitude, and Heap all have cohort tools built in. A data warehouse with SQL works too. The query is a window function over user events grouped by signup date.
- How should I slice cohorts?
- By acquisition channel first. Different channels often have very different retention curves. Then by persona or use case. The slices reveal which segments compound and which destroy value.
Frequently asked
What is cohort analysis?
Cohort analysis is the practice of grouping customers by a shared starting characteristic (usually the month or week they joined) and tracking behavior across the cohort over time. The most common output is a retention curve.
Why does cohort analysis matter more than aggregate retention?
Aggregate retention can hide everything important. A business reporting 70 percent retention might have flat retention across all cohorts (healthy) or might have 90 percent for old cohorts and 40 percent for new ones (broken). Aggregate cannot tell these apart.
What are the three retention curve shapes?
Falling to zero: one-and-done categories like furniture. Falling to a flat line: subscription with stable retention. The smile curve: network-effect products where retention dips then climbs back as surviving users get more engaged.
What is a healthy retention curve?
Either falling to a flat line at a meaningful percentage (typical for subscriptions) or showing the smile pattern (typical for network-effect products). Curves that fall to zero indicate one-and-done categories that require perpetual acquisition.
How do I slice cohorts effectively?
First by acquisition channel. Different channels often have very different retention curves. Then by persona, use case, or pricing tier. The slices reveal which segments compound.
How granular should the cohorts be?
Monthly cohorts are standard. Weekly works for high-volume businesses (consumer apps, high-velocity DTC). Quarterly works for low-volume B2B with long sales cycles.
What is the smile curve?
A retention curve that drops in the first weeks or months, then climbs back up as the surviving cohort becomes more engaged. Common in network-effect products like Facebook, LinkedIn, and marketplaces. Rare and extremely valuable when it appears.
Which analytics tools support cohort analysis?
Mixpanel, Amplitude, Heap, and most product analytics tools have native cohort views. In a data warehouse, the query is a window function over user events grouped by signup cohort and time-since-signup.
Sources cited on this page
- Tomasz Tunguz — Theory Ventures blog on SaaS cohort analysis.
- David Skok — For Entrepreneurs blog on retention curves.
- Amplitude — Product analytics blog on cohort retention.
- Mixpanel — Cohort analysis documentation and case studies.
- Lenny Rachitsky — Growth-leader interviews on retention.
- Reforge / Casey Winters — Essays on retention and cohort dynamics.