"Retention is 82%" is a sentence that can mean almost anything. Eighty-two per cent of whom, measured when, after how long? Average it across everyone who ever signed up and you get a comforting number that tells you nothing about whether your product is getting better or quietly bleeding out. Cohort and retention-curve analysis is the discipline of refusing that comfort, of cutting the blur into honest slices and watching what each one actually does.

The quick version

  • A cohort is a group of users who started at the same time (signed up in March, say). You track each group separately instead of blending everyone into one average that hides the trend.
  • Plot the share of a cohort still active over weeks or months and you get a retention curve. It almost always drops fast, then bends.
  • The number that matters is where the curve flattens, the plateau, not the day-one figure. A curve that flattens above zero means a real population finds the product valuable. A curve that slides to zero means it doesn't, yet.
  • Stack newer cohorts against older ones and you can see whether your product is genuinely improving, even when the blended average looks flat.

The idea in depth

Cohort analysis isn't a leadership theory; it's a measurement habit that good leadership decisions depend on. Its modern popularity in product and operating teams comes largely from Eric Ries's The Lean Startup (2011), which argued that most dashboards report vanity metrics, cumulative totals like "registered users" that only ever go up and never tell you whether the product is working. Ries's alternative was actionable metrics, and his recommended tool for them was cohort analysis: stop looking at the whole pile, line up each week's or month's joiners, and compare like with like. The move that follows is simple and unglamorous, before you celebrate a topline, ask which cohort produced it and whether the newest cohorts are doing better or worse than the old ones.

Why the curve bends, and why that's the whole point

Plot any cohort's retention over time and you nearly always see the same shape: a steep early drop, then a flattening. The instinct is to read the early drop as failure. The more useful reading comes from marketing scientists Peter Fader and Bruce Hardie, whose paper "How to Project Customer Retention" (Journal of Interactive Marketing, 2007) explains why curves bend even when no individual customer changes their behaviour at all. Customers differ in how likely they are to stick. The flighty, the curious and the mistaken churn first; the people for whom the product fits stay. So the surviving cohort is, period by period, made of steadier and steadier members. The retention rate rises among survivors not because anyone became more loyal, but because the disloyal already left. Fader and Hardie call this a sorting effect driven by heterogeneity, and they show a simple "shifted-beta-geometric" model can fit and project these curves from just a few periods of data.

So the move is: don't panic at the early cliff and don't over-read it. The cliff is partly your product sorting tourists from residents. Judge the product by where and how high it settles, and by whether you can lift the plateau, not by the slope of week one.

flowchart TD
    A(["A cohort signs up together"]) --> B(["Curious, flighty & mistaken users churn fast"])
    B --> C(["A steep early drop"])
    C --> D(["Survivors are the people it fits"])
    D --> E(["The curve flattens, the plateau is your real retention"])
The early cliff is partly a sorting effect: the curve flattens because the people it doesn't fit have already left. Leaders Loop

The plateau is the closest thing to a product-market-fit reading

Once you accept that the plateau is the real signal, retention curves become a way to read product-market fit. Sequoia Capital's data-science team puts it plainly in their essay Retention: "The higher the level at which the curve flattens, the higher the long-term retention and the healthier the product." A curve that keeps declining toward zero is the signature of a product that hasn't found a population it serves; a curve that flattens above zero says it has found one, and the height of that flat line is roughly how good the fit is. The same logic runs through investor and operator writing, Brian Balfour calls the retention curve "the best proof" of product-market fit in his essay The Never Ending Road To Product Market Fit, and consumer-growth writers like Lenny Rachitsky and Casey Winters use a flattening cohort curve plus growing new-user numbers as their working definition of fit.

In practice that means picking the one action that represents your product's core value, a purchase, a saved item, a logged session, not just "opened the app", and measuring retention of that action by cohort. The flat part of that curve is your scoreboard. Lifting the plateau is a different job from improving acquisition, and most teams confuse the two.

"The higher the level at which the curve flattens, the higher the long-term retention and the healthier the product.", Sequoia Capital

This connects directly to two other Toolkit ideas. Reading a plateau honestly means knowing your distributions, percentiles and quartiles, a single average curve can hide a bimodal reality where one segment retains beautifully and another never does. And the leap from "this cohort changed" to "our new feature caused it" is exactly where correlation gets mistaken for causation: a cohort improving could be your product, or it could be that you started acquiring a better-fitting kind of user.

Where it breaks down, name the limitation

Cohort analysis is descriptive, not magic. Three honest limitations. First, recent cohorts are incomplete, a cohort from last month hasn't had time to reveal its plateau, so comparing a young curve's early points to an old curve's is a common self-deception. Second, the sorting effect cuts both ways: because survivors look better over time by construction, a flattening curve is necessary but not sufficient evidence that you changed anything. Establishing that an intervention worked needs a controlled comparison, not a prettier curve. Third, retention is a lagging measure of a moving target, Balfour's "treadmill" point is that markets shift, so a plateau you earned last year can erode without any decision on your part. The tool tells you the temperature; it doesn't tell you the cause or guarantee tomorrow.

A worked example

Imagine a mid-sized B2B scheduling tool. The headline says monthly active accounts are up and blended retention is "around 70%," and the team is pleased. Their head of product splits the data into monthly signup cohorts and plots each one's retention of the core action, an account that actually builds a schedule that month. (Figures below are illustrative, to show the shape of the reasoning, not real company data.)

flowchart LR
    M0(["Month 0: 100%"]) --> M1(["Month 1: 62%"])
    M1 --> M3(["Month 3: 41%"])
    M3 --> M6(["Month 6: 38%"])
    M6 --> M9(["Month 9: 37%"])
One cohort's retention of the core action (illustrative). The steep early drop, then a plateau near 37–38% by month six. Leaders Loop

Two things jump out. First, the "70%" was a vanity blend, it leaned on long-tenured accounts and counted any login as "active." The real, durable retention of the action that matters settles closer to 37%. That's not a disaster; it's a population for whom the tool genuinely sticks, and now they know its true size.

Second, the head of product lines up six monthly cohorts side by side and notices the plateau is creeping up: older cohorts flatten near 33%, the newest near 38%. That is the signal worth acting on. But here's the discipline, before crediting the onboarding redesign shipped in month four, she checks whether the company also changed who it was marketing to. It had: a new campaign was bringing in operations teams rather than solo admins. The curve improved; the cause was ambiguous. So the move isn't "declare victory on onboarding." It's to run a clean comparison, hold the new onboarding back from a random slice of incoming accounts for a few weeks, and let the cohorts settle the argument. The plateau told her where to look. It couldn't tell her why on its own.

Frequently asked questions

What's the difference between cohort analysis and a retention curve?

A cohort is the slice, a group defined by a shared start (or a shared trait). A retention curve is what you plot once you have a cohort: the percentage still active at each point in time. Cohort analysis is the broader habit; the retention curve is its most common chart. You can cohort by signup month, acquisition channel, plan tier, or first action.

How long do I need to wait to read a curve?

Long enough to see it flatten. For many consumer products that's a few months; for low-frequency or B2B products it can be a year or more. The practical rule from Sequoia and others: look for three or more near-flat points in a row before you trust a plateau. Judging a six-week-old cohort against a two-year-old one is the most common mistake, the young cohort hasn't finished falling yet.

What counts as a "good" retention plateau?

It depends entirely on category and frequency, so resist a universal number. Published operator benchmarks (Rachitsky and Winters' work) put roughly 25% at six months as solid and ~45% as strong for consumer social, with transactional consumer products higher, but a weekly-use product and a once-a-year product cannot be compared on the same axis. The honest answer is to benchmark against your own past cohorts and your nearest category, not a headline figure.

My curve is flattening, does that prove my feature worked?

No. Curves flatten partly by the sorting effect Fader and Hardie describe, high-churn users leave early no matter what you do, and partly because you may be acquiring better-fitting users. A flatter or higher plateau is a reason to investigate, not proof of cause. To attribute it to a change, you need a controlled comparison: hold the change back from a random group and compare their curves.

What's a "smile" curve?

A rarer pattern where retention dips, flattens, then rises as lapsed users return because the product got better. Andreessen Horowitz's Santiago Rodriguez and Alex Immerman documented it for some AI-native products (ChatGPT among them) in their 2025 piece Retention Is All You Need. It's a genuinely good sign, resurrection beating churn, but it's the exception, not what most teams should expect to see.

Related in the Toolkit

Where to go next