Cohort Analysis 101: How to Track Retention That Actually Drives Decisions

Most product teams track retention by looking at one number: monthly active users. That number rises, the team celebrates; it falls, the team panics. But MAU is a stock metric -- it tells you how many people are around right now without telling you anything about why. A product can grow MAU while losing every cohort of users it acquires, simply by acquiring faster than it churns. By the time the curve flattens, the underlying retention problem has been masked for quarters. Cohort analysis is the antidote: instead of asking "how many users did we have in March," it asks "of the users we acquired in January, how many were still around in March, and how does that compare to December's cohort?"

This guide explains what a cohort actually is, the critical distinction between acquisition and behavioral cohorts, how to read a retention curve without fooling yourself, the difference between N-day and unbounded retention, the five cohort views every product team should run weekly, and how Gurulu's audiences and funnels produce these views without writing SQL.

What a Cohort Really Is

A cohort is a group of users defined by a shared event in a shared time window. The most common definition is the acquisition cohort: every user who first signed up in the same week. But cohorts are more general than that -- they can be defined by any qualifying event. "Users who completed onboarding in week 12" is a cohort. "Users who upgraded to paid in March" is a cohort. "Users who first hit the new dashboard after the April release" is a cohort. The shared event and the shared window are what make a cohort, not the user property.

Why this matters: when you measure retention against a non-cohort base -- like "all paying users right now" -- you mix old loyal users with new fragile ones, and the average tells you nothing actionable. Cohorting forces you to compare apples to apples: this week's signups against last week's signups at the same lifecycle stage. That is the only honest way to know whether your product is getting better or worse at keeping users.

Acquisition vs. Behavioral Cohorts

Acquisition cohorts. Group users by when they first arrived. This is the default cohort and the right one for measuring overall product-market fit. The classic view is a triangle table: rows are weekly signup cohorts, columns are weeks since signup, cells are the percentage still active. If your product-market fit is improving, the curves of newer cohorts should sit above older ones at every age.

Behavioral cohorts. Group users by an action they took, not when they joined. Examples: users who invited a teammate within their first 7 days; users who imported data; users who hit the API once. Behavioral cohorts are how you find activation signals -- the early actions that correlate with long-term retention. If "invited a teammate in week 1" cohorts retain at 80% after 90 days while "never invited" cohorts retain at 12%, you have just found your North Star activation event.

The two cohort types answer different questions. Acquisition cohorts answer "is the product getting better?" Behavioral cohorts answer "what behavior should we drive to make users stick?" Mature product teams run both continuously. Skipping behavioral cohorts is the most common analytics mistake -- it is also the easiest one to fix.

Reading a Retention Curve Without Lying to Yourself

A healthy retention curve has three regions. The first 1-7 days show steep decline as casual signups bounce -- this is normal and unavoidable; trying to flatten this region usually means restricting signups. Days 7-30 are the activation valley, where users either find value or churn for good. After day 30, a healthy product shows a flat tail: the surviving cohort continues using the product indefinitely, and the curve approaches a horizontal asymptote rather than a zero. The asymptote height is your true retention rate -- the percentage of acquired users who become long-term value.

An unhealthy product has no flat tail. The curve continues sloping toward zero forever, meaning every cohort eventually disappears. This is called terminal churn, and it almost always indicates that the product is not solving a recurring problem. If your day-180 retention is 8%, day-360 is 4%, day-720 is 2%, you do not have a retention problem you can fix with onboarding tweaks; you have a value problem that needs a product strategy answer.

The retention curve also reveals seasonality and product changes. A V-shaped dip at week 8 across multiple cohorts often correlates with a release that broke a key flow. A sustained lift in newer cohorts after a feature launch is the cleanest evidence that the feature actually moved retention -- much more credible than week-over-week MAU.

N-Day vs. Unbounded Retention

N-day retention. A user is retained on day N if they performed a qualifying action exactly on day N. Used by Facebook and most consumer apps. It punishes apps with weekly or monthly usage cadences (like banking or HR tools), so it is a poor fit for B2B SaaS. The advantage is that it is unambiguous and easy to graph.

Bracket (rolling) retention. A user is retained for week N if they were active on any day in that week. Much more forgiving for natural-cadence products. This is the right default for B2B and most subscription products. Gurulu's audience builder defaults to this when you ask for "users active in the last X days."

Unbounded retention. A user is retained for week N if they were active in week N or any later week. Tells you what fraction of the cohort eventually returned, regardless of consistency. Useful for low-frequency products (tax software, travel) where weekly activity is not the goal.

Range retention. A user is retained for the range [N, N+k] if they had at least one active day in that range. The compromise between bracket and unbounded; common in 28-day MAU calculations.

Pick one and stick to it. The retention number you cite changes by 30-60% depending on which definition you use. Mixing definitions across teams or quarters is how you end up arguing about whether the product is healthier than last year when nobody is computing the same number. Document your definition in your dashboard and never silently swap it.

The 5 Cohort Views Every Product Team Should Run Weekly

First, weekly acquisition cohort triangle for your North Star action -- the activation event that defines a real user. Second, behavioral cohort comparison: users who completed activation within their first session vs. those who did not, retention at days 30 / 60 / 90. Third, paying-customer cohort retention -- this is the only number that matters for revenue, and it should be computed off subscription state, not engagement. Fourth, feature-adoption cohort: users who first used a specific feature this month, tracked against a baseline cohort that did not. Fifth, channel cohort: users acquired from each major source (organic, paid, referral, partnership) measured at the same age, to know which channels deliver retainable users vs. tourists.

In Gurulu, each of these is an audience definition plus a retention chart. Audiences are persistent: define "completed activation in week 1" once and it auto-updates every week. Funnels handle the multi-step versions: "signed up -> activated -> purchased -> renewed," with cohort-aware drop-off so you see how each weekly intake performed at each step. The system also surfaces vanity-vs-decision-grade warnings: a cohort view based on a single click that 95% of visitors take is flagged as low-signal because it cannot discriminate between cohorts. Decision-grade cohorts have asymmetric outcomes -- they separate users who will be valuable from users who will not -- and that is the bar to hold every cohort report to.