Cohorts, not totals: how a flat dashboard hides a dying product
A steady active-user count can mask catastrophic churn — and the only way to see the rot is to stop looking at totals and start looking at cohorts.
Somewhere inside a Series A deck there is a slide showing monthly active users as a smooth, slowly rising line. The founders are proud of it. The number has not dropped once in two years. An investor flips past it in four seconds and asks the one question the founders were hoping to avoid: “Can you show me the retention curves by cohort?”
The founders go quiet.
The line was not a lie. It was just a vantage point so high that all the damage was invisible.
The bucket with a hole
Imagine a bucket. Every month, 5,000 new users sign up and pour in from the top. Every month, 5,000 users lose interest and drain away from a hole in the bottom. The water level — your monthly active user count — stays perfectly flat. The dashboard looks fine. The product is dying.
This is not a hypothetical. It is the default failure mode of early-stage SaaS, consumer apps, and subscription products. Growth teams optimise for acquisition because acquisition is fast and measurable. Retention is slow, subtle, and shows up in data that most teams are not even collecting correctly.
The tragedy is not that teams are dishonest about their numbers. It is that total active users is genuinely the wrong unit of analysis. It conflates two completely separate phenomena: how many new people you attracted this month and how many people you kept from last month. Mixing them into one number is like measuring the temperature of a room by averaging the thermometer reading from the freezer and the oven — the result is technically accurate and completely meaningless.
What a cohort actually is
A cohort (from the Latin for a division of soldiers) in product analytics means a group of users who joined during the same time window — usually the same calendar month. Cohort analysis tracks what happens to that fixed group over time. The population of the cohort never grows; it can only shrink.
January cohort: 1,000 users signed up.
Month 1 (February): 600 of them opened the product at least once. That is a 60 percent month-1 retention rate.
Month 2: 450 of the original 1,000 are still active. 45 percent retention.
Month 3: 380. 38 percent.
Month 4: 350. 35 percent.
Month 5: 340. 34 percent.
Write it as a table and the pattern leaps out:
| Month after signup | Active users | Retention |
|---|---|---|
| 1 | 600 | 60% |
| 2 | 450 | 45% |
| 3 | 380 | 38% |
| 4 | 350 | 35% |
| 5 | 340 | 34% |
The drop from month 1 to month 2 is steep: 150 people gone. Month 2 to month 3 loses another 70. But month 3 to month 4 loses only 30, and month 4 to month 5 loses 10. The rate of loss is decelerating. The curve is flattening.
That flattening is everything.
The shape of the curve is the signal
A retention curve that decays to zero is a product with no loyal users. Every single person who tries it eventually leaves. You are running on a treadmill. Growth only continues for as long as you can acquire faster than you lose, and acquisition does not scale forever.
A retention curve that decays and then flattens — even at 30 or 34 percent — means there is a core of users who have decided this product is genuinely part of their life. They churned out at month 1 or month 2 because the product was not right for them, but the ones who stayed are sticky.
This is the earliest quantitative signal of product-market fit (the degree to which a product satisfies a strong market demand). Not the NPS survey. Not the Twitter praise. The flattening tail of the monthly retention curve.
Andrew Chen at Andreessen Horowitz has written about this, Brian Balfour at Reforge has built a curriculum around it, and every practitioner who has worked at a high-retention consumer product eventually arrives at the same conclusion: retention curves are not a reporting exercise. They are a diagnostic tool.
The question they answer is not “how many users do we have?” but “do users who try this product find enough value to keep coming back?” Those are completely different questions with completely different implications.
Why totals feel true and lie constantly
There is a psychological reason product teams default to total active users. The number goes up. Going up feels like progress. There is a visceral pleasure in watching a metric increase, and there is a social cost to questioning it — nobody wants to be the person who ruins the good mood at the weekly all-hands.
But the fundamental issue is statistical: an aggregated total obscures the distribution of its parts. If you have five January cohorts each retaining at 34 percent at month 5, and you add 5,000 new signups every month like clockwork, your total active users in May will look healthy by almost any snapshot metric. What the snapshot hides is that only a third of any given month’s signups are still around six months later.
For subscription businesses (where revenue is tied to continued usage), this matters enormously. Customer lifetime value (LTV) — the total revenue expected from a single customer over their relationship with you — is directly a function of retention. If month-1 retention is 60 percent and the average subscription costs 20 dollars per month, the expected revenue from a new customer is not 20 dollars times however many months you hope they stay. It is 20 dollars times the sum of your retention probabilities at each month. With our cohort numbers, that works out to roughly 20 dollars times (1.0 + 0.60 + 0.45 + 0.38 + 0.35 + 0.34 …) — every improvement to that curve compounds directly into LTV.
This is why a venture investor’s first question after seeing a revenue chart is not “what is your growth rate?” but “what does retention look like by cohort?” Growth rate tells you how fast the bucket is filling. Retention tells you how large the hole is.
The leaky bucket in practice
Consider two companies at the same total monthly active user count — say 50,000 each.
Company A acquired 50,000 users over four years and most of them are still active. Its oldest cohort retains at 60 percent after 36 months. Customer acquisition cost (CAC — the fully-loaded marketing and sales spend per new customer) is high but infrequent, because the company does not need to constantly replace its base.
Company B grew to 50,000 users by acquiring 10,000 new users every month and losing 10,000 old ones. Its 36-month retention is 2 percent. The 50,000 headline number is accurate. But every single quarter Company B must spend CAC times 30,000 just to stand still, because a third of its base churns within the first 90 days. The burn rate required to maintain a flat user count is staggering, and it grows with every increment in size.
On a single dashboard showing total active users, both companies look identical.
In a board meeting showing cohort retention grids, they look like completely different species.
Reading a retention grid
The retention grid is the standard format for communicating cohort data. Rows are cohorts (January 2025, February 2025, …), columns are months of age (M1, M2, M3, …), and each cell contains the retention percentage for that cohort at that age.
A healthy grid looks like a heat map that goes dark orange on the left (high early retention) and stays reasonably warm all the way to the right (flat tail). An unhealthy grid fades to near-white within two or three columns — everyone is leaving early, and there is no stable floor.
You can read two things from a retention grid that you cannot read from any aggregate:
Cohort-over-cohort improvement. If the February cohort’s M1 retention is 65 percent and the January cohort’s was 60 percent, a product change that shipped in late January probably helped. This is how product teams close the feedback loop between a feature ship and user behaviour — not through surveys, but through watching whether the next cohort’s curve changed shape.
The level of the floor. Where does the curve stabilise? 40 percent is exceptional in most consumer categories. 20 percent is viable for many SaaS products if the economics work. 5 percent is a signal that the product has not found its audience, or that the free-to-paid conversion is broken, or that the onboarding experience is bleeding users before they see any value.
The onboarding problem hiding in month 0 to month 1
The steepest drop in almost every retention curve happens between signup and month 1. In the January cohort above, 40 percent of users — 400 people — never came back after their first session. This is not churn in the traditional sense. It is abandonment. The product never had a chance to demonstrate its value because the user left before the value moment arrived.
Product teams call this “time to value” — the elapsed time between a user signing up and the moment they first experience the core benefit the product promises. For a note-taking app, time to value is how long before you create your first useful note. For a data pipeline tool, it might be how long before your first successful pipeline run. The longer this takes, the steeper the M0-to-M1 cliff.
The implication is practical: if your M1 retention is 40 percent and your M5 floor is 30 percent, your onboarding problem is actually larger than your long-term retention problem. Fixing the first-session experience — compressing time to value, reducing friction, removing steps that require the user to do work before they receive benefit — would shift the entire curve upward without touching the product’s core loyalty mechanics.
What the flat tail actually proves
Back to the January cohort. By month 5, 340 of the original 1,000 users are still active. The curve went from 60 percent at month 1 to 34 percent at month 5, but it lost only 10 users between month 4 and month 5. If the next measurement comes in at 33 or 34 percent, you have found your floor.
That floor is the product’s natural user base — the people for whom this tool is genuinely the right answer. They are not using it out of inertia. They are not trapped by a yearly contract. They signed up in January, survived the chaos of month 1 and month 2, and are still showing up in June.
These are the users you should be talking to. Their behaviour tells you what the product is actually for, which is often subtly or dramatically different from what the team thought they were building. They have implicitly answered the question “is this worth my ongoing time?” in the affirmative, and understanding why — what job the product does for them, what alternative they would use if it disappeared — is the most valuable research a product team can conduct.
The flat tail is also the projection you can compound. If 34 percent of each January class will still be active 18 months from now, and you know your average revenue per user, you can model forward revenue with reasonable confidence. Businesses with flat retention floors are fundamentally more predictable, and predictability trades at a premium in every financing environment.
The metrics that actually matter downstream
Once you are looking at cohorts, several other numbers start making sense in ways they did not before.
Net revenue retention (NRR) is what happens when you track not just whether users are active but how much revenue they generate over time within a cohort. If your average customer expands their usage — upgrades their plan, adds seats, buys adjacent features — the revenue from a cohort can actually grow even as some users churn. NRR above 100 percent means expansion revenue exceeds churn revenue, and the business literally grows without acquiring a single new customer.
Payback period is how many months of gross margin from a customer it takes to recover CAC. With good cohort data, you can calculate payback per cohort class and see whether your economics are improving over time. A January cohort where payback is 14 months and a May cohort where payback is 9 months is a direct signal that either acquisition costs fell or early retention improved.
Engagement depth before churn — looking at what actions churned users took (or did not take) in their first session — turns the retention curve into a diagnostic for the onboarding flow. This is where qualitative insight and quantitative cohort analysis converge.
None of these numbers are computable without cohort thinking. They all require tracking the same group of people across time, rather than snapshotting the total population at a moment.
The organisational will to look
The hardest part of cohort analysis is not technical. Pulling a cohort retention grid from most modern analytics stacks — Amplitude, Mixpanel, even a well-structured data warehouse query — takes an afternoon. The hard part is creating the organisational will to look at what the grid reveals and act on it.
A flat or declining total active user count is bad news that is obvious. A rising total active user count with a collapsing retention curve is bad news that requires interpretation, and interpretation creates room for denial. Teams can argue about whether the cohort definition is right, whether the retention event is correctly specified, whether the newest cohort just needs more time. The conversation can absorb weeks before anyone accepts what the data is saying.
This is why the discipline of cohort analysis is as much cultural as it is analytical. It requires agreeing, before the data is uncomfortable, on what the metrics mean and what thresholds would constitute a problem. It requires treating a flattening retention curve as a strategic asset worth protecting and a decaying one as a crisis worth interrupting roadmaps over.
The insight that a steady headline number can hide catastrophic health is not new. Anyone who has read about the leaky bucket has encountered it. The more important insight is that the bucket is always leaking to some degree, and the only honest question is how fast.
The retention curve answers that question. Everything else is theatre.