A retention curve plots the share of one cohort still active against the number of days or weeks since that cohort arrived, and its shape is the most diagnostic picture in product analytics.
Reading a cohort table across a row already produces the numbers. The curve is that row drawn as a line, and drawing it makes one question answerable that the table only hints at, which is whether the falling ever stops.
Two shapes carry almost all of the meaning. A curve that falls steeply and then levels off has found a group of people who keep coming back, and the height of that level is the share of every future intake the product can expect to hold. A curve that keeps falling until it touches zero has no such group, so every customer the product has at any moment is a customer it recently acquired.
This matters to a product manager because the two shapes imply completely different work. A product with a floor grows by pouring more people into the top, since a predictable share of them stick. A product with no floor cannot grow that way at any price, because acquisition spending buys a population that has already left by the time the next invoice arrives. So the shape of the curve decides whether growth spending is worth authorising at all.
The sections below read the two shapes off a chart. They then set out the three definitions of retention in common use and what the choice between them does to the number. Finally they cover the rare third shape and what a retention curve cannot say.
Two shapes and what the floor means
Both cohorts lose more than half of their people in the first week, so the early fall says very little. The two products separate after day thirty, when one curve stops moving and the other carries on down to nothing.
The first fortnight of any retention curve looks alarming and carries almost no information. People try things, find them unsuitable and leave, and a steep early drop is what every product shows whether it is durable or not. A team that reads the first week as a verdict will conclude that every product is failing.
What separates the two lines above is what happens after the steep part. The darker curve settles at 22 per cent and stays there through days 60 and 90, which means roughly a fifth of each intake has made the product part of how they work. The lighter curve never settles. It passes through 6 per cent at day 30 and 2 per cent at day 60, and by day 90 the cohort has effectively gone.
A floor is what turns a growth plan into arithmetic. If a product holds 22 per cent of each intake and acquires ten thousand people a month, it adds about two thousand two hundred durable customers a month and its base compounds. The same ten thousand a month against a curve that reaches zero produces a base that stops growing the moment acquisition stops.
Sean Ellis is usually credited with the argument that a flattening curve is the honest evidence of product market fit, and the reasoning behind it is worth stating without the label. A curve that flattens shows that a stable share of arriving people find the product worth returning to unprompted. That is a stronger claim than any survey produces, because months of behaviour stand behind it.
Day N retention, unbounded retention and bracket retention
Behaviour over months is only as solid as the rule deciding who counts as active. Three definitions of retention are in common use, and the vocabulary for them came from Amplitude's analytics guidance. They differ only in which days count as a return, and the gap between them is large enough to change a decision.
Day N retention counts the people who were active on day N exactly. If somebody came back on day 29 and on day 31 and missed day 30, this definition does not count them. It suits a product somebody is meant to open every day, such as a habit tracker or a news app.
Unbounded retention counts the people who were active on day N or on any day after it. If a person's next visit falls on day 44, they still count as retained on day 30, because the definition asks only whether they ever came back. It suits a product with an irregular rhythm, such as a booking tool or a tax return.
Bracket retention counts the people who were active at least once inside a window the team defines, such as days 24 to 30. It sits between the other two and it matches how most business software is actually used, since a weekly rhythm makes any single day arbitrary.
What the choice of definition does to the number
One cohort of a thousand people shows how far apart the three definitions can land. Of those thousand, 180 opened the product on day 30 itself, 265 opened it at least once between day 24 and day 30, and 310 opened it on day 30 or on some later day inside the measurement period.
| Definition | People counted | Reported retention |
|---|---|---|
| Day 30 retention | 180 | 18 per cent |
| Bracket retention, days 24 to 30 | 265 | 26.5 per cent |
| Unbounded retention from day 30 | 310 | 31 per cent |
Every figure in that table is correct and they describe the same thousand people using the same events. The spread between the lowest and the highest is a factor of about 1.7, which is larger than almost any product change a team will ship in a year.
Two consequences follow. A benchmark quoted without its definition is worthless, so a board slide comparing a product's 18 per cent against a competitor's 31 per cent may be comparing a strict measure against a generous one. And a team that changes its own definition halfway through a year has manufactured an improvement, which is why the definition belongs in the tracking plan alongside the events it is built from.
The smile curve and the products that produce one
A third shape appears occasionally. Retention falls, flattens and then begins to climb, because people who had drifted away come back as the product becomes more useful. Santiago Rodriguez and Alex Immerman of Andreessen Horowitz described this pattern on 10 September 2025 in an analysis of retention among companies building on large language models, noting that most curves flatten around the third month and that a small number turn upwards afterwards.
A rising tail is worth understanding before anybody claims one. It requires former users to notice a change, which means the product has a way of reaching people who stopped using it. It also requires the change to be large enough to overturn a judgement somebody already made. Most products have neither, so a flat line remains the realistic good outcome.
What a retention curve cannot show
A retention curve is silent about why anybody stayed. Twenty two per cent of the cohort persists through day 90, and the chart reports that share as a count while saying nothing about the people inside it. What those people have in common, and what the product did for them that the other 78 per cent never got, is absent from the picture entirely.
The curve is also an average over a mixed population. A single line at 22 per cent can hide one group retaining at 60 per cent and another at 4 per cent, and those two groups usually want different things from the roadmap. Splitting the curve by plan, by company size or by how somebody arrived is the standard repair, and the split often matters more than the headline.
There is one more limit that catches teams out. A retention curve describes the cohorts it was built from, which are cohorts that already arrived. A product changing its acquisition channels is changing the population the next curve will describe, so last quarter's floor is a forecast only while the intake stays the same.
The flattening curve leaves one question behind and it is the useful one. A fifth of each intake keeps coming back, and the next thing worth knowing is what those people did in their first few days that the rest never did.