"Retention is 40%" is not a statement about your product. It is a statement about your product and a definition, and the definition is doing at least half the work. Three different teams can compute retention from the same event table and get 22%, 40% and 61% without any of them making a mistake.
The three numbers are not competing estimates of one quantity. They answer three different questions, and knowing which one you are looking at is most of the skill.
The three curves people confuse
N-day retention
A user counts as retained on day N if they came back on that exact day. Day 7 means day 7 — not day 6, not day 8.
This is the strictest definition and the right one for products with a genuine daily rhythm: a habit tracker, a game with a daily reset, a messaging app. It is the wrong one for anything used weekly or occasionally, where it will report catastrophic numbers for a perfectly healthy product. A B2B tool that people open every Monday will show a near-zero day-3 retention and a spike every seventh day. That sawtooth is the definition talking, not the users leaving.
Unbounded, or rolling, retention
A user counts as retained on day N if they came back on day N or any day after it. Equivalently: their last recorded activity is on or after day N.
This is the closest thing to "have they churned", and it is the most useful curve for weekly or irregular products. It is monotonically decreasing by construction, which makes it easy to read, and it flatters the numbers compared with N-day. It also cannot be computed honestly for recent cohorts: a user who signed up ten days ago cannot yet be retained at day 30, so recent cohorts must be excluded rather than plotted.
Bracket, or bucketed, retention
A user counts as retained in a period — week 1, week 2, month 3 — if they did anything at all inside it. This is what nearly everyone means by "weekly retention", and it is what most dashboards show by default.
Brackets are forgiving and readable, and they hide the boundary problem: a user active on Sunday and the following Monday appears in two consecutive weeks, while a user active twice on Wednesday appears in one. Fixed calendar weeks also mix cohorts with different amounts of elapsed time. Anchoring brackets to each user's signup date instead of the calendar removes most of that noise.
| Definition | Answers | Best for | Typical day-7 reading |
|---|---|---|---|
| N-day | Did they come back on this exact day? | Daily-use products | Lowest |
| Bracket | Were they active at all this period? | Weekly or monthly products | Middle |
| Unbounded | Have they stopped coming back yet? | Churn and lifetime questions | Highest |
The denominator decides the story
Two more choices matter as much as the curve type, and they are usually made by accident.
What starts the clock? Account creation includes people who never used the product at all, which drags every curve down and makes marketing changes look like product changes. Anchoring to first meaningful use instead measures the product rather than the funnel in front of it. Both are defensible. Mixing them between two charts is not.
What counts as coming back? Opening the app, doing anything, or doing something that matters? "Any event" quietly counts background pings, push-notification opens and the session where someone came back to cancel. A retention curve built on a meaningful action is lower, more stable, and far more useful for deciding what to build.
Only one thing on the curve matters
The height of week 1 is mostly acquisition quality. The interesting question is whether the curve flattens.
A curve that decays and then goes flat has found a group of people for whom the product is genuinely part of their routine. The flat portion — its height and how early it arrives — is the sustainable core of the business, and it is what determines whether growth compounds or leaks. A curve that keeps sloping downward has no floor: every cohort eventually goes to zero, and growth is a treadmill.
This is why comparing single points across cohorts is so misleading. Two cohorts can have identical day-7 retention and completely different shapes by day 60. Always plot the curve, always plot several cohorts on the same axes, and always exclude cohorts too young to have reached the right-hand edge.
Segment before you conclude
An aggregate retention curve is a weighted average over populations with different shapes. The most common finding when you split it is that one segment is flat and healthy while another decays to nothing, and the blend looks like slow steady decline that belongs to nobody.
Split by acquisition source first — paid and organic cohorts almost never retain alike — then by platform, plan and whether the user arrived alone or as part of a team. If one segment has a flat curve, the strategic question is usually how to acquire more people like them, not how to rescue the segment that never had a floor.
Revenue retention is a different animal
Net revenue retention answers a related but separate question: of the revenue you had from a cohort a year ago, how much do you have now, including expansion and contraction? It can exceed 100% while user retention is poor, because a handful of growing accounts can carry a cohort that is losing members.
Both numbers are real; they just describe different risks. Logo and user retention tell you whether the product holds people. Revenue retention tells you whether the accounts that stay are worth more over time. Report them side by side, and be suspicious of any deck that shows only the more flattering one.
The takeaway
Pick the curve that matches how your product is meant to be used, anchor it to first meaningful use, require a meaningful action to count as a return, and judge it by where it flattens rather than how high it starts. Then write that definition next to every retention number you publish — most disagreements about retention are disagreements about the denominator.