Retention Curves
By the end of this lesson, you should be able to: measure retention three different ways and know which one fits which question, read what the shape of a retention curve is telling you, read a cohort table, and avoid the mistake that makes new cohorts look worse than they are.
Retention usually sets the ceiling on how big a product can get. Duolingo found that improving current user retention had the largest impact on its daily users.
What retention measures
Retention asks a simple question: of the people who started, how many are still here?
To answer it, you group users into a cohort, a set of people who started at the same time, like everyone who signed up in the first week of March. Then you follow that one group over time and count how many are still active on day 1, day 7, day 30 and so on.
Why a cohort, and not all users at once? Because "all users" mixes people who joined yesterday with people who joined two years ago. Following one group that started together is the only way to see how people behave as they age with the product.
Three ways to define "still here"
There are three standard ways to define retention, and they give different numbers for the very same cohort.
1. N-day retention. Was the user active on exactly day N? For day 7: did they show up on day 7 itself?
2. Unbounded retention. Was the user active on day N or any day after it? This counts anyone who hasn't left for good yet.
3. Bracket retention. Was the user active at any point in a window, like days 7 to 13? This is how weekly retention is usually measured.
Say you're measuring day 7 for someone who used the app on days 0, 3 and 12.
| Definition | Counts this user as retained on day 7? | Why |
|---|---|---|
| N-day | No | They weren't active on day 7 itself |
| Unbounded | Yes | They came back later, on day 12 |
| Bracket, days 7 to 13 | Yes | Day 12 falls inside the window |
So which is right? It depends on the product and the question.
- N-day fits products people use every day, like messaging or games, where missing a specific day really means something.
- Bracket fits products used every week or so, like food delivery, where missing one exact day means nothing.
- Unbounded fits questions about whether people have left for good, since it only drops users who never come back.
Whichever you pick, say which one it is, because two people using different definitions can report very different numbers for the same product.