Retention Curves
By the end of this lesson, you should be able to: name the three standard retention definitions and say which question each answers, read a curve for the only thing that matters about its shape, build and read a cohort triangle, and spot the censoring mistake that makes most triangles lie.
Three definitions, one cohort, three different answers
Take the 7,415 Basket users who have a full 56 days of history. Ask a simple question: how many were retained on day 30?
| Definition | What it counts | Day 1 | Day 7 | Day 30 |
|---|---|---|---|---|
| Classic day-N | Active on exactly day 30 | 21.9% | 17.7% | 10.9% |
| Rolling 7-day | Active at any point in days 30 to 36 | 59.8% | 54.1% | 41.9% |
| Unbounded | Active on day 30 or any day after | 88.7% | 84.9% | 69.8% |
11%, 42%, and 70%. The loosest answer is 6.4 times the strictest, on the same users, the same table, the same day. If lesson 1 felt familiar, this is the same problem wearing different clothes.
None of them is wrong. They answer different questions:
- Classic day-N suits products with a daily rhythm. A news app should be opened on day 30. A grocery app has no reason to be, so day-N retention makes Basket look near-dead when it isn't.
- Rolling suits products with a weekly or monthly rhythm. This is the right one for Basket, because "did you shop this week" is the real question.
- Unbounded flatters everything and is the only honest one for churn, because it's the only definition where a user who comes back was never really gone.
Pick from the product's rhythm, say which you picked, and never compare a number across two of them.
The shape matters more than the level
Here's the thing to actually look for, and most people miss it because they're reading the height instead of the slope.
Basket's rolling retention, week by week, with how much it fell each time:
| Week | Retained | Change |
|---|---|---|
| 1 | 54.1% | |
| 2 | 48.3% | −5.9 pp |
| 3 | 44.2% | −4.1 pp |
| 4 | 42.7% | −1.5 pp |
| 5 | 40.0% | −2.7 pp |
| 6 | 39.9% | −0.1 pp |
| 7 | 37.9% | −2.0 pp |
The drops shrink. Nearly six points in week two, then four, then one and a half, then wobbling around two. The curve is flattening, and that's the single most important thing a retention chart can tell you.
A curve that flattens has found a group of people for whom the product genuinely works. They fade a little and then they stay. A curve that keeps falling at the same rate has no such group, and every user you acquire is on a conveyor belt to the exit. The first product has something to grow; the second has a bucket with no bottom, and pouring more users in is setting money on fire.
When an interviewer shows you a retention curve, don't open with the level. Open with "is it flattening, and where". The level tells you about the product category. The asymptote tells you whether there's a business.