Retention and cohorts
New customers fill the bucket; retention decides whether it holds water. This lesson shows you how to define retention so the number means something, build and read a cohort table, judge whether a retention curve flattens and compare it fairly with benchmarks, and measure retention for partner shops and internal tools.
By the end you can
- Define retention precisely: the active event, the period and the denominator
- Read a cohort table along rows, down columns and across diagonals
- Judge whether a retention curve flattens, and compare it with benchmarks on the same definition
- Measure retention for business customers and internal tools, including revenue and channel retention
- Write a weekly cohort retention query in SQL
The everyday picture
A language school starts a new evening class every month. The teacher keeps a notebook with one line per class: how many students still came after one week, two weeks, three weeks. She never adds the classes together into one attendance number, because the class that started last week still has everyone and would make the total look wonderful.
Comparing classes at the same week tells her whether the new textbook is helping. Once she noticed that the May class shrank twice as fast as the others. It was the one filled by a half-price promotion. And she watches for the week when attendance stops falling. The students left after that point are the ones who came to learn, and the size of that group is the real measure of her school.
Each class is a cohort. The notebook is a cohort table. The week attendance stops falling is where the retention curve flattens.
What retention measures, and three choices that change it
Retention is the share of a group who are still doing the core action some time after they started. It is the leak in the bucket. If you pour in 62,500 new customers a week and most leak out by week 4, growth spending only refills a leaking bucket.
Three choices change the number a lot, so write them next to every retention figure.
- The active event. Count the core action, such as placing an order, not opening the app. For partner shops you may want two versions: still ordering at all, and still ordering in the app.
- The period. Match how often people naturally do the job. Mangrove customers order about 1.7 times a week, so weekly periods fit. A corner shop restocks every week or two, so monthly fits partner shops. A messaging app would use days.
- "In that period" or "in that period or later". Classic retention counts a customer in week 2 only if she ordered in week 2. The other version, often called unbounded or rolling retention, counts her if she ordered in week 2 or any time after. Unbounded numbers are higher and never rise from one week to the next; classic numbers are lower and bumpier. Both are fine. Mixing them is not.
Then the denominator: everyone who signed up, or everyone who placed a first order? The second is higher, because it skips people who never started. Neither is wrong, but a number without its denominator can't be compared with anything.
Quick check
A customer orders in week 1 and week 3 after signing up, but not in week 2. Is she retained in week 2?
Classic retention asks "active in that exact period?". Unbounded asks "active in that period or later?". Her week-3 order makes her count in week 2 only under the unbounded rule, which is why the two versions give different numbers from the same data.