Analysis
Cohort analysis
Also called retention table.
Cohort analysis is watching how a fixed group behaves in later periods rather than blending all users together. The usual output is a retention table with one row per group.
How it is measured
Build rows from cohorts by start period and columns from time since start: week 0, week 1, week 2. Each cell is the share of that row's group who did the action in that later week.
Read it along the diagonal as well as across. The diagonal shows what happened in the same calendar week to different groups, which exposes a release or an outage that hit everyone together.
Worked example
A meal-kit startup builds a table of weekly signup cohorts. The week-4 reorder rate runs 31, 30, 29 percent for three weeks, then drops to 19 percent for the cohort that started the week of a price change.
The same drop does not show in the sitewide order total, which kept growing from new signups. Only the table showed that recent customers behaved worse than earlier ones.
How it differs
Cohort analysis follows groups across time. A cohort is one such group. Another view, such as a plain trend line, mixes people who joined in different weeks.
Common errors
Reading old and new users as one. Comparing cohorts of different size without noting it. Using calendar months of different length. Ignoring seasonal starts. Declaring a result from one small row.
In practice
Build one retention table for your most important repeat action. Check the newest row against the older ones every week. If one row differs, look up what changed in that start week.