A cohort chart earns its space on a dashboard by showing that two groups behave differently. If every curve traces the same path with a little noise, you have built an expensive way to display your average retention rate.
Signup month, the default cut in most tools, is usually the worst offender. It separates when something changed, which is occasionally what you want and mostly is not.
Three cuts that tend to work better.
By acquisition path
Accounts that arrived through a search for your category behave differently from accounts that arrived through a colleague’s invitation, and both behave differently from accounts your sales team sourced. The gaps here are frequently twenty points or more at ninety days.
The value is not the chart. It is that the cost of an acquisition channel becomes arguable in retention terms rather than in cost per signup, which changes where the budget goes.
By first-week depth
Not whether an account did the activation event, but how many distinct meaningful actions they took in week one. Bucket into one, two to three, and four or more.
This almost always separates cleanly, and unlike acquisition path it points at something you can influence. If the four-or-more bucket retains twice as well, the onboarding question stops being how do we get them through the wizard and becomes how do we get them to a fourth action.
By seat count at day thirty
For anything sold per seat, the number of humans actually using the product a month in is a blunt instrument that works embarrassingly well. Single-seat accounts churn. Accounts with four or more rarely do.
Resist the obvious conclusion, which is to push seats. Seat count is usually a symptom of whether the product became part of a team’s routine or stayed one person’s side experiment. Pushing seats without the routine produces accounts with four inactive users and the same churn.
The rule
Before you put a cohort chart anywhere permanent, check the spread at your reporting horizon. If the best and worst cohorts sit within ten points of each other, drop the cut and try another. The chart is not the deliverable. The separation is.
