Cohort

How it works

Pick a starting event and a time bucket. Everyone who installed in the first week of March is one cohort; everyone who installed in the second week is another. Membership never changes after that. A cohort report then lines the groups up by age, counting day 0 as the start, and shows what each one did on day 1, day 7, day 30 and so on.

The alternative is a calendar report, which shows everything that happened on a date regardless of when the users arrived. Calendar totals mix new and old users, so a rise in revenue could come from last month's installs or last year's. A cohort view separates them.

When it matters

Use cohorts whenever the question is about quality over time. Say you changed your onboarding in April. Comparing the March and April cohorts at day 7 tells you whether the change helped, because both groups are measured at the same age. Cohorts are the base of retention reports and of LTV estimates, which both ask what a fixed group did after it arrived.

Small cohorts are noisy. A group of forty users can swing widely from week to week, so widen the bucket from days to weeks before you read a trend into it.

How it relates to attribution

An MMP adds the source to the cohort. Because each install is credited to a channel and campaign, you can split one install week into a cohort per campaign and compare them. That turns a count of installs into a view of which campaign's users stayed and paid, which is the comparison you need when you decide whether an MMP is worth adding. Check that a vendor's reports group this way when you send the ten questions for every vendor.