Reading retention analysis

Retention answers one question: of the people who started on a given day, how many are still here later? It can be more specific too — of the people who did one thing, how many come back to do another.

Where: Product analytics › Behavior › Analysis, with the analysis type set to Retention.

Setup

Start and return events are both optional:

• Neither: any activity counts. You're asking whether new users are still active days later.
• Start event only, e.g. "completed sign-up": retention of registered users, without visitors who only browsed.
• Both, e.g. start "first order" and return "order": repeat purchase.

Period: day, week or month. Up to 30 daily periods or 12 weekly or monthly periods; the default is 14 days.

Three counting methods

• Returned in period N — only whether they came back in that exact period. The curve can dip and rise again, since some people skip day 3 and return on day 7.
• Returned in period N or later — that period or any later one. The curve only goes down; use it to see how many you keep in the end.
• Number of periods returned — a distribution of how many periods each person came back in, showing usage frequency.

When in doubt, use the first. It's the most common, and the easiest to compare with other people's numbers.

Reading the table

• Each row is a cohort: start date, size, and how many came back in each period, as a count and a share.
• Cells that haven't matured are empty, not zero. A cohort that started today has no day-7 figure until a week from now. Empty and 0% mean very different things.
• The average line below is weighted by cohort size, so large cohorts move it more.

Common shapes

• Fast drop, then flat. The healthy shape. The height where it flattens is your core-user share; the higher the better.
• Keeps falling, no plateau. Users haven't found a reason to keep coming. Find what the people who stayed did differently — a feature they used in week one, for example.
• One cohort clearly above or below the rest. Compare with releases and campaigns at the time; the cause is often there. Add a date annotation so you can find it again later.

What did the people who stayed do right?

Save the people still active in week 4 as a segment and compare their first week with people who left in week 1: which features they used, which pages they read. The behaviour with the biggest difference is the one to push in onboarding.

Retention vs lifecycle

Retention shows the share of one cohort in period N. Lifecycle shows, for each period, how active users split into new, retained, resurrected and churned — better for watching the shape of your whole user base change. Switch between them on the same analysis page.

Prerequisite

Retention has to recognise the same person over time. With the project's privacy set to strict, the visitor identifier changes daily and retention can't be calculated. With identity mode on, retention by signed-in user is the most accurate — see Connecting anonymous visitors to signed-in users.