Choosing the cohort start event: install or first purchase

What is a cohort start event?

A cohort start event is the action that puts a user into a cohort and fixes their Day 0, so every later Day N is counted from that moment. It is the first choice to make in any cohort analysis.

Start event: the anchor action, such as install, first open, sign-up or first purchase. The day it happens is Day 0 for that user.

Return event: the action that counts a user as active on a later day, such as any app open or a meaningful action like placing an order.

Every retention number on this site, whether classic, rolling or cumulative, divides users retained on Day N by the users who did the start event. Change the start event and you change both the top and the bottom of that fraction. Day 0 activity is never a return, whatever anchor you choose.

The four common start events

Pick the question first, then the event.

• Install cohorts are the default for acquisition work, because they include the people who bounced on Day 0.
• First-purchase cohorts answer a customer question such as repeat purchase or renewal. They are a product and CRM view, not an acquisition view.

The longer argument for keeping both an install view and a purchase view is set out in this guide to defining cohorts by first install or first purchase.

Why do first-purchase cohorts show higher retention?

Because a first-purchase cohort leaves out everyone who never bought, it only contains your most committed users, so its retention is always flattering next to an install cohort.

This is survivorship bias. Hypothetically: say 1,000 people install on Monday and 100 of them buy that week. If 40 of those buyers open the app on Day 7, the first-purchase cohort shows 40% Day 7 retention. The install cohort for the same Monday might show far lower, because it also counts the 900 who never bought and mostly drifted away.

• Neither number is wrong; they describe different groups.
• Never put an install-cohort curve next to a first-purchase curve and call one better.
• If you report both, label the start event on every chart and table.

Choosing the return event

The return event matters as much as the anchor.

• Any open is the broadest choice. It is easy to compute and easy to compare across cohorts and channels.
• A meaningful action (a search, an order, a lesson finished) is stricter and closer to value. It will always give lower numbers than any open.
• Count a user once per day, however many times they do the event that day.

Keep the return event fixed across cohorts and channels, and start a new series if you change it. The SQL and spreadsheet formulas on this site count an app open as the return; to use a different action, filter the activity table to that event first.

Day boundaries and time zones

Day N has to mean the same thing for every user, and there are two ways to define it.

• Calendar days: Day 1 is the next calendar day after install, in a fixed time zone. A user who installs at 11:50 pm can be a Day 1 return ten minutes later.
• 24-hour windows: Day 1 is the 24 to 48 hours after the install moment. Fairer to late-night installs, harder to line up with daily reports.

Practical rules:

• Pick one time zone for the whole report and state it.
• A cohort only has a Day N value once Day N has fully passed for every user in it.

Assigning the cohort to a channel

To compare channels you need each user's acquisition source attached to their start event.

• For install cohorts, attach the attributed channel or campaign at install.
• For later anchors such as first purchase, keep the original install channel. Otherwise a paid user who buys after an email looks like an email user.
• Watch the share of installs from each channel over time, because a shift in mix moves blended retention even when each channel is flat. The channel mix section of the reading guide shows this with the site's sample app, a made-up app.

Events attached to installs are also what ad networks optimise on, which is covered in this explainer on events, postbacks and ROAS. For channel-level reading, see retention by channel.

A decision table

Try different definitions on the interactive cohort explorer to see how much the choice moves the numbers.