How to Run a Retention Cohort Analysis in Mixpanel That Matters
Retention in Mixpanel is defined by two events: a birth event that starts the clock and a returning event that counts as retained. Get those two wrong and the curve is meaningless, no matter how it is rendered. Start with signup and your first core action, then compare weekly cohorts.
Most retention reports are unreadable because the team never decided what “using the product” means. That is a product question, not an analytics one.
What Events Should I Use for Retention?
The birth event should mark the start of a user’s relationship: Account Created or Trial Started. The returning event should be your core value action — the thing that makes the product worth paying for. For a project tool that is Task Completed; for a writing app, Document Edited.
Using Session Start or a login event as the returning event produces flat, flattering curves. Logging in is not value. Every product I have seen with a suspiciously high retention number was counting logins.
How Do I Read the Retention Curve?
Three readings matter, in order.
| What to look at | What it tells you |
|---|---|
| Day 1 to Day 7 drop | Onboarding quality |
| Week 8 to Week 12 flattening | Whether a real habit formed |
| Cohort-to-cohort uplift | Whether product changes worked |
The shape beats the number. A curve that drops to 20% by day 7 and holds flat at 18% by week 12 is healthier than one at 35% by day 7 that decays to 4% by week 12. The flat tail is the durable user base.
Should I Use N-Day, Unbounded, or Week-Based Retention?
Mixpanel offers all three. N-day counts a user as retained only if they return on exactly day N. Unbounded counts them if they return on or after day N. Week-based buckets the same way over seven-day windows.
Unbounded looks better and is less useful for products with weekly usage patterns — email or invoicing tools get inflated scores. Use N-day for daily-use products and week-based for weekly ones. Then keep it consistent, because switching mid-quarter makes every comparison invalid.
How Do I Split Cohorts to Find What Actually Works?
Break the report down by a property captured at signup or at first use. The useful splits are behaviour, not demographics:
- Did the user complete setup in the first session?
- Which acquisition channel brought them?
- Did they invite a teammate in week one?
- Which plan or trial type did they choose?
Invite-a-teammate splits tend to show the biggest gap in B2B products — often 3-5x week-8 retention. That finding usually survives the correlation-versus-causation objection better than most, though forcing invites is not guaranteed to work.
How Do I Use Cohorts Without Getting Fooled?
Compare like windows. A cohort created three weeks ago has less data and always looks better at week 4 than a cohort from nine months ago — survivorship hides in the shorter window.
Also watch mix effects. If you changed pricing in March, every cohort after March reflects the change, so a product comparison across that line measures price, not product.
Finally, the free plan keeps 90 days of history. Cohort analysis is the report most damaged by that limit — a 12-week retention curve barely fits, and year-over-year comparison is impossible without upgrading.
When Is Retention Analysis a Distraction?
Before you have 500 users. At 50 users per cohort, the noise is larger than any trend you will detect, and three bad signups move the curve by six points. Ship, instrument, and revisit retention once the denominators are above a few hundred.
Quick Summary
- Retention needs a birth event (signup) and a core-value returning event, not a login.
- The flat tail of the curve matters more than the day-7 number.
- Use N-day for daily products, week-based for weekly ones, and never switch mid-quarter.
- Behaviour splits (setup completed, teammate invited) outrank demographic ones.
- Below 500 users per cohort, the noise exceeds the signal.
