DM6: Analysing behaviour: metrics, funnels, cohorts and attribution
Cohorts and retention
🎯You'll be able to
Group customers by when they started, build a retention table, and read it without fooling yourself.
🏠The everyday picture
A coaching class that wants to know whether its new teaching method keeps students compares this year's batch with last year's, month by month, not all students mixed together.
📘What it really is
- A cohort is a group that shares a starting point: customers whose first order was in September, people who signed up in a campaign.
- A retention table has cohorts as rows and periods since start as columns, with the share still active (for example, ordered again) in each cell.
| First order | Customers | Month 1 | Month 2 | Month 3 |
|---|---|---|---|---|
| July | 400 | 18% | 11% | 9% |
| August | 450 | 21% | 13% | — |
| September | 520 | 24% | — | — |
- Read down the columns to compare cohorts at the same age ("Month 1 rose from 18% to 24%"), and across rows to see each cohort fade.
- Recent cohorts have fewer columns: don't compare a September cohort's month 1 with July's month 3.
- Seasonality: a festival cohort (Diwali buyers) may behave differently from a monsoon cohort.
- Small cohorts are noisy: a cohort of 40 customers swings a lot; show the count beside each rate.
- Cohorts beat averages: an overall "repeat rate" mixes old and new customers and can rise just because the shop is older.
📝A worked example
Input: the table above (illustrative figures).
Steps:
- Compare month 1 down the column: 18%, 21%, 24%.
- Check what changed: MittiMart started including a care card and a reorder link in August.
- Write the finding with its limits.
Expected output: "Month-1 repeat orders rose from 18% (July, 400 customers) to 24% (September, 520). The care card started in August, but September also had pre-festival demand, so the card isn't proven to be the cause."
🌍Where it's used
- Subscription apps track monthly retention by sign-up month.
- Shops compare repeat purchase by first-order month and by channel.
- Banks follow card activation and use by the month a card was issued.
📝Case scenario
A grocery app's overall repeat rate was flat for a year. A cohort table showed new customers from discount campaigns rarely returned, while referral customers returned twice as often. Budget moved from discounts towards referrals (DM9).
🖱️Try it
Sketch a cohort table for something you do regularly: for example the month you joined a gym, an app or a library, and whether you still used it in later months. Success: rows are start months, columns are months since start.
⚠️Common mistakes
- Comparing cohorts at different ages. Sign: month 3 of one cohort against month 1 of another. Fix: compare along the same column.
- Hiding cohort sizes. Sign: only percentages. Fix: show the count beside each rate.
- Claiming a cause from a cohort change. Sign: "the care card raised retention." Fix: say what else changed, or test it (DM7).
🛠️For your project
Your capstone can include a cohort view in its design notes; the kit's data covers only a two-week test, so a full retention table would be illustrative.
✅Quick check
What's a cohort?
Show the answer
A group that shares a starting point, such as the month of their first order.
Why show the cohort size beside each rate?
Show the answer
Small cohorts swing a lot, so readers need the size to judge the rate.