पाठशाला Pathshala · हिसाब Hisāb, Unit economics · Lesson 11 · Build
Cohort analysis for founders who are not analysts
One table, built from the payments export in an afternoon, shows whether customers stay, whether they spend more over time and when each month’s customers pay back what they cost.
Pathshala, The Founder Library · 11 October 2026 · 6 min read
An average hides the thing a founder most needs to see. Churn of 6 per cent a month for the company can mean every customer is slowly drifting away, or that the customers who arrived in a festive-season promotion left in a body while everyone else stayed. The two call for opposite decisions. A cohort table separates them, and it can be built from the payments export in an afternoon by anyone who can use a pivot table.
This lesson builds the table step by step from raw transactions, shows how to read the three numbers it holds, retention, revenue retention and payback, sets out what healthy looks like according to the people who publish it, and ends with the monthly hour that keeps the table current. The figure in the middle builds a table from a list of payments in front of you.
What a cohort is, and why averages lie
A cohort is a group of customers who started in the same period, usually the same month. The table puts cohorts down the side and the months since they started across the top, and each cell holds what that cohort did that many months in. Sequoia’s data science team, in its essay on retention, describes how to read the result: each row is one cohort, each column is a fixed age since sign-up, and a pattern running diagonally usually comes from something that happened to every cohort on the same calendar date, a release, a price change, an outage.
The company-wide average blends young cohorts, which churn fastest, with old ones, which have already lost their tourists. When a company is growing quickly most of its customers are young, so the blended churn looks worse than the product deserves; when growth slows the mix ages and churn appears to improve without anything changing. Only the table shows which is happening.
From raw transactions to the table, in four steps
Step one: export. From the payment gateway, the billing system or the accounting software, export one row per successful payment with three columns: customer ID, payment date and amount ex-GST. Remove refunds and reversals, or net them against the original payment. Use a stable customer ID, not an email address that changes.
Step two: assign each customer a cohort. Add a column holding the month of each customer’s first payment. In Excel or Google Sheets, MINIFS over the date column for the same customer ID gives the first date; EOMONTH or a month-start formula turns it into a month. Every row for that customer carries the same cohort.
Step three: compute the age of each payment. Add a column with the number of whole months between the cohort month and the payment month; DATEDIF with the "M" unit does it, or twelve times the year difference plus the month difference. A payment in the first month has age zero.
Step four: pivot. Build a pivot table with cohort as rows and age as columns. Make one with a distinct count of customer IDs and one with the sum of amount. Then divide each row by its own age-zero cell. The first is logo retention, the second revenue retention. That is the whole method, and it is the same one the figure below runs on its synthetic payments.
Switch between the three views. Logos shows how many customers each cohort keeps. Revenue shows what they pay, and climbs where some customers upgrade. Payback divides each cohort’s cumulative gross profit by what it cost to acquire, and turns green in the month it has paid for itself; move CAC and gross margin and watch the green move.
Reading the table: three numbers
Logo retention and churn. Read down a column to compare cohorts at the same age and along a row to watch one cohort decay. Monthly churn for a cohort is one minus the ratio of consecutive cells. A Jaipur company selling scheduling software to clinics wins 40 customers in January; 34 are paying in February, 31 in March and 30 in April. Month-three logo retention is 75 per cent, first-month churn is 15 per cent and the next two months lose about 9 and then 3 per cent. The shape, a steep first month and a flattening tail, is the one the [retention curves lesson](/library/retention-curves-and-flattening-test) teaches you to hope for.
Revenue retention. The same January cohort paid ₹2.4 lakh in its first month and ₹2.3 lakh in April, because a few clinics added doctors. Revenue retention at month three is 96 per cent against logo retention of 75: the customers who stay are growing. When revenue retention runs above logo retention the business has expansion, and when it crosses 100 per cent for a cohort that cohort is growing on its own.
Payback by cohort. Suppose winning the January cohort cost ₹6 lakh, ₹15,000 a clinic, and gross margin is 70 per cent. Cumulative gross profit is about ₹1.7 lakh after the first month, ₹3.3 lakh after the second, ₹4.9 lakh after the third and ₹6.5 lakh after the fourth. The cohort pays back in its fourth month. Do the same for every cohort and you can see whether payback is getting shorter as the company learns, or longer as it reaches beyond its best customers.
What healthy looks like
Andreessen Horowitz’s 16 More Startup Metrics names two healthy trends: retention in each cohort stabilising after a period such as six or twelve months, and newer cohorts performing progressively better than older ones. The first means the product has a core of customers who stay; the second means the company is getting better at finding and keeping them. Its companion essay, 16 Startup Metrics, asks for engagement expressed as cohort retention on the metrics that matter for the business, which for a paid product is payment.
For levels, Lenny Rachitsky’s survey What is good retention gives six-month benchmarks: around 40 per cent is good and 70 great for consumer subscriptions, around 60 good and 80 great for small and mid-market business software, and higher again for enterprise. Use them as a sense of scale, not a pass mark; a clinic-software company in tier-two cities and a consumer app in metros are different businesses, and the most useful comparison is always your own newer cohorts against your older ones.
The blended number tells you how the company is doing. The cohort table tells you why, and which customers to go and talk to.
Five ways to build it wrong
Annual plans counted as twelve monthly payments, or as one. A customer who prepays a year appears once in the export and then vanishes for eleven months. Spread annual payments across the months they cover before pivoting, and track logins alongside payments for those customers.
Reactivations treated as new customers. A customer who leaves in March and returns in July belongs to their original cohort with a gap, not to July. Cohort by first payment ever, not first payment this year.
Free trials in the denominator. For a paid product the cohort starts at the first payment. A trial cohort is useful too, but it is a conversion table, not a retention table.
Cohorts too small to read. A cohort of eight customers swings twelve points when one leaves. Group into quarters until each cohort holds at least thirty.
Revenue that is not the customer’s. Exclude GST, one-time setup fees and pass-through costs, so that revenue retention measures what the customer chooses to keep paying for.
The monthly cohort hour
On the first Monday after the month closes, refresh the export, rerun the pivot and add the new column. Then look at three things in order. Read the newest diagonal: did every cohort drop at once last month? If so find the event. Read the month-one and month-three columns from top to bottom: are newer cohorts holding better than older ones? Read the payback view for the last three cohorts: is the month they pay back moving earlier or later?
Write one sentence for each answer and one action, and pick five customers from the cohort that looks worst and call them this week. The table is the cheapest research instrument a company owns; it only needs someone to read it every month and act on what it shows. Feed the churn and payback you find into the [ten-input model](/library/one-spreadsheet-model-every-founder-should-build) so that the forecast learns from the customers too.
Nothing here is legal, tax or investment advice. The Jaipur cohort and the figure’s payments are illustrative.
Sources
- Sequoia Capital data science team, Retention, Sequoia Capital — Triangle retention charts: rows as cohorts, columns as ages, diagonals as calendar events; flattening, declining and smiling curves.
- Anu Hariharan, Frank Chen and Jeff Jordan, 16 More Startup Metrics, Andreessen Horowitz, September 2015 — Healthy cohorts stabilise after six or twelve months; newer cohorts outperform older ones.
- Jeff Jordan, Anu Hariharan, Frank Chen and Preethi Kasireddy, 16 Startup Metrics, Andreessen Horowitz, August 2015 — Engagement ideally expressed as cohort retention on the metrics that matter for the business.
- Lenny Rachitsky, What is good retention, Lenny’s Newsletter, June 2020 — Six-month retention benchmarks: consumer SaaS ~40 per cent good, ~70 great; SMB and mid-market SaaS ~60 good, ~80 great.