Cohort Analysis for Founders

What is a cohort?#

A cohort is a fixed membership set created by a rule. A January paid cohort might include every account whose first recurring invoice started in January. "Month 3" then means the third completed month after each account's own start date, not March for everyone.

Three measures do most of the work:

The gap between GRR and NRR is the expansion story. The gap between logo retention and GRR is the account-size story: losing many small accounts and keeping a few large ones produces weak logo retention and strong GRR.

Definitions must specify whether reactivations, usage overage, foreign exchange, acquired customers, and one-time services are included. Public companies disclose these choices precisely because the choices move the number. Vertex, for example, reports NRR of 105% and GRR of 94% for the same customer base at the same date — one number above 100% and one below, describing the same year. (Vertex Q4 2025 results)

The exact policy matters less than having a stable, versioned, auditable one. Never silently change the definition when the chart becomes inconvenient.

Why does cohort analysis matter to founders?#

It separates acquisition from product quality. Total active customers can rise every month while each successive cohort retains worse. Only an age-aligned grid shows whether growth comes from more entrants or better survival.

It makes go-to-market changes testable. Channel, segment, contract type, implementation path, and even the closing rep produce different retention and payback. A cheaper lead source that produces customers who churn in month four is not cheaper.

It exposes changing unit economics before the P&L does. Acquisition cost belongs to the cohort that generated it. Blending an old, efficient cohort with a new, expensive motion conceals the deterioration for several quarters.

It gives fundraising claims a denominator. A single company-wide retention number cannot show trajectory. Investors want to see whether recent cohorts activate faster, retain better, and expand more than older ones — and whether the recent cohorts are old enough to say.

How do you build a cohort analysis?#

1. Write the cohort contract#

Before any SQL, write down: the entity (user, account, workspace, location, contract); the entry event and timestamp; eligibility and exclusions; the period grain (day, week, month, quarter); the outcome (activity, logos, recurring revenue, gross profit); the treatment of upgrades, downgrades, pauses, reactivations, and mergers; and the data-maturity lag. Version it.

2. Build an age-based matrix#

Rows are entry periods; columns are age since entry. Show the eligible denominator and the absolute count next to every percentage.

3. Plot survival, handling censoring properly#

Naive retention rates work only when every account has been observed for the full window. When accounts enter at different times, or some are still active at the data cutoff, those accounts are censored — you know they survived at least this long and nothing more. Treating them as churned understates retention; leaving them in the denominator overstates it.

The Kaplan–Meier product-limit estimator handles this by re-computing the at-risk population at each event time:

S(t) = Π over event times j ≤ t of (1 − d_j / n_j)

where d_j is the number of churn events at time j and n_j is the population still at risk immediately before it. Kaplan and Meier formalised exactly this treatment of incomplete observations in 1958, and it remains the correct default for young cohorts. (Kaplan & Meier, 1958)

4. Attach economics#

cohort payback month = first month in which
    cumulative cohort gross profit ≥ cohort acquisition cost

Use gross profit, not revenue — revenue does not repay acquisition cost when delivery has a cost. Include onboarding and channel cost consistently. If payback has not occurred inside the observed window, report that fact rather than extrapolating a straight line through the censoring boundary.

5. Segment only after the base view works#

Useful cuts: segment, use case, channel, geography, plan, implementation type, contract length, campaign. Stop slicing while the cells still contain enough accounts to mean anything.

Related concepts#

• Customer Acquisition Cost — match acquisition spend to the cohort it actually produced.
• CAC Payback Period — convert cumulative cohort gross profit into a recovery date.
• Customer Lifetime Value — use realised cohort survival instead of an assumed churn rate.
• ARR and MRR — define the recurring-revenue base the cohort measures.
• Churn Rate — estimate customer and revenue loss with consistent denominators.
• NRR and GRR — build the existing-customer revenue bridge.
• Gross Margin — set the delivery-cost boundary used in cohort economics.
• Product-Market Fit — read retention curves as evidence of durable value.
• Sales Funnel & Pipeline Metrics — follow acquisition cohorts from lead through renewal.
• Usage-Based Pricing — interpret revenue cohorts when consumption varies.
• Switching Costs — separate retention that reflects value from retention that reflects friction.

Sources#

1. E. L. Kaplan and Paul Meier, "Nonparametric Estimation from Incomplete Observations", Journal of the American Statistical Association 53(282), June 1958, 457–481. Original derivation of the product-limit (Kaplan–Meier) estimator and the correct treatment of censored observations. Open-access copy.
2. Vertex, Inc., "Vertex Announces Fourth Quarter and Full Year 2025 Financial Results", 11 February 2026. Primary disclosure of NRR (105%), GRR (94%), ARR ($671.0m), AARPC ($137,867), and the company's written ARR/MRR and NRR definitions.
3. Similarweb Ltd., "Similarweb Announces Fourth Quarter and Fiscal 2025 Results", 17 February 2026. Primary disclosure of overall NRR (98%) versus NRR for customers with ARR ≥ $100k (103%), customer counts (6,128 total; 454 ≥ $100k), ARR concentration (63%), and the company's ARR, NRR, CAC, and CAC-payback definitions.
4. U.S. Securities and Exchange Commission, Commission Guidance on Management's Discussion and Analysis (Release No. 33-10751), issued 30 January 2020, effective 25 February 2020. Interpretive guidance on disclosing key performance indicators: definition, method of calculation, usefulness, management use, and methodology changes.

Source-use note: Northwind and every figure in the worked example are hypothetical. The company disclosures above are primary statements by the reporting companies about their own metric definitions; they establish that such definitions vary, not that any particular definition is correct for your business.