Cohort Analysis for Marketers: Seeing What Averages Hide
Averages lie by blending: a 'stable' monthly revenue can hide new customers churning faster every quarter, propped up by an aging loyal base. Cohort analysis un-blends — grouping customers by when (or how) they arrived and following each group through time — so you see whether the machine is actually improving.
Here's how to build, read, and act on cohorts without a data science team.
Key takeaways
- A cohort is a start-group followed over time — signup month, first-purchase month, acquisition channel — measured on retention or value at each age.
- The triangle chart is the tool: cohorts as rows, age as columns, retention or revenue in the cells — patterns jump out that averages bury.
- Compare curves, not points: where the curve flattens (and at what level) is the health signal; month-one numbers alone mislead.
- Cohorts settle arguments averages can't: did onboarding changes work, which channel brings keepers, is LTV actually rising.
Build your first cohort table
Pick the start event (first purchase or signup), the grouping (monthly is the workhorse), and the measure — repeat-purchase rate, active rate, or revenue per cohort member at each month of age. Lay it out as the classic triangle: each row a cohort, each column months-since-start, each cell that cohort's value at that age. Most analytics platforms and even spreadsheets can produce this from order data. Resist the urge to over-segment immediately — time-based cohorts first, then split by channel or product once the baseline shape is understood.
Read the shapes
Down a column: are newer cohorts performing better or worse at the same age than older ones? That diagonal comparison is the 'is the business improving' answer no monthly dashboard gives. Along a row: where does the curve flatten — the plateau is your retained core, and its level times cohort size is the future you're compounding. Watch for the classic patterns: curves that flatten high (healthy, invest in acquisition), curves that decay to near-zero (leaky bucket — fix retention before scaling spend), and newer cohorts bending below older ones (quality of acquisition or experience degrading, hidden by the blended average until it's expensive).
Decisions cohorts unlock
Channel truth: cohort revenue curves by acquisition source reveal which channels deliver keepers versus one-and-done bargain hunters — reallocate CAC accordingly, because payback period per channel is a cohort number, not an attribution number. Change verification: compare cohorts before/after the new onboarding, pricing, or welcome flow at the same age — the honest A/B when randomized tests aren't possible. LTV grounding: project lifetime value from observed curve shapes instead of optimistic multiples, and let that bound what acquisition can rationally pay. The habit that matters: one monthly cohort review, same charts every time — trends across cohorts are the point, and they only appear to people who keep looking.
Frequently asked questions
What's the difference between cohort analysis and segmentation?
Segments group by attribute at a point in time; cohorts group by start moment and follow behavior across time. Cohorts answer 'is it getting better' — segments answer 'who is different'.
How many months of data do cohorts need?
Enough ages to see curve shape — typically several months minimum, more for long purchase cycles. Even partial triangles reveal whether newer cohorts beat older ones early.
Which cohort metric matters most for ecommerce?
Repeat-purchase rate and cumulative revenue per cohort member — together they show retention behavior and payback timing, the two numbers that govern how much acquisition you can afford.