Meerkats AI
← All posts
Customers·21 May 2026·8 min read

Cohort analysis: know your buyers better than the platforms do

Blended averages are where truths go to hide. Three cohort views — by month, by first product, by channel — that show whether your brand is compounding or quietly decaying.

TL;DR
  • Blended metrics mix your best old customers with your newest ones — repeat rate can look stable for months while every new cohort is worse than the last.
  • Three splits do most of the work: acquisition month (health trend), first product (which SKU creates loyalists), acquisition channel (who sends buyers vs tourists).
  • The action isn't reporting — it's re-pricing acquisition: two channels with identical CPAs can deserve completely different budgets once cohort quality is priced in.

A brand's blended repeat rate held at 31% for three straight quarters — stability everyone read as health. The cohort view told the real story: customers acquired 18 months ago repeated at 45% and were slowly aging out, while last quarter's cohort was repeating at 19%. The blend averaged a strong past against a weak present. By the time the blended number finally moved, the damage was six months deep.

What a cohort actually is (fifteen seconds)

A cohort is every customer who did a thing in the same window — usually "first order in June." You then follow that group forward: how many bought again by day 30/60/90, what their cumulative contribution looks like. Nobody enters or leaves the group, which is exactly why it can't lie the way an average can.

View 1 — by acquisition month: is the machine improving?

CohortBuyersRepeat by day 90Contribution/buyer by day 90
February1,14027%₹492
March1,38024%₹455
April1,90520%₹410
May2,45016%₹360

Read the diagonal: volume doubling while day-90 quality drops a third. This is the signature of scaling into weaker audiences — each incremental rupee buys a worse customer. Blended dashboards celebrate this exact chart ("record new customers!"). The cohort view prices it: May's buyers are worth ~27% less each, so May's real CAC ceiling should have been ~27% lower. Was it?

View 2 — by first product: which SKU creates loyalists?

From a real audit: buyers whose first order was the mango product repeated at roughly twice the rate of buyers who entered on honey. Same store, same ads budget — one entry product creates customers, the other creates transactions. The implications cascade: hero the loyalist-maker in prospecting creatives, gate the tourist product behind bundles, and judge campaigns by which product's buyers they recruit, not just how many.

View 3 — by channel: who sends buyers vs tourists?

ChannelCPA90-day repeat90-day contributionVerdict
Meta prospecting₹1,10024%₹520Real customers — defend it
Google brand₹35038%₹640Great cohort, but see incrementality — many were coming anyway
Discount-code influencer₹7009%₹310Cheap tourists — the CPA is a trap

The middle column is the one platforms will never show you, because order two happens outside their window and off their properties. Two channels at identical CPA can deserve opposite budget decisions — cohort quality is the missing price tag.

Reading a cohort triangle without a data team

  • Down a column (same age, successive cohorts): the machine's health trend. Down-and-worsening = the February–May table above.
  • Along a row (one cohort aging): where repeat flattens — that's your realistic LTV horizon, not the aspirational one.
  • The two traps: judging young cohorts too early (give them 60–90 days), and celebrating a great month that was really one viral spike of tourist buyers.
The operating loop
Monthly, three questions: Is cohort quality trending up or down? Which first-product and channel made the best 90-day buyers? Does each channel's CAC ceiling reflect its cohort value? That's the entire practice — twenty minutes, if the joins exist; impossible, if they don't.

FAQ

How is this different from RFM segmentation?
RFM scores customers today for targeting; cohorts follow groups through time for diagnosis. RFM tells you who to email this week; cohorts tell you whether acquisition is buying good customers at all.
Shopify shows a returning-customer rate. Isn't that enough?
That's the blend — the exact number that hid the decay in our opening example. You need repeat by acquisition month, which Shopify's native reports don't give you directly.
How many customers before cohorts are readable?
Rule of thumb: 150–200+ per cohort for repeat-rate movements to mean anything. Smaller brands should cohort quarterly instead of monthly rather than reading noise.
What tool do I need?
Mechanically it's one join: orders × customers × first-order attributes (month, product, channel from the ad-click match). Spreadsheet-possible, painful monthly — which is why it belongs in the same automated spine as CAC and MER.

See your real numbers, not the platform's.

Connect Shopify, Meta and Google. Reconciled truth in 20 minutes.