Cohort analysis for small business, without a data team

Aug 22
3 min read

Updated: 4 days ago

Introduction

A cohort is simply a group of customers who arrived at the same time. Cohort analysis means following each group forward separately instead of blending everyone into one average.

It sounds like something requiring an analyst. It requires a spreadsheet and the date each customer first bought.

1. Why blended averages hide the truth

Suppose your repeat purchase rate is 30% and has been for a year. That looks stable.

Now split it by when customers arrived. Customers from January repeat at 45%; customers from November repeat at 12%. The average is unchanged and the business has quietly got much worse — recent customers are not coming back, and older loyal customers are propping up the figure.

A blended number cannot show that, because new arrivals are always mixed in with people who have had years to return.

2. Cohort analysis for small business needs three columns

From your own records: customer, date of first purchase, dates of later purchases. That is it.

Group customers by the month they first bought. For each group, count how many bought again within 30 days, 60, 90. Put the groups down the rows and the time windows across the columns.

The result is a triangle of percentages, and reading down any column tells you whether the business is improving or decaying.

3. Read down the column, not across the row

This is the part people get wrong. Reading across a row just shows that customers buy more over time, which is unsurprising.

Reading down a column compares like with like: every cohort at the same age. If the 30-day repeat rate falls from 40% for January arrivals to 20% for June arrivals, something changed around the arrival experience — and you can date it.

That dating is what makes cohorts worth the effort. A blended figure tells you there is a problem; a cohort table tells you roughly when it started.

4. Use it to judge whether a change worked

This is the most practical use. You change something — a follow-up message, an entry offer, an onboarding step — and you want to know if it helped.

Compare the cohorts who arrived before the change with those who arrived after, at the same age. If the later cohorts retain better at 30 days, the change worked. If they do not, it did not, whatever the overall monthly total is doing.

Without cohorts, a busy season will happily take credit for a change that did nothing.

5. Match the window to your purchase cycle

Thirty, sixty and ninety days suit businesses bought from monthly. They are meaningless for a service bought annually.

Set the windows to roughly one, two and three purchase cycles for your business. A café might use weeks. A training provider should use terms. An annual service should use years and accept that the table takes a long time to fill.

6. Segment by acquisition source once the basics work

The next layer, and worth it: build the same table per channel.

Channels differ enormously in the quality of customer they produce, and cohorts are how you see it. A channel producing cheap customers who never return is usually your most expensive, and this is the report that proves it rather than implying it.

7. Keep it in the same sheet as everything else

Cohorts belong beside acquisition cost and lifetime value, not in a separate file.

Once the sales export is already feeding your reporting, the cohort table is a pivot away and recalculates itself. Free tools handle this perfectly well; the work is in getting first-purchase dates recorded reliably, which is a discipline problem rather than a software one.

Conclusion

Group customers by the month they arrived, track each group forward across fixed windows, and read down the columns rather than across the rows.

Match the windows to your purchase cycle, use the table to judge whether a change actually worked, and segment by channel once the basics are running. It needs three columns and a spreadsheet — and it answers the one question a blended average never can.