Cohort analysis

What a cohort is

A cohort is a group of customers who acquired a shared characteristic at the same point in time. In retail that is usually the month of the first purchase. All customers who ordered for the first time in March form the March cohort. The analysis then tracks how that group develops over the following months.

The difference from an aggregate metric is the heart of the method. A shop can report rising repeat purchases and still have a retention problem, because many new customers arrive at the same time and distort the mix. Cohort analysis separates the groups and shows whether a younger cohort really retains better than an older one.

How the report looks

The usual presentation is a matrix. Rows are the cohorts, columns the months since the first purchase. Each cell holds the share of customers who bought again in that month. Column 0 is 100 percent by definition, because every customer bought at least once in it.

• Read a row: How one group behaves over time. A steep drop in the first two months is normal, what matters is where the curve settles.
• Read a column: How a young group performs against an older one in the same month of life. This is the real quality comparison.
• Read a diagonal: The view of a single calendar month, useful for spotting one-off effects.

What typical curves mean

A strong drop after the first purchase is not a defect, it is the normal case. Many customers buy once and never again, for example with gifts, a move or a one-off replacement. What is interesting is the area behind it: a curve that stays at a stable level after month three suggests a settled customer base. A curve that keeps falling suggests that even regular customers drift away.

Effects worth knowing: seasonal items show a return in an annual rhythm, which a short matrix will not reveal. Small cohorts are statistically unreliable, and below roughly 100 customers no conclusion should be drawn. And a change in assortment can make a cohort look permanently different from its predecessors without any change in customer behaviour.

What the report is used for

Three applications are especially useful. First, evaluating campaigns: does an acquisition push bring customers who stay, or only first orders? Second, evaluating assortment decisions: does a new category retain better than the old one? Third, planning follow-up activity, because the curve shows at which point a second purchase is most likely.

For implementation, order data with a customer identifier and a date is enough. A table with first purchase month, order month and the number of customers per cell is sufficient to compute the matrix. It matters that the identifier stays stable. Guest orders are otherwise counted as new customers every time, and the curve loses its meaning.