Customers

Everything on this page is restricted to customers with an id, which is not the same population as the one revenue is measured over.

Do they come back

Retention starts near 20% in month one, holds through month three at 21.5%, and then drifts down to about 15% by month ten. There is no cliff. Shops either turn into accounts or they never order twice, and the ones that stay keep ordering at a fairly steady rate — which is what a wholesale book looks like and not what a consumer app's decay curve looks like.

One cohort behaves nothing like the rest. December 2009 reads 35.0%, 42.5% and 37.6% at months one, three and six, against 15–25% for everything after it.

That is not a better cohort. It is the shop's existing account base appearing for the first time on the day the extract begins — customers who had been ordering for years get labelled "new" because the file starts there. It is the same left truncation that produces the customers who only ever appear as returns, and it is the reason the first row of the table below is a floor rather than a benchmark.

Blanks in the last rows are not zeros. A cohort that first bought three months before the file ends cannot have a month-six number yet, so months_observable travels with every row and the chart above drops any cohort with fewer than six months. Plotting an unobservable 0% beside a real one would say they mean the same thing.

Where the money actually is

Concentration is severe even by wholesale standards: a quarter of the customer base produces about seven pounds in ten. The practical reading is that account management and reactivation are aimed at different rows of this table, and "At risk" — few customers, high spend, long silence — is the row worth a phone call.

Two caveats the table cannot show. Recency is measured against the last day in the data rather than today; scored against the real current date, every customer in a 2011 file is equally stale and all five quintiles collapse into one. And the 5×5 cutoffs are the usual convention, not something derived from this business — a starting point for a conversation with whoever owns the accounts, not a finding.