79,340 promotions, each charged for the discount it gave and credited with the margin the extra units earned. The average one returned 80 cents on the dollar.
Most retailers know what their promotions cost. Almost none can say what those promotions bought. The discount is visible in the margin line the week it runs. The extra units are mixed in with everything else. So I billed each promotion for its own discount and the quiet week that followed it, then credited it with the margin on every extra unit it sold, one promotion at a time.
Not a category average and not a modelled uplift. Each promotion measured against the same product's own quiet weeks, then charged and credited on its own terms.
Shallow cuts earn. Past 15% the return goes negative. At 30% and deeper the business loses 35 cents on every discounted dollar, and 85% of those events lose money.
Every other money figure in this study can be argued about through the cost file. This one cannot.
A promotion that sold more units than a normal week but took less money through the till. Whatever that product costs to buy, it lost gross margin.
11.1% of all 79,340 promotions fail exactly that test. On the deepest cuts it is one in six.
It is arithmetic, not a model. No cost assumption is involved, so it survives every objection the cost file can raise.
It is the first thing anyone says to a study like this, and it is fair: the discount has been counted and the rest of the trolley has not. The file carries 5.2 million transaction lines across 3.3 million baskets, so it can be answered rather than argued about.
Per $1 of discount, before any credit for the rest of the basket.
The same basket carrying more besides the promoted item. Still negative.
Every companion in every extra basket, treated as a trip that would not otherwise have happened.
Stated here rather than left to be found.
Products without enough promotions to read are left out rather than estimated on a handful of weeks.
Stockpiling is charged for the week after. If it borrows from further out than that, this study understates the cost.
The public file carries none, so a cost model stands in and is labelled everywhere money appears. The split between winners and losers does not depend on the cost level, and the one-in-nine finding needs no costs at all.
Every money figure here rests on an assumed cost, and a client can argue with all of them. The one-in-nine finding does not: more units, less money, therefore less margin, whatever the cost. One such test is worth more in a meeting than ten that depend on a model.
"Promotions bring people in" would have sunk this study in the first ten minutes. Measuring it from 3.3 million baskets, and crediting the promotion with the entire rest of the trolley rather than a modelled share, turns the objection into a number the client can weigh.
Whether an extra purchase occasion is a new trip is the client's judgement, not mine. So the study says what share it would take to break even and stops. That is the line between analysis and overreach.
Calendars argue about which products and which weeks. The variable that actually decides the outcome is how deep the cut goes, and it is usually set by habit.
Start with the question, not a proposal. If your export cannot support the study you had in mind, that is worth knowing in twenty minutes.