Written by JD Davenport, who still sells for a living. Updated 4 November 2026.
A figure computed correctly on the wrong population looks exactly like a figure computed correctly. Nothing about it looks wrong. That is the whole problem, and it is why wrong-denominator findings cost more than arithmetic errors ever do.
An arithmetic error announces itself eventually. Somebody adds the column by hand, or the total stops matching the bank. A wrong population never announces anything. It produces a clean number, in the right format, on the usual schedule, and people build decisions on top of it for months.
The unmatched window. A spread of 145 jobs between the busiest and the quietest salesperson, used to argue that work was being handed out unevenly. The real spread was fifty-four. The 145 compared nine months of one person against four and a half months of another. That is not a comparison. It is a start-date table.
The filter that guarantees its own answer. A finding that the primary supplier accounted for 98.5% of material. The population had first been filtered to jobs where that supplier was at least 70% of material. The filter excluded exactly the jobs that would have held the spend the analysis existed to find. The headline could not have come out any other way.
The selected cohort. A 7.57% material return rate, circulating in the business and used to set a company-wide target. The company-wide figure was 5.67%. The 7.57% covered only jobs that had filed a credit, and was inflated further by one job that returned 147% of what it bought. A target set on an unchecked denominator generates a miss nobody can verify. That one was about a third too high, and everybody was going to be measured against it.
The export window. A large gap between material booked to jobs and material bought from the supplier, published twice, once at six figures and once at five. Both were phantoms. The supplier export began on a fixed date rather than at the start of the year, so any job with a long lag between sale and production had bought its material before the window opened.
Any export that starts on a fixed date manufactures a phantom gap on every record whose activity predates it. Long-lag jobs are exactly the records that predate it. Long-lag jobs are also exactly the ones you most want to look at.
The trap selects for your most interesting jobs. It is not random noise. It systematically produces false findings about the records that matter most, which is the opposite of what an error is supposed to do.
Re-denominate before you diagnose. Convert to the unit that makes comparison possible before drawing any conclusion. Per job. Per dollar. Per rep. Per matched window.
Most disagreements about a number are disagreements about its denominator, and neither side knows it. One person is holding a rate per job and the other is holding a rate per dollar, and they argue about the business for an hour when they should have spent a minute agreeing what they were counting.
Before any figure leaves your hands, answer three questions about it in writing. What population was this computed on. What window does it cover. What was filtered out to get there.
If you cannot answer all three, you do not have a number yet. You have a number-shaped object.
The work of checking a denominator takes about ten minutes. The work of unwinding a target that was a third too high takes a quarter, and you spend it arguing with people who were measured against it.
A fuller version of this sits inside [how to run a profit and loss for each job](/faq/profit-and-loss-per-job/), which is where most denominator errors get introduced in the first place.
Adapted from The Audit, "The denominator". More in the books. Other notes on the notes page.