Main idea

To find out which words customers really use, you count the occurrences of each business term across several hundred sets of interview notes. The first result announces a two-letter acronym appearing several thousand times. In reality there were two: the search was counting every occurrence of those two letters inside other words. Search for "CA" that way and you harvest "cadre", "cas", "scanner", "application".

So you have to count whole words, and account for capitalisation when the term is an acronym. A wrong counter is worse than no counter at all, because it looks like data: it gets displayed, it gets quoted in meetings, and nothing in its form flags it as false.

The reverse mistake is sneakier. A term believed absent from the product was in fact there, used in a completely different sense, in a technical corner unrelated to the business. Verify an absence before concluding that something is missing.

Why it matters

These figures don't stay descriptive: the frequencies of a customer vocabulary steer the roadmap, the documentation and the sales pitch. An overcounted word displaces a priority; a word believed absent opens a design project that had no reason to exist.

And the countermeasure is cheap — look at the raw occurrences before looking at the total. It is the check you skip because the total, unlike the occurrences, is already ready to be presented.

Nuances and limits

Rigour in counting doesn't make the corpus representative: what is never said in an interview is never counted, however good the measurement.

And two corpora aren't counted together: interviews say what customers want to build, tickets say what breaks. The same word can dominate one and be absent from the other.

Open questions

• On a polysemous term, at what volume do you have to disambiguate by hand rather than accept an approximate count?