Main idea

The two large corpora a company holds about its customers — interview notes and support tickets — don't record the same thing. Interviews carry the projects, the intentions, the vocabulary of what customers are trying to put in place. Tickets carry the breakdowns, the blockages, the vocabulary of what doesn't work.

The same word can therefore dominate one and be absent from the other, without either of them being wrong. Counting them together amounts to adding up two measurements of different objects: you get a ranking that corresponds to no reality, and you prioritise wrongly.

Carried separately, the two frequencies each have their destination: the interview frequency steers the roadmap, the ticket frequency steers documentation and training.

Why it matters

This gives a simple rule to anyone working with customer verbatims: two corpora, two counts, two readings — and an explicit refusal to merge them, even when the tool offers to.

It also guards against a volume bias: support produces far more text than interviews, so that a global count amounts to measuring only what breaks.

Nuances and limits

The distinction isn't watertight: a customer in an interview also describes what blocks them, and a ticket sometimes carries an intention. These are dominant tendencies, not labels.

And a third corpus is missing from both: what customers have never put into words, neither as a project nor as an incident.

Open questions

• What third source would gather what neither corpus contains, without relying on the memory of those who talk to customers?