Re-ask the whole portfolio monthly and score the answers
Three clean sessions per prompt, with the model version logged: one answer a month is variance, not movement. Then score position, recommendation, accuracy and the sources behind it.
Step 3.4.6
Time: 2h
Done when
Report how many of the twenty prompts named the product and how many CITED it with a link, across the engines, for this month's run. Named and cited are different numbers and both are needed — one is training data bleeding through, the other is the retrieval layer finding a page. (Proof: a count.)