A product recommendation depends on the question. A retailer requirement, a budget or a compatibility constraint can change which alternatives are relevant. These are different test conditions, not interchangeable prompts.
I treat prompt design as part of measurement design. Keep a baseline, document the conditions and change one important constraint at a time when investigating a difference.
Repetition matters too. A single answer is an observation. Repeated answers help describe consistency under the recorded conditions. They still do not represent every real user or every future answer.
This perspective also informs our research at Hyper Agency. Our Swedish election-compass study used repeated API calls to examine answer consistency. It is a study of that specific setting, not a benchmark for shopping outcomes.