Start with genuine buying questions
Use enquiries, sales conversations and approved research to identify what people ask before choosing a supplier. Remove unnecessary personal information from the examples. Include the service, situation and constraints that matter to the buyer rather than adding your brand name automatically. A fictional maintenance business might examine questions about finding a provider for a small commercial property. That reflects a buying task. A prompt instructing the assistant to recommend that same business measures compliance with the prompt, not independent discovery.
Cover different stages deliberately
Separate questions about understanding a service, comparing options and choosing a provider. Keep branded questions in their own group if you want to inspect what an assistant says about a known business. Do not combine their results with unbranded discovery and present one unexplained score. Include realistic constraints such as service area or job type only where relevant. Record why each category matters to the business. A broad library should cover the intended decision space without becoming a long collection of nearly identical wording variants.
Preserve exact wording and identifiers
Give each prompt a stable identifier and save the exact text. Record its category, intended buyer need and any assumed context. Small wording changes should remain visible rather than silently replacing the earlier test. If the question includes a date or location, document why. Do not use hidden follow-up instructions to steer an unfavourable answer into a mention and then retain only the last response. The record should show what was actually asked so another reviewer can interpret the result.
Separate exploration from measurement
Exploratory prompts can help discover missing topics or confusing language, but they should not silently enter the fixed comparison set after producing a favourable result. Keep them in a separate group with their purpose stated. When a new question belongs in ongoing measurement, add it through a documented revision and establish its own starting point. Retain older results under the version of the library used. This prevents a changed question mix from appearing to be improved visibility when the underlying tests are no longer comparable.
Agree result categories before testing
Define what you will record: brand mentioned, recommendation made, source cited, unsupported statement or no relevant answer. A citation to your website and a name appearing without a link are different observations. A visit attributed to an assistant in analytics is another separate measure and does not establish what answer a person saw. Keep these distinctions in the review form. Avoid treating any appearance of a business name as proof that it was recommended positively or that the answer matched the buyer need.
Review relevance without chasing favourable results
Ask someone familiar with the buyers to review the library before running the baseline. Remove irrelevant questions with a recorded reason, not because the answers are disappointing. Keep the chosen product settings and test procedure alongside the prompts. Revisit the library when the business offering or target buyer changes, while preserving the earlier comparison boundary. Report which questions were tested and what they represent. The library supports a defined observation exercise, not a claim to measure every possible answer or predict future recommendations.