By Samuel Anan

A question set is useful when it represents decisions the business actually cares about and can be repeated without changing its meaning. It is not useful merely because it produces a large spreadsheet of answers.

I would build a small, documented set before expanding the volume. That makes it easier to inspect the responses and understand what a change might mean.

Separate discovery from branded accuracy

Use unbranded questions to explore whether a business appears in responses to a relevant customer need. Use branded questions to check whether the business is described accurately.

Do not combine the two categories into one visibility percentage. A question containing the company's name gives the answer a different starting point.

The distinction between mentions and recommendations helps define what to record for each category.

Choose customer situations, not flattering prompts

Start with the service, customer type, location and constraints that make the offer relevant. A useful question might concern a provider suitable for a particular task, not simply a request to praise the brand.

Use language a customer could plausibly use. Avoid loading the question with the company's preferred claims or details that make it the only possible answer.

Keep several different decisions when appropriate: initial discovery, comparison and a question about a specific requirement. Label them so later reporting does not assume they are equivalent.

Record the intended interpretation

For each question, write why it belongs in the set and what a relevant answer would need to consider. This does not mean prescribing the winner.

An illustrative service-area question might require an answer to distinguish remote delivery from an in-person visit. If the response overlooks that distinction, the error matters even when the brand appears.

The business-facts audit supplies the approved information needed to assess accuracy. Without it, the scoring can drift towards whatever sounds favourable.

Keep the method stable enough to compare

Record the service, date, available mode and relevant context. Decide the number of observations in advance and retain all of them.

Where a question must change because the offer or customer situation has changed, create a new version and explain the reason. Do not quietly rewrite it and continue the old trend line.

OpenAI's search guidance and Google's AI features guidance describe their respective experiences. They should not be treated as evidence that every AI service produces directly comparable outputs.

Use a simple observation sheet

Keep the exact question, full response or saved record, visible sources, brand appearance, factual accuracy and any suitability claim. Add a note for uncertainty rather than forcing every response into a clean pass or fail.

A short set reviewed carefully can reveal more than hundreds of answers reduced to an unexplained score. The important work is inspecting the claims and the sources.

As explained in documenting an AI answer, the record describes observed responses. It is not a universal ranking report.

Review the set without chasing the result

Periodically check whether the questions still reflect the business and its customers. Remove obsolete situations and add new ones with a clear version change.

Do not keep only the questions where the brand performs well. A useful audit includes the difficult questions that expose a missing explanation or an inaccurate public source.

The set should help the business make better decisions about its public information. Its value comes from relevance and consistency, not from engineering a favourable answer.