Build a prompt set to assess your AI reputation.

A useful panel starts with a question about the business, not a collection of attractive AI answers. A founder preparing for due diligence may need to check identity, career history and current role. A professional-services firm may need to understand whether buyers encounter it when exploring a particular need in a particular market.

• Name the entity and the audience whose questions are being represented.
• State the decision: evaluating an advisor, checking a leadership appointment, comparing service providers or another concrete task.
• Choose the language, market and systems relevant to that task. Record the interface and search conditions separately.
• Define what the initial panel can and cannot establish. Avoid treating a small working panel as representative of all buyer questions.

The prompt library supplies the site’s canonical templates. This guide explains how to turn appropriate templates and documented audience questions into an operational register. The AI reputation audit guide describes the wider assessment workflow.

Assign one primary family to each prompt and record separately whether it names the target. Keep comparison questions and source-review questions out of an unprompted discovery rate. Mark recommendation eligibility before testing; a question asking for a current role is not asking the system to recommend a provider.

For a leader without a credible comparison category, allocate the work to identity, factual representation and source evidence. Inventing a league table of executives does not make the measurement more useful.

A question may come from a documented buyer conversation, an enquiry, available search-query data, or the analyst’s design. Record that provenance. An analyst-written question can test an important hypothesis, but it should not be presented as a measured popular query without supporting data.

• Keep the original question or its documented origin and explain any editing needed for a repeatable test.
• Retain meaningful distinctions in need, market and decision stage; remove accidental duplicates that would overweight the same question.
• Do not select or weight a prompt merely because an earlier run produced a favorable mention.
• If weights are used, state their rationale and retain the unweighted observations. Unsupported search-volume assumptions should not become hidden weights.

Google documents that AI Overviews and AI Mode may issue multiple related searches when developing an answer. That underlying query fan-out is different from the question the evaluator supplied. Do not count those searches as additional independent prompts in the panel; see Google’s explanation of AI features.

Context can help resolve ambiguity, but it also changes the test. Asking who a person is and asking who they are in a role at a named organization are two different conditions. The second question already supplies the role and organization.

The existing library pairs a basic identity question with a contextual identity question for this reason. Record supplied facts in a conditioning-facts field and state which claims remain eligible for assessment. Use the executive digital footprint guide to establish the reference record before checking a leader’s roles and dates.

Use fresh conversations for independent runs and document account context, prior conversation, attachments and search mode. A biography pasted earlier in a conversation is part of the test conditions, not evidence that the system knew those facts independently.

This is a hypothetical design exercise for an organization with a real service category, not a completed study. The IDs below are illustrative register IDs, not replacements for the canonical library’s template numbers. Each row shows one wording example for a group of slots; the full wording of every slot must be written before any collection begins.

For planning only: 12 prompts × 3 repetitions × 3 systems = 108 planned attempts. The numbers describe the workload of this example, not a sample size validated for every audit. Log failures, refusals and eligible answers separately rather than reporting every attempt as a usable observation.

The pilot checks whether the questions are understandable, the conditions can be recorded and the coding rules can be applied. It is not a basis for publishing a competitive market-share claim. For any public research on this site, follow the declared research methodology, including its publication requirements.

• Write the exact questions, primary families and measurement targets before the pilot.
• Check that placeholders are resolved consistently and the wording does not accidentally supply the fact being tested.
• Confirm that the question is relevant to the chosen language and market. Treat translated panels as distinct versions and review their meaning, rather than pooling them automatically.
• Have another reviewer apply the coding rules to a sample when feasible. Record unresolved classification differences instead of assuming that subjective labels are reproducible.
• Revise ambiguous or accidental duplicate questions because of their design, not because an answer omitted the target.
• Approve and archive a panel version before the baseline used for reporting. Keep pilot observations distinct from that baseline.

The baseline should keep the declared number of repetitions and the recorded conditions consistent across systems. Where a product does not expose a model version or search behavior, record that limitation. Do not fill an unknown field with an assumed value.

Download the blank prompt register to capture the questions, provenance, supplied context, recommendation eligibility, repeat count, approval and version history. It contains column headings only.

Keep each approved version available with its exact wording and scope. A change in question text, language, category, comparison set or test context can affect the answer independently of any change to the entity’s sources.

• Give the panel a version and every prompt a stable register ID. The combination of panel version and prompt ID must identify one exact question.
• Retain an ID for continuity only when its measurement purpose remains the same; record a replacement relationship when that purpose changes.
• Record the reason and date for a change. A leadership transition may require new factual context; it should not be hidden inside an apparently unchanged panel.
• Keep old and new versions separate in reporting. Where continuity matters, define a controlled overlap using unchanged anchor questions and label the limits of that comparison.
• Never reconstruct missing historical answers from current runs or merge pilot revisions into a trend as if they were the original baseline.

The AI reputation dashboard guide explains how version changes, eligibility and denominators should appear beside reported rates. The AI search intelligence guide covers the broader record of conditions and sources.

An operational handover should let another person repeat the intended test and understand what each result means. It needs more than a spreadsheet of questions.

• The approved prompt register, with exact wording, provenance, language, market and the intended measurement for every question.
• The verified entity reference record and rules for matching names, relationships and material facts.
• The collection protocol: systems and interfaces, session conditions, repetitions and treatment of unavailable responses.
• The coding rules for mentions, identity, citations, recommendations and factual checks, including supplied facts that must not be scored independently.
• A version history, approval owner and evidence-log location, with access appropriate to the information being collected.

A well-designed panel can expose gaps, errors and conflicting sources. It does not control the systems’ wording or establish what every real user sees. Its value is a repeatable, interpretable record that helps decide what to investigate and correct.

What is a prompt set for an AI reputation audit?

It is an approved, versioned collection of exact questions used to observe how AI systems describe and present a defined entity. Each question has a stated audience, intent and measurement target, with the conditions and repetitions documented so results can be interpreted and repeated.

Should the questions include the company or person’s name?

Include the name when testing recognition, identity or factual representation. Leave it out when testing discovery within a relevant need or category. Keep those question families separate, because an answer to a named question does not show that the system would introduce the entity unprompted.

How many prompts are needed?

There is no universal number that validates every assessment. Scope the panel around its decisions, coverage and collection workload, then explain its limits. A small pilot can test the process, but it does not by itself establish a representative benchmark or justify a competitive share-of-voice claim.

Can contextual facts be used to help identify a person?

Yes, when the purpose is documented identity disambiguation. Record the facts supplied in the question or conversation. A role or organization provided by the evaluator must not then be scored as if the system had recalled it independently.

Can prompts be edited while tracking a trend?

Keep the original version and document the change. New wording or context may alter the measurement, so it should receive a new panel version and be reported separately. Any comparison with the old panel needs explicit continuity rules and limits.

Are these prompts intended to make an AI system recommend the entity?

No. They are observation tools. Questions that presuppose praise, supply an expected recommendation or are repeatedly tuned until the entity appears should not be presented as neutral evidence of visibility or reputation. The objective is to record interpretable answers and supporting sources.

This guide presents a proposed working method. The pilot panel and workload are illustrative, not observations from a completed study or client engagement. Documentation and the existing prompt library were checked on 27 October 2026. Recheck their current requirements before implementation.

• Mohammed Teto — Prompt library: The canonical template catalogue, including paired identity questions and their interpretation.
• Mohammed Teto — Research methodology: Requirements for fixed prompts, recorded conditions, coding and publication.
• Google Search Central — AI features and your website: How Google AI features may use multiple related searches while developing an answer.
• Mohammed Teto — AI search intelligence: The wider discipline of recording representation, citations and test conditions.

Related guidance: Prompt library, AI reputation audit, Executive digital footprint, AI reputation dashboard.