Skip to content

Baseline method · published August 31, 2026

How to Run a Five-Prompt AI Search Baseline for a Consultant

Five prompts are enough to create a useful starting record when the questions stay fixed and the limits stay visible. The set is a diagnostic sample, not a universal visibility score.

By Christopher Buchner, founder of PruneQ

Why start small

A consultant can generate hundreds of questions and run them across several services. The resulting spreadsheet looks comprehensive and can exceed the time available for checking factual accuracy, sources, screenshots, and changes.

A small frozen set forces a useful decision about buyer intent. It also makes the same questions practical to repeat after a website change and again 30 to 45 days later.

If you want a faster first check before building a measurement routine, start with the simpler guide to checking what AI says about your business.

Research on generative-search measurement shows that answers and citations vary across runs and time. Five prompts do not solve that statistical problem. They provide a transparent diagnostic sample whose wording and evidence can be inspected.

Choose one branded prompt and four buyer questions

1. Branded accuracy

Who is [founder], what is [company], and what services does the business offer?

Checks whether the answer resolves the correct person and organization and uses current visible facts.

2. Problem recognition

What should a [buyer type] check when their website is unclear or difficult to find?

Shows the source set used for the problem the consultant addresses, without asking for a recommendation.

3. Service category

What does a good [service category] include for a [buyer type]?

Tests whether the consultant's method and category language appear in the answer or cited sources.

4. Evaluation

How should a [buyer type] evaluate a provider for [service]?

Maps the credentials, proof, directories, publications, and criteria used in a buyer-style comparison.

5. Recommendation context

Which providers help [buyer type] with [specific problem] in [market, if relevant]?

Observes recommendation-style results and sources. This is the noisiest prompt and should never be treated as a stable rank.

Replace the brackets with language the buyer would use. Avoid stuffing the company name into all five questions. The branded prompt checks identity. The other questions should test the problem and category without forcing the answer to mention the consultant.

Freeze the conditions before the first run

Choose one primary answer service and product surface. Record whether web search is enabled. Use the same account state and locale when practical.

Add a second service only when it produces evidence that changes the decision. More engines create more observations, more source differences, and more review work. They do not automatically create a better conclusion.

Keep the exact wording. If the business, buyer, or service changes enough to require a new prompt, preserve the old prompt and start a new series rather than rewriting prior results.

Record the answer like an audit observation

  1. Exact prompt and the reason it belongs in the set
  2. Date, time, locale, and signed-in state when known
  3. Engine, product surface, and whether search or browsing was enabled
  4. Mentioned, missing, cited, or factually confused
  5. Correct and incorrect founder, company, service, location, price, and credential facts
  6. Every cited domain and exact cited page
  7. A screenshot or saved answer
  8. Which owned or third-party source appears to support each important fact
  9. Any site, profile, or editorial change since the prior run
  10. A limitation sentence for that observation

Read the result in plain language

A useful baseline summary might say:

The branded answer identified the correct company but used an outdated founder title. Two buyer prompts cited broad marketing publications. One cited a competitor's guide. The recommendation prompt did not mention the company in this run. These five answers are dated observations and do not establish a stable rank or the cause of inclusion.

That summary produces a work list. Correct the title where it is wrong. Compare the cited guides with the owned service page. Qualify the publications and profiles that repeatedly appear. Leave the absent recommendation alone until the evidence points to a legitimate action.

Repeat without declaring causation

  • Run the same set immediately after implementation as a quality check.
  • Repeat it after 30 to 45 days for a later observation.
  • Compare factual accuracy, mentions, citations, and source types.
  • Review Search Console, referral traffic, and qualified inquiries separately.
  • Record other changes that could influence the comparison.
  • Report changed, unchanged, and variable results without assigning a universal score.

References