Stage 2: How should you check whether ChatGPT or another assistant recommends your company?

Use a small, disclosed test before purchasing broad monitoring. Repetition matters: a March 2026 research preprint found substantial citation variability across repeated observations on three platforms and three consumer-product topics. That supports caution about single observations; it does not establish a universal sample size for B2B software recommendations. (arxiv.org)

Start with buying questions, not flattering prompts

Use the following as templates, not claims about measured search demand:

Separate branded comparison and accuracy prompts from unbranded discovery prompts. Do not combine them into a headline score without showing the components.

What is a practical starting sample?

For an initial diagnostic, we recommend:

• Six unbranded buying prompts.
• Two clearly identified surfaces, selected from customer evidence or explicitly labelled as assumptions.
• Three observations per prompt and surface, collected in fresh conversations across separate collection dates.

That produces 36 planned observations. It is a screening exercise, not a representative survey of US buyers.

If the concern merits a pilot, expand to a proposed 24-prompt core set. At two surfaces and three observations per prompt, that is 144 planned observations per measurement wave. Expand further only when additional coverage or precision would change the decision.

A fresh conversation alone is not a complete control: OpenAI documents that memory can affect search-query rewriting and that location can influence search results. Record these settings rather than assuming identical prompts mean identical conditions. (help.openai.com)

Collection checklist

• Save exact prompt wording and its provenance: customer-derived, sales-derived or inferred.
• Record the product surface, access route, date, time, language and geographic settings.
• Record available model, account, memory and personalisation settings; mark unknowns explicitly.
• Label API observations as API observations. Do not present them as consumer-interface measurements without a separate equivalence assessment.
• Store the full answer, visible citations and collection status.
• Use a written rule for aliases, subsidiaries and similarly named companies.
• Human-review ambiguous recommendations and important factual claims.
• Keep the core prompt set fixed; place exploratory prompts in a separate series.

What should the scorecard measure?

Use these operational definitions for the pilot:

Show counts alongside percentages. “Recommended in 4 of 18 collected answers” is more interpretable than an unexplained visibility score.

Also show prompt coverage: how many fully tested prompts produced at least one recommendation. Do not allow that measure to conceal instability across repetitions.

How should missing results be handled?

Our recommended treatment is:

• Collection failure: unknown, not brand absence.
• Successful search with no AI Overview: a recorded feature-availability outcome, not a collection failure. Report it separately from inclusion conditional on an Overview appearing.
• Valid answer with no recommendation: a genuine zero for the recommendation measure.
• Inaccessible cited page: record the citation, but mark its contents unverified.

Stage output: an evidence-linked AI search gap analysis, not just a dashboard screenshot.

Move forward when: the gap concerns a validated buying requirement, recurs sufficiently to justify investigation, and has evidence that someone can act on. If the result remains mixed, report that uncertainty rather than manufacturing a pass/fail threshold.