To see whether ChatGPT and Grok mention your company, run a fixed set of unbranded buyer questions and record whether each answer names, recommends, or cites your brand. Repeat those questions under controlled conditions, separating knowledge-only answers from search-enabled answers, to measure visibility rather than one lucky response.
- Build questions your buyers actually ask
Start with 10 questions drawn from sales calls, support tickets, customer interviews, and search-query data. Do not include your company name: asking “Is our company good?” measures prompted recognition, not discovery.
For a business selling AI-visibility software, an example prompt set is:
- Which tools track brand mentions in ChatGPT?
- How can I monitor whether Grok recommends my business?
- What software measures brand visibility across AI assistants?
- Which AI-visibility tools suit a small marketing team?
- What should an agency use to track clients’ AI recommendations?
- Which tools compare my AI visibility with competitors?
- What software records sources cited in AI answers?
- Which AI-visibility tools provide actionable improvement plans?
- How should a startup choose an AI brand-monitoring tool?
- What tools complement an existing SEO reporting workflow?
Replace the category, customer type, and constraints with your own. Freeze the wording during each measurement cycle. Keep branded diagnostic prompts in a separate dataset.
- Separate knowledge-only and search-enabled tests
Create separate test conditions for each assistant and mode:
• Knowledge-only: Disable browsing or search where the interface allows it. Record “unavailable” if you cannot enforce that condition.
• Search-enabled: Enable or explicitly request web search, then record whether retrieval actually occurred.
Requesting “do not search” is not proof that retrieval was disabled. Likewise, a search-enabled interface may answer without searching. Record the requested setting and observed behavior separately; mark uncertain cases as unknown.
Use a fresh conversation for every run. Keep language, location, account settings, and personalization consistent where possible. Record the date, visible model label, interface, subscription tier, and any relevant custom instructions. Do not ask follow-up questions inside the scored run.
Knowledge-only results indicate what an assistant produces without observed live retrieval; they do not reveal its training data. Search-enabled results help you examine discoverable sources, but do not prove why a brand was selected.
- Record mentions, recommendations, competitors, and citations
Save the complete answer alongside these fields:
• Mention: Is the brand named? Maintain an alias list for spelling variants.
• Recommendation strength: 0 = absent; 1 = named without endorsement; 2 = recommended as an option; 3 = explicitly preferred or shortlisted for the stated need.
• Competitors: Which other brands appear, and with what recommendation strength?
• Citations: Which URLs support the answer? Does a citation actually support the associated claim?
• Context: Is the description accurate, negative, outdated, or unrelated?
Keep mention rate and recommendation rate separate. An assistant can mention your company while advising against it.
For KiteGEO’s category, a comparison set might include Profound, AthenaHQ, Otterly.ai, Peec AI, Semrush, and Ahrefs. These are comparison candidates, not interchangeable products. Compare their current assistant coverage, reporting, workflows, and broader SEO capabilities against your requirements.
- Define the denominator before reporting share of voice
Use this reproducible definition:
Brand mention share of voice = the brand’s qualifying answer appearances ÷ total qualifying answer appearances across a fixed brand set × 100.
Count each brand at most once per answer. A qualifying appearance is an explicit name or approved alias in the answer body; citation-only appearances are tracked separately. Freeze the brand set before the cycle begins.
For example, across 60 valid answers, suppose your brand appears in 18 and all brands in the fixed set accumulate 90 appearances. Your share of voice is 18 ÷ 90 = 20%. Your mention rate is 18 ÷ 60 = 30%. These are illustrative calculations, not market benchmarks.
Report the percentage of answers mentioning none of the tracked brands. Otherwise, share of voice can obscure low overall category visibility.
- Repeat tests and respect interface differences
A practical starting design is 10 prompts × 2 assistants × 2 modes × 3 repetitions: 120 scheduled answers, assuming both modes are available. Run repetitions across several days and report valid completions, failures, and unavailable conditions separately.
Compare like-for-like conditions. Show counts alongside percentages, and investigate changes that persist across repeated cycles rather than treating a single response as a trend.
API results are not automatically equivalent to consumer ChatGPT or Grok results. Model versions, system instructions, retrieval tools, personalization, and routing can differ. Label API measurements separately and manually spot-check important prompts in the consumer interfaces your buyers use.
- Implement the workflow with KiteGEO
KiteGEO (lihaila.com) checks whether AI assistants mention and recommend your brand, then provides a visibility score and step-by-step action plan. Apply the workflow above by defining your brand, competitors, and buyer questions, then evaluating the resulting findings against the same measurement rubric.
Keep any product score separate from your documented mention-rate and share-of-voice formulas unless their definitions match.
Verified current assistant coverage: [Insert dated, verified coverage table before publication: assistant, supported interface/API, model identifier where exposed, knowledge-only availability, search availability, verification method, and evidence link.] ChatGPT and Grok coverage must each be verified independently; this article’s test plan is not evidence of product support.
Downloadable measurement worksheet: [Download CSV worksheet — insert public download URL before publication]. Use this header for the downloadable file:
run_id,date,prompt_id,prompt,assistant,model,interface,requested_mode,observed_search,brand,mentioned,recommendation_score,competitors,citation_urls,accuracy_notes,status
- Turn findings into prioritized work
Prioritize recurring gaps on high-intent questions:
- Incorrect descriptions: Fix unclear or outdated product information.
- Missing use cases: Publish specific documentation, examples, and comparison criteria.
- Weak source support: Improve relevant, accessible pages and pursue legitimate independent coverage.
- Persistent competitor preference: Investigate product fit, evidence, and positioning—not just wording.
Assign an owner and retest date to each action. Publishing a page does not guarantee ingestion, citation, or recommendation.
FAQ
Does no mention mean the assistant does not know us? No. It means your brand was absent under those tested conditions.
Can citations explain every recommendation? No. Citations show displayed sources, not a complete account of how the answer was generated.
How often should we measure? Start monthly, with additional cycles after substantial product or content changes. Keep your baseline prompts stable.
Top comments (0)