To find out what ChatGPT or another AI answer system says about your company, run the same decision-relevant prompts under recorded conditions, save the exact responses, separate factual accuracy from brand presence, inspect any exposed sources, and repeat the test over time. One answer is an observation, not a stable ranking or proof of cause.
Start with a question set tied to real decisions
Use prompts a buyer, partner, candidate or analyst might genuinely ask. Keep the wording stable so later runs are comparable.
- Identity: What does [Company] do, and who is it for?
- Category: Which companies help [audience] solve [specific problem]?
- Comparison: How does [Company] differ from [named alternatives] for [use case]?
- Evidence: What sources support the description of [Company]?
- Freshness: What are [Company]'s current products, leadership and public business terms?
Do not rewrite a prompt just to force a mention. That produces a different test rather than evidence that visibility changed.
Use the FACTS audit
The FACTS framework turns each response into a reviewable record:
| Check | What to record |
|---|---|
| F — Framing | How the answer categorizes the company, product and audience. |
| A — Accuracy | Correct, incorrect, incomplete or outdated statements, with the owned source used to verify each fact. |
| C — Citations | Every exposed source and whether it actually supports the nearby statement. |
| T — Test context | Platform, date, account or region when relevant, prompt text, search or browsing state and conversation history. |
| S — Stability | Whether the observation repeats across clean sessions and later collection dates. |
This is a measurement framework, not a score invented from one screenshot.
Keep an evidence log
For every run, store the exact prompt and response, the collection timestamp, the visible product or model label, whether web search was active, exposed citations, and the reviewer. Anthropic's current documentation, for example, says Claude web-search responses include citations and also lists availability and connectivity limits. Those conditions matter when interpreting a result. See Anthropic's web-search guidance.
A simple log can use one row per prompt and collection date. Add columns for mentioned, factual errors, outdated facts, cited domains, comparison set and reviewer notes. Preserve the raw evidence separately instead of compressing everything into a single number.
Separate presence, accuracy and source coverage
- Presence: Was the company named for the prompt?
- Accuracy: Did the description match the current owned facts?
- Framing: Did the answer place the company in the right category and use case?
- Source coverage: Were relevant owned or independent sources exposed?
A company can be mentioned inaccurately, absent from one answer but present in another, or described correctly without a visible citation. Treat those as different findings.
Correct the public evidence, not the model directly
When an answer is wrong, first identify the authoritative public source for the fact. Correct inconsistent product names, descriptions, leadership details, pricing or policy information on owned pages. Add a clear update date where freshness matters. Then review important independent profiles or articles and request factual corrections through their normal editorial process.
Do not claim that schema, a crawler directive or a content edit will force an answer system to mention the company. Google states that its established SEO practices remain relevant for its AI features, that there are no special additional requirements, and that eligibility does not guarantee crawling, indexing or serving. See Google Search Central's AI-feature guidance.
Manual and product-assisted paths
A manual audit works for a small, fixed prompt set: collect responses in clean sessions, apply the same FACTS review, and compare the dated records. For recurring work, KNWN Visibility can organize observed mentions, prompts, citations, competitors, topic gaps and changes where those product features are available.
- Use the measurement guide to interpret observations and limitations.
- Use competitor monitoring when the same decision prompt needs a consistent comparison set.
- Use citation analysis to review exposed sources separately from the answer text.
Recheck after 7 and 28 days
Repeat the unchanged prompt set after seven days to catch immediate movement and after 28 days to look for a more durable pattern. Annotate any website release, source correction or product change between runs. A changed answer after an edit is correlation; it does not by itself prove that the edit caused the change.
Audit worksheet
- Choose five to ten decision-relevant prompts.
- Record the collection conditions before testing.
- Save exact responses and citations.
- Apply the FACTS review.
- Assign each inaccurate statement to an authoritative source owner.
- Publish only verified corrections.
- Repeat at 7 and 28 days with the original prompts.
The result is a defensible account of what was observed, what was wrong, which public evidence changed and whether later observations moved. It is not a guarantee of future inclusion, sentiment or recommendation.
Interpretation note
This article is educational material, not a product commitment or a guarantee of rankings, citations, traffic or commercial outcomes.