First, identify the answer mode
Check whether web search is visibly enabled and whether the response includes citations or a Sources panel. ChatGPT Search, Claude web search, Perplexity and Microsoft Copilot can all expose source links in supported modes. Gemini may show sources or related links, but Google notes that not every response includes them.
Then test each material claim
- Open the cited page rather than trusting the citation label.
- Confirm that the page actually supports the nearby claim.
- Check publication date, author, primary evidence and geographic scope.
- Look for omitted conditions, outdated prices or mismatched products.
- Use a primary source for company, product, legal and technical facts whenever possible.
- Record unsupported claims separately from missing citations.
Do not infer hidden causation
If a directory is cited, you can report that it appeared in the observed answer. You cannot conclude from one answer that the directory is a permanent ranking factor or that creating a listing will cause future mentions. Repeat observations and use cautious language such as “appeared in sampled citations” or “may be worth validating.”
Platform-specific cautions
- ChatGPT: OpenAI says models can still make mistakes and recommends opening cited sources for important decisions.
- Gemini: Google says related links may not be the exact sources used to generate every statement, and the double-check feature finds similar or conflicting web content.
- Perplexity: citations are central to the product, but each cited page still needs to be read and assessed.
- Copilot: Microsoft warns that responses and third-party web content may be incomplete or inaccurate even when citations are present.
A reliable audit trail
Save the prompt, answer, product, date, locale, search mode and source URLs. Mark interpretations as interpretations. This makes the report reviewable and protects clients from overconfident conclusions.