What ChatGPT, Perplexity and other assistants say about local businesses, how that is measured, and what changes it. Based on dated measurements.
An AI assistant says your company is: “award-winning”; “ISO certified”; “licensed by” a regulator; a member of an association; winner of a prize you never received. The claim sounds positive. That does not make it safe to use. An invented qualification can create more risk than an invented negative fact because someone inside the business…
A customer says: ChatGPT says you are permanently closed. The business is open. That can cost an enquiry before the customer ever reaches the website. This is more serious than an old price because the false statement questions whether the business exists at all. The useful question is: Which source or public signal is supporting…
Usually, no. Create a separate public page only when it answers a distinct customer question or consolidates useful information that is otherwise hard to find. Google says AI Overviews and AI Mode use the same core Search requirements and do not need special AI markup, schema or machine-readable AI files. OpenAI says any public website can appear in ChatGPT search and focuses on crawl access rather than an AI-only page. A page made only to target AI systems, with no separate value for people, is more likely to duplicate existing content than solve a real information problem.
No. A citation shows that an AI answer links to or uses a source; it does not by itself mean the system recommends that source’s business. OpenAI describes citations as links to sources behind search-backed answers, and Google describes AI features as surfacing supporting links. To measure recommendation, inspect the actual wording: is the business named as a provider the user should choose, merely used as evidence, or only linked for further reading? Record citation and recommendation as separate observations.
An AI assistant can quote an old price because the answer may rely on an outdated page, a stale third-party source, an older indexed version or a response generated without a sufficiently current source. First capture the exact answer and citations, then make the current price clear on the authoritative page you control, update any applicable structured data and important third-party listings, and re-test later. A corrected source improves the available evidence but does not guarantee when or whether an external AI system will change its answer.
AI-generated answers can be wrong, and source-backed systems can still misinterpret or combine information. If two businesses have similar names, first document the exact wrong attribution, the source or link shown with it, and the identifying facts that distinguish the businesses: official name, website, address, phone and other public details. Then correct inconsistent public information you control. You can improve the evidence available to AI and search systems, but you cannot guarantee that an external AI system will stop making the mistake or use your preferred source.
Sometimes an AI-assisted search answer may reflect business information or material from the web, but current official documentation does not justify saying that every assistant reads or uses a particular Google review. Google says Business Profiles can help local business information appear in its generative AI search experiences, while OpenAI says ChatGPT Search may use web search and show citations. The checkable result is a specific answer and its sources at a specific time, not a guarantee of future recommendations.
Partly yes, and not in the way it is sold. AI answers are assembled from pages a system can reach, so what is published about you elsewhere is how you exist to it, but no documentation says that a count of mentions moves an answer.
ChatGPT SEO prompts are useful for drafting, grouping supplied data and checking whether a page answers a customer’s question. They are not search-volume data, ranking instructions or a way to submit information to Google. The safest prompts provide real source material, ask for a limited output and include a human verification step before anything is published.
You will rarely prove it, and you do not need to: the tells are the ones a reader feels within a paragraph. From our practice the strongest one is that everything is too neat, because a person writes with digressions and small mistakes and a machine leaves none.