What is hallucination?
A hallucination is an AI answer that states something false or unsupported, such as a feature, price, customer, or integration that does not exist. The wording can be fluent. Fluency is not evidence. A citation can still point at a page that does not say what the answer claims.
Why it matters for AI visibility
Hallucinations are an AI visibility issue when buyers meet a wrong version of your product, and a trust issue when they meet a wrong version of a rival you are compared with. You cannot switch them off from the outside. You can make the correct fact easy to retrieve: a clear page, consistent details on sources you influence, and a correction when a cited page is stale. Grounded answers reduce the room a model has to invent, but a retriever can select a wrong page, and a model can misread a right one. When you spot a bad answer, save the prompt, engine, date, and the exact sentence. Check whether any link supports it. If a cited page is wrong, fix that page. If there is no source, there is nothing to “update” except the public facts a future retrieval might use. Re-test the same prompt. Do not report a single corrected answer as proof the model is fixed.
Example
An assistant says your software syncs with a payroll system you do not integrate with, and links to an old blog post that speculated about a partnership. The sentence is a hallucination relative to the product, grounded in a page that should be corrected or ignored. Log both the claim and the URL.
Related terms
Grounding and retrieval-augmented generation are how answers get attached to pages, which is the main counter to invention. An AI citation lets you audit the attachment. Sentiment labeling does not catch a hallucination that sounds positive and is still false.