A hallucination occurs when an AI system such as ChatGPT, Gemini, Claude, or Perplexity confidently states something that is not true: a wrong price, a nonexistent feature, or a source that does not back up the statement. The term is a loose analogy to hallucinations in humans; it is also called confabulation. For businesses, such errors matter because they shape how people perceive an offering.
Two kinds of hallucination
Research broadly distinguishes two cases:
- Wrong or invented facts: The answer contradicts verifiable facts or makes up details, such as a location that does not exist.
- Departing from the question or source: The answer strays from the question or from the content available to the system, for example by misrepresenting a retrieved web page.
Hallucinations also occur when an AI system retrieves current content from the web and bases its answer on it (grounding), which usually lowers the risk but does not eliminate it. A 2023 audit of several such systems found that cited sources do not always support the statements they are attached to. An AI citation alone therefore does not prove that a statement is true.
How hallucinations arise
A language model does not look up facts in a stored directory; it generates text, word by word, that fits the question. When it lacks reliable knowledge, it often still produces a fluent, confident answer. A September 2025 research paper argues that language models hallucinate because their training and evaluation reward guessing over admitting uncertainty. Hallucinations are especially likely for facts about little-known entities that are rarely written about, which can particularly affect mid-sized businesses and their products.
Why hallucinations matter for brands
In an October 2025 study by the European Broadcasting Union and the BBC, journalists from public service media reviewed more than 3,000 answers from ChatGPT, Gemini, Perplexity, and Copilot to news questions. Almost half had at least one significant issue, for example with sourcing. Around a fifth had major accuracy issues, such as invented or outdated details. The figures apply only to this study but show that errors are not isolated cases.
For businesses, accuracy is therefore a distinct part of their AI visibility: an offering can have a high mention rate and still appear with wrong prices, services, or contact details. Such errors go unnoticed unless someone regularly checks what AI systems say about the business; how closely their statements match the verified facts is described by brand accuracy.
A business cannot rule out hallucinations entirely. But clear, current information that reads the same everywhere gives AI systems a correct source to draw on; this is what brand fact consistency is about. A company’s own website can also provide this information as structured data.