GEO glossary

AI hallucination

AI hallucination is when an AI assistant like ChatGPT or Gemini states something false or invented as if it were fact, for example a wrong price, a made-up statistic or a source that does not exist. OpenAI defines hallucinations as plausible but false statements generated by language models.

Reviewed by Gareth Hoyle, Managing Director
Last reviewed 31 October 2026

Why do AI models hallucinate?

A language model generates likely words, not a checked answer. In a paper published on 5 September 2025, OpenAI gave two reasons. Pretraining has no true or false labels, so rare, arbitrary facts such as a person's birthday cannot be learned from patterns. And most evaluations score accuracy alone, which rewards a model for guessing rather than saying it does not know.

Search reduces the problem but does not remove it. An assistant that retrieves pages can still misread one, mix up two similarly named companies or rely on an out-of-date source.

How hallucinations affect brands

The errors that hurt a business are specific: old prices, discontinued products described as current, invented features, a former address, or a competitor's details attached to the brand name. OpenAI's point about rare facts applies here. A brand mentioned in only a few places gives a model little consistent information to learn from.

What to do when an assistant gets your brand wrong

Start by finding where the wrong fact comes from. If the answer cites sources, the error often traces back to one page. Feedback reports go to the platform and are not a guaranteed correction, so fixing the sources the assistant reads matters more.

Record the prompt, platform, date and the wrong statement
Check the cited pages and ask the owner of an inaccurate one to correct it
Publish the correct facts plainly and with a date on your own site, with Organization structured data
Make sure profiles, directories and your knowledge panel say the same thing
Use the platform's feedback tools, such as Report a problem on Google AI Overviews
Re-test the same prompts each month to see whether the answer changes

Example

An assistant tells users that a software company's cheapest plan costs £49 a month, a price it dropped a year ago. The cited source is an old comparison article. The company dates its own pricing page, asks the comparison site to refresh its table and tracks the prompt until the answer changes.