AI hallucination a confident answer with nothing true behind it.

An AI hallucination is a confident answer from an AI model that is false, or that no source supports. Prefer shows what each AI engine says about your brand and which pages it cites, so you can catch a hallucination before your buyers do. The term covers anything from a wrong price to a product feature, founder or customer that does not exist.

Why models hallucinate#

A large language model writes by predicting likely next words. It does not look facts up unless the system around it retrieves sources first. When the model has little or conflicting information, it still produces fluent text, and fluent text can be wrong.

Researchers have studied this for years. A widely cited survey of hallucination in natural language generation (Ji et al.) describes models that are “prone to hallucinate unintended text” across tasks from summaries to question answering. A 2025 paper by OpenAI researchers, Why Language Models Hallucinate (Kalai et al.), argues that training and evaluation reward guessing: a model that guesses when unsure scores better on most tests than one that says “I don’t know.”

How retrieval reduces it#

The main defense is grounding: retrieving real sources at query time and writing the answer from them. The original retrieval-augmented generation paper (Lewis et al., 2020) found that retrieval models produced “more specific, diverse and factual language” than a model working from memory alone. Google says the same of its own tooling: its Grounding with Google Search documentation states that grounding helps “reduce model hallucinations by basing responses on real-world information.”

Retrieval does not remove errors. It moves them. If the retrieved page is wrong, satirical or thin, the answer can be wrong too. In May 2024 Google explained odd AI Overviews by pointing to misread queries, satirical and forum content, and “data voids,” where little good content exists on a topic.

Why it matters for AI visibility#

For a brand, a hallucination is a visibility problem with your name on it. A buyer asks an assistant about you, gets a confident wrong answer, and has no reason to doubt it. Common forms:

• Identity errors. The model confuses you with a company that has a similar name.
• Invented facts. Features, integrations or customers you do not have.
• Wrong numbers. A price or plan that was never real.
• Missing context. A competitor described in detail while you get one vague line.

Most of these trace to the same cause: there is too little clear, crawlable, consistent text about you, so the model fills the gap. That is a data void, and it is fixable.

Example#

We saw this on our own brand. On our day-zero baseline, ChatGPT answered “What is Prefer?” with fabricated details. After we published plain facts on our site, added structured data and corrected outside listings, the same prompt returned an accurate answer cited to our own domain by 2 September 2026. The full story is in how to correct what ChatGPT says about your brand.

Common confusions#

• Hallucination vs outdated answer. An outdated answer repeats something once true, usually from an old page. A hallucination has no true source at all. The fix differs: update the old page, or create the missing one.
• Hallucination vs bias. A model that prefers a competitor is not hallucinating if what it says is true. That is a share of voice problem, not an accuracy problem.
• Hallucination vs a bad citation. Sometimes the answer is right but the cited page does not say it. That is still worth flagging, because the claim has no real support.

How to reduce hallucinations about your brand#

1. Ask the engines your buyers’ questions and read the answers, not just whether you are named.
2. Check the sources behind each answer. A wrong fact with a cited page means fix that page. A wrong fact with no source means you need a clear page the model can find.
3. State key facts plainly on server-rendered pages: what you do, who it is for, what it costs.
4. Be a consistent entity across your site, directories and profiles, so the model has one story to repeat. See knowledge graph entity.
5. Re-check weekly, because answers change between runs.

How Prefer helps#

Prefer runs your buyer prompts on ChatGPT, Gemini, Perplexity, Google AI Overviews and AI Mode, and shows each answer with the sources behind it. That makes a hallucination about you visible, and tells you which page to fix. Start with the free AI Visibility Checker to see what AI says about you today.