GEO / AI search

Hallucination (LLM Hallucination)

Q: What is an LLM hallucination?

A hallucination is a factually incorrect statement from an LLM that sounds grammatically correct and confidently phrased. The cause: LLMs generate text on the basis of probability, not through truth verification. They ”invent" facts that sound plausible.

Q: How often do hallucinations occur?

With modern models, 3–15 % per factual query. This depends on the model, the quality of the prompt and the availability of context. RAG systems reduce hallucinations considerably, because the LLM quotes from retrieved documents rather than from training memory.

Q: How do you recognise hallucinations?

Through fact-checking. Always verify LLM answers that contain specific figures, names or quotations. Red flags: invented studies, non-existent people, incorrect dates. In coding: non-existent library functions.

Also known as: Hallucination, LLM hallucination, AI hallucination

A hallucination occurs when a large language model produces factually incorrect statements with high confidence — invented statistics, misattributed quotations, non-existent sources, incorrect historical dates. Hallucinations are not a bug but a consequence of how LLMs fundamentally work: models generate statistically probable text, not verified facts. They are unavoidable, but can be reduced considerably through retrieval augmented generation (RAG), source linking and user awareness.

Typical hallucination patterns

Invented sources — names of studies that do not exist, with author names that appear ”reliable".
Incorrect figures — statistics that sound plausible but come from no real source.
Incorrect date/name attributions — different people conflated, wrong years of birth.
Plausible-sounding but incorrect code snippets — function calls with invented parameters.
Incorrect domain/tool attributions — features ascribed to brand A that only brand B has.
Excessive self-assurance — the answer appears definitive even though the evidence is uncertain.

What reduces hallucinations

RAG with authoritative sources — when the model answers from linked sources rather than from its training memory, hallucinations are rarer.
Schema.org markup on sources — AI engines can parse facts better and cite them correctly.
Clear statements with evidence on your own website — prevents the model from interpreting ”creatively".
Current data with date markup — prevents blending with outdated information.
Wikidata entity linking — an unambiguous identity reduces hallucinations caused by confusion.
User education — users should check sources, particularly on critical topics.

Hallucination risks for brands

Concrete brand risks: (1) Incorrect feature attribution — the AI claims a competitor has a feature that only your own brand offers. (2) Invented prices or terms in AI answers. (3) Incorrect brand history or founding dates. (4) Confusion with similar-sounding brands. Strategy: maintain your own brand fact page with clear Schema.org markup, keep the Wikidata entity correct, and link sources for your own claims — this gives AI engines the right ”fact sheet" and reduces hallucination risks.

Example from practice

Example: A SaaS brand found that ChatGPT answers about tool features regularly claimed a feature the competition did not have — the ChatGPT user was frustrated because the facts were wrong. Research showed that a Wikipedia disambiguation page had conflated the brand name with a similar-sounding brand, and the LLM had mixed the two. Solution: the Wikidata entry was disambiguated unambiguously, Schema.org sameAs was added with clear URL markup, and the Wikipedia page was made more precise. After 8 weeks: hallucinations effectively ended, and tool feature answers in ChatGPT consistently correct.

Frequently asked questions

What is an LLM hallucination?

A hallucination is a factually incorrect statement from an LLM that sounds grammatically correct and confidently phrased. The cause: LLMs generate text on the basis of probability, not through truth verification. They ”invent" facts that sound plausible.

How often do hallucinations occur?

With modern models, 3–15 % per factual query. This depends on the model, the quality of the prompt and the availability of context. RAG systems reduce hallucinations considerably, because the LLM quotes from retrieved documents rather than from training memory.

How do you recognise hallucinations?

Through fact-checking. Always verify LLM answers that contain specific figures, names or quotations. Red flags: invented studies, non-existent people, incorrect dates. In coding: non-existent library functions.

How do you minimise hallucinations?

Through RAG, specific context prompts and specialised models. For critical applications: only models with high factual accuracy (Claude Opus, GPT-4o) plus RAG. Content Studio uses RAG with a curated library for each project.

Used in Rankmio for

Citation accuracy tracking in Rankmio

Go to the feature →

Last updated: 2026-06-17 · Browse all glossary entries