The essentials
- Definition: a hallucination is false, invented or distorted information presented with confidence by a language model: a figure, a quotation, a date, a company, a source that doesn't exist.
- The cause: the model predicts the most likely text, it doesn't verify; on a rare, recent or poorly documented subject, the plausible continuation isn't the true one.
- Frequency in 2026: from 1 to 3% of statements on common questions with web search, to more than 20% on niche subjects with no sources.
- For a business: an AI can invent a price, an address or an offer; the defence is a documented, consistent presence on the web.
A hallucinationin artificial intelligence is false, invented or distorted information that a language model presents as fact, often with confidence and in credible phrasing: a figure that doesn't exist, a quotation wrongly attributed, an invented study, a company described with the wrong address or offer. The term describes normal behaviour in LLMs, not a fault: the model produces the most likely sequence of words, and on a subject its data covers badly, the most likely sequence isn't the truth.
Why models hallucinate
- Prediction, not verification: the model doesn't consult a database of facts; it generates plausible text. A study name, a year and a percentage form a plausible sentence even with no real study behind it.
- Gaps in the data: rare subjects, local businesses, recent events, less-represented languages; where data is missing, the model fills in.
- Leading questions: a question that presupposes a fact ("what does X's premium plan cost?") pushes the model to supply one, even if the plan doesn't exist.
- Contradictory sources: when the web says two different things about a company, the model can produce a third.
- Learned confidence: models are tuned to answer fluently and confidently; admitting ignorance is less common than producing a plausible answer.
The rates observed in 2026
| Context | Rate of false statements (order of magnitude) | Factor |
|---|---|---|
| General questions, a recent model with web search | 1 to 3% | Plentiful, consistent sources |
| Summarising a document you provide | 1 to 5% | The model is reading rather than inventing |
| Figures, dates, precise references without search | 10 to 20% | Details poorly represented in training |
| Niche subjects, local businesses, current events | 20% and above | Gaps in the data, scarce sources |
| Quotations and bibliographic references | High, up to 30 to 50% without search | A highly predictable form, invented content |
The 2026 models hallucinate markedly less than those of 2023, and web search with citations ( RAG) reduces the phenomenon by grounding answers in documents. It doesn't disappear: a false or out-of-date source produces a false but sourced answer.
Examples in digital marketing
- The invented figure: "the average e-commerce conversion rate is 4.7% according to a 2025 study"; the study doesn't exist. Every figure destined for publication must be traced back to its source.
- The distorted company: an AI describes an agency with an old address, a discontinued offer or an invented price, because the web contains contradictory or outdated information.
- The imaginary feature: "GA4 lets you…" followed by a feature that doesn't exist, or no longer does; frequent with fast-moving tools.
- The attributed quotation: a sentence credited to an expert or to Google that they never said.
- The fictional competitor: a list of "best tools" mixing real products with invented names.
Protecting yourself from hallucinations in your work
- Supply the sources: give the model the document, the report or the page to answer from, rather than relying on its memory.
- Demand the references: ask for the source of every figure, then open it; a source that doesn't open is a hallucination.
- Ask open questions: "what does X offer?" rather than "what does X's premium plan cost?".
- Cross-check: two assistants, or the assistant and a classic search, before publishing anything.
- Have an expert review it: human review remains mandatory for any published content, particularly figures, names and tool features.
Protecting your company from what AI says about it
AI assistants describe companies from what they find: your site, your Google Business Profile listing, directories, press, social media. A company whose sources are complete, current and consistent gets described correctly; a company with contradictory or missing information gets filled in by the model, and therefore hallucinated. The defence is a well-established entity : a reference About page, structured data, consistent information everywhere, and regular monitoring of what assistants say about the brand, described in monitoring your brand's presence in generative AI.
How GreenRed helps
Rather than juggling several tools, GreenRed's GEO and AI module brings these metrics together in a single dashboard, compares them over time and tells you which actions come first. You can try it free, with no card, from the Pricing.
Frequently asked questions
What is an AI hallucination?
False, invented or distorted information that a language model presents as fact with confidence: a figure, a study, a quotation, a date, an address or an offer that doesn't exist. It is normal behaviour for models, which predict the most likely text without verifying it.
Why does ChatGPT invent sources?
Because a bibliographic reference has a highly predictable form (author, year, title) that the model reproduces without any real content behind it. Without web search, the rate of invented references can exceed 30%. With search and citations, the model points to real pages — which should still be checked.
How do you avoid hallucinations in content written with AI?
By supplying the sources to the model rather than relying on its memory, by demanding a reference for every figure and opening it, by asking open questions, by cross-checking against a classic search, and by having an expert review it before publication.
Can AI give false information about my company?
Yes, if the web's sources are contradictory, outdated or incomplete: the model fills in with an invention (an address, a price, an offer). The defence is a documented, consistent presence (site, Google listing, directories, structured data) and a quarterly check on what assistants say.