What is an AI hallucination?
OpenAI defines hallucinations as "plausible but false statements generated by language models" (OpenAI, 5 September 2025, read 1 November 2026). Anthropic describes text that is "factually incorrect or inconsistent with the given context" (Claude docs).
Google does not use the word on its AI Overviews help page. It does say: "AI Overviews can and will make mistakes" (Google Search Help, read 1 November 2026).
Why do AI models hallucinate?
A 2025 paper by OpenAI researchers argues that models hallucinate because training and evaluation "reward guessing over acknowledging uncertainty" (Kalai and others, arXiv). A model that guesses scores better on accuracy-only tests than a model that says "I don't know".
OpenAI's summary of the paper also points to rare facts (OpenAI). Spelling follows patterns, so spelling errors fade as models grow. A one-off fact, such as a pet's birthday, cannot be predicted from patterns alone. Your plan names, your price and the year you dropped a feature are rare facts too.
Why do hallucinations matter for B2B SaaS?
Buyers now ask assistants about software. G2's 2026 survey of 1,076 B2B decision-makers found that 51% of B2B software buyers now start their research in an AI chatbot more often than in Google.
So a wrong answer about your product reaches buyers before they visit your site. An assistant can state an old price, list a feature you removed, or mix you up with a company that has a similar name.
How do you reduce hallucinations about your brand?
- State your key facts in plain sentences on your own site: what you sell, who it is for, how pricing works and where you are based.
- Keep the same facts on your profiles and in the third-party lists that mention you.
- Date your pages and update them when a fact changes.
- Ask the main assistants the same questions each month and log every wrong answer.
- When an answer is wrong, find the page it cited and fix or update that page.
These steps make correct facts easier to find. They do not control the answer. The model provider decides what the assistant says.
If you build with AI models yourself, Anthropic suggests letting the model say "I don't know", grounding answers in direct quotes and checking each claim against a citation. Its docs add that these techniques "don't eliminate them entirely".
What does a hallucination look like? (hypothetical example)
This example is hypothetical. A buyer asks an assistant: "Does Acme CRM have a free plan?" Acme dropped its free plan in March. Three list articles from last year still say it has one. The assistant reads those lists and repeats the old fact with confidence.
The fix is not only on Acme's site. Acme adds a dated sentence about its plans to its pricing page, and asks the three publishers to update their entries.
Hallucination vs grounding vs retrieval
- A hallucination is the false answer itself.
- Grounding ties an answer to sources the model can point to. It lowers the risk but does not remove it, because a model can still misread or mix its sources.
- Retrieval-augmented generation fetches documents before the model writes, so it works from text instead of memory alone.
- A large language model predicts the next word. That is why its errors sound fluent.
For how assistants pick the sources they quote, see LLM SEO: how B2B SaaS brands get named in AI answers.