An AI hallucination is a confident but false output from a language model — an invented fact, a fabricated quote, a citation to a study that doesn't exist. The text reads fluently and authoritatively, which is exactly what makes hallucinations hard to catch.
A large language model predicts the next most-likely token, not the truth. When its training data is thin, outdated, or contradictory on a topic, it fills the gap with something statistically plausible rather than admitting uncertainty. Ask for "three peer-reviewed 2023 studies on X" and a model may generate real-sounding author names, journals, and DOIs that lead nowhere — the format is learned, the content is confabulated.
As AI answers replace some clicks, brands care about being represented accurately. Two failure modes hurt: the model states something wrong about your business (a discontinued product, a stale price), or it credits a claim to you that you never made. Retrieval-augmented generation — grounding answers in fetched source documents — sharply reduces both, which is why clear, well-structured, factual pages are now a defensive asset.
A SaaS company found ChatGPT telling users its free plan included a feature that was actually paid-only. The cause: outdated pages and a vague pricing table the model stitched together. Rewriting the pricing page with explicit, unambiguous plan tiers — and adding structured data — corrected the AI answer within weeks and cut a wave of misdirected support tickets.
Reducing the odds that AI misrepresents you is a core aim of generative engine optimization, and it leans heavily on the grounding techniques behind RAG.