Definition
A hallucination is an answer from a language model that sounds convincing but is factually wrong or fabricated: a non-existent product, an incorrect price, a quote that was never said, a source that does not exist. The model does not 'lie' deliberately; it predicts the most likely next words and has no built-in sense of true or false. Without a reliable source it fills the gaps with whatever sounds plausible.
For marketers this matters for two reasons. First, an AI engine can say incorrect things about your brand: outdated prices, services you do not offer, a location that has closed, or a competitor presented as your product. Second, you can publish hallucinations yourself if you put AI-generated content live without checking — a fabricated statistic or a non-existent source in an article damages your credibility with readers and with the engines that evaluate you later.
Hallucinations about your brand usually arise because the model cannot find a good source. Retrieval systems (RAG) reduce the problem considerably: if a clear, current page exists that states the facts, the engine retrieves it and uses it instead of guessing. If that page is missing, outdated, blocked for AI crawlers or unreadable without JavaScript, the model falls back on what it 'thinks' it knows from training.
The antidote is therefore publishing and maintaining unambiguous fact pages: about us, pricing, product specifications, locations, frequently asked questions, with a last-updated date and structured data. Consistency between your site and external profiles (Google Business Profile, LinkedIn, directories, review platforms) helps too: models weigh sources against each other, and contradictory information leads to errors.
Check regularly what engines say about you through prompt sampling: ask questions about your brand, prices and services and compare the answers with reality. Traze tracks which facts AI engines state about your brand, flags deviations, and runs every page the platform writes through a fact check and human review before it goes live.