Measurement & risk
Hallucination
An AI stating something confidently that is wrong, unsupported or made up.
A hallucination is a confident statement from an AI model that is factually wrong, unsupported by any source, or entirely invented. It happens because models generate plausible text from patterns, not from a checked record of truth — so fluent nonsense can come out sounding authoritative.
Hallucination is relevant to brands in two ways. First, grounding and retrieval reduce it, which is why engines that cite sources are generally safer and why being a clear grounding source protects accuracy about you. Second, models can hallucinate facts about your company — wrong pricing, features or claims — and the fix is to make the correct facts unmissable and well-corroborated so the engine grounds on them instead of guessing.
Monitoring what AI engines say about you, and strengthening the authoritative sources for your key facts, is the practical defence.
Key takeaways
- A hallucination is when an LLM states something false with full confidence and no warning that it made it up.
- For your brand this is a direct risk: an engine can tell a buyer the wrong price, wrong feature, or wrong founder, and sound certain doing it.
- Strong entity presence and authoritative, consistent grounding sources are your best defense against being described wrong.
Why models confidently make things up
An LLM generates the most statistically likely next words. It isn't checking a database of truth. When the training data is thin, conflicting, or silent on a topic, the model fills the gap with something that reads plausibly, and it has no built-in signal to tell you the answer is invented. A 2025 survey describes hallucination as output that is fluent and syntactically correct but factually inaccurate or unsupported by evidence.
That's why common fixes only go so far. Turning temperature to zero doesn't eliminate it, and retrieval doesn't fully solve it either. Researchers increasingly frame hallucination as a property of how these models work, not a bug a single patch removes. For you, that means it's a standing risk to manage, not something you wait for the vendors to fix.
What hallucination costs a brand
When buyers ask ChatGPT or Perplexity about your company, the engine answers with whatever it can assemble. If your presence is weak, it guesses, and it can state your pricing wrong, attribute a competitor's feature to you, place your headquarters in the wrong city, or invent an integration you don't offer. The buyer has no reason to doubt it. A confident wrong answer about you is worse than silence, because it shapes a decision before you ever enter the conversation.
The counter is to make the correct facts the easy, consistent answer. Build a clear entity: consistent name, description, and key facts across your site, your schema, and high-authority third-party sources the engines trust, like your Wikipedia-adjacent references, G2 profile, and Crunchbase. Use grounding sources that say the same thing everywhere so the model isn't choosing between conflicting versions of you. When the authoritative, repeated signal is accurate, the model has far less room to improvise.
This is also why prompt visibility tracking matters as a safety net. You won't know an engine is describing you wrong unless you're running the prompts and reading the answers.
Example
We tested a RegTech client by asking ChatGPT and Perplexity to describe their product. ChatGPT confidently listed a SOC 2 integration they'd never built and quoted a price tier that didn't exist. Nothing in the answer hinted it was uncertain.
The fix wasn't arguing with the model. We tightened their entity: matched product facts across the site, added Organization and Product schema, and got two authoritative sources to state the same accurate details. On the next tracking run, both engines described the product correctly. The hallucination came from an information vacuum, and we filled it.
Frequently asked questions
Can I stop an AI engine from stating wrong facts about my company?
You can't force a model to be perfect, but you can starve the hallucination. Make the correct facts consistent and authoritative across your site, your schema, and trusted third-party profiles so the accurate answer is the easy one for the model to reach.
Does retrieval or RAG eliminate hallucination?
No. Grounding in retrieved sources reduces it meaningfully but doesn't remove it. The model can still misread, blend, or overstate what it retrieves. That's why consistent, high-authority sources about your brand matter so much.
How would I even know if an engine is hallucinating about my brand?
By tracking it. Run buyer prompts about your company across the major engines on a schedule and read the actual answers. Hallucinations about pricing, features, or facts show up fast once you're watching instead of assuming.
Sources
Related terms
Grounding
Tying a model's answer to fetched source documents instead of its own memory.
Grounding Source
The exact page an engine cites as the proof for a specific claim in its answer.
Fact Density
How much specific, sourced detail a passage carries per sentence, versus filler.
Entity Optimization
Making your brand resolve as one clear, consistent entity everywhere engines look.
Prompt Injection (SEO)
Hiding instructions in a page to try to steer what an AI answer says.
Want to know where you stand in AI answers?
We’ll show you which engines cite you, and which don’t.
Get a free AI citation audit