A company doesn’t need to do anything wrong for an AI assistant to invent a fact about it — stale training data, a same-named competitor, or a retrieval miss is enough, and courts have already started making brands pay for what the chatbot said.
“Hallucination” sounds like a model problem you can benchmark away. For brand facts specifically, it’s closer to a plumbing problem: three separate failure points, each with a different fix, each showing up differently across ChatGPT, Perplexity, and Gemini. Below is what causes brand-specific hallucinations, how often they happen in measured studies, what it has already cost real companies, and which mitigations actually move the number.
Three distinct ways a brand fact goes wrong
Parametric staleness. A model’s training data freezes at a cutoff date. Ask about a rebrand, a pricing change, or a discontinued product that happened after that cutoff, and the model answers from memory — confidently, and wrong. A June 2026 arXiv study on knowledge conflicts and hallucination correlation found that outdated parametric knowledge and hallucination activation patterns are related but not identical, meaning stale training data is one driver among several, not the whole story. This is the same failure mode our hallucination benchmarks breakdown calls “closed-book” hallucination — the model has no source document to check itself against.
Entity conflation. The model resolves your brand name to the wrong node — a similarly named company, a discontinued product line, or a person who shares your name. We cover the mechanics of this in our piece on entity resolution: Google’s Knowledge Graph, Wikidata, and an LLM’s pretraining each anchor identity differently, and when they disagree, the model picks one arbitrarily.
Retrieval failure. When an assistant does search the live web before answering, it can still pull the wrong page, an outdated cached version, or a low-authority source that contradicts your official one. This is a grounding failure, not a memory failure, and it’s measurable: retrieval-augmented generation dropped hallucination rates from 68% to 10% on a structured-output benchmark in a widely cited NAACL industry-track paper — a large effect, but proof that grounding alone doesn’t get you to zero.
How often it actually happens
The most rigorous public measurement of brand-adjacent hallucination comes from journalism, not marketing. Columbia’s Tow Center for Digital Journalism ran 1,600 queries across eight AI search engines — ChatGPT Search, Perplexity, Perplexity Pro, Gemini, DeepSeek, Grok-2, Grok-3, and Copilot — each time giving the engine a direct quote and asking it to identify the source. The engines got it wrong more than 60% of the time overall, per the Columbia Journalism Review write-up of the study (March 2025). Performance varied enormously by engine: Perplexity was wrong 37% of the time, the best of the eight; ChatGPT Search was wrong 67% of the time; Grok-3 was wrong 94% of the time and pointed users to dead links more often than live ones.
That study measured citation accuracy for news content, not brand facts specifically — but the mechanism is identical to what happens when someone asks an assistant to describe a company: the model is reaching for a source, and reaching wrong, at a rate that varies by 2.5x depending purely on which assistant answered the question.
What it has already cost companies
The clearest precedent remains Moffatt v. Air Canada: in February 2024, the British Columbia Civil Resolution Tribunal held Air Canada liable for negligent misrepresentation after its website chatbot invented a bereavement-fare policy that didn’t exist. The tribunal’s reasoning was blunt — Air Canada was responsible for everything on its website, whether a static page or a chatbot wrote it — and it awarded the customer CA$650.88 in damages. The case has become the standard citation for “you own what your chatbot says,” and 2026 has added company-specific rather than travel-specific sequels: on May 12, 2026, Germany’s Higher Regional Court of Hamm ruled a medical company liable for its own chatbot’s erroneous statements about a physician’s professional qualifications, a decision covered by the Library of Congress Global Legal Monitor.
Scale-wise, legal researcher Damien Charlotin maintains a public database tracking court decisions where hallucinated AI material became part of a case; it stood at roughly 1,490 decisions as of May 2026 and had grown toward 1,600 by June 2026, according to reporting that cites the tracker. The trend line, not any single number, is the finding: courts are seeing more of these cases every month, not fewer.
What actually reduces brand hallucination
| Failure mode | What it looks like | Primary mitigation |
|---|---|---|
| Parametric staleness | Model states a pre-cutoff fact (old pricing, discontinued product) as current | Force retrieval/grounding rather than relying on memorized answers |
| Entity conflation | Facts from a same-named company or product attached to your brand | Structured Organization markup with disambiguating IDs (see entity resolution) |
| Retrieval failure | Assistant cites an outdated cache, a low-authority page, or a third-party summary over your own | Machine-readable canonical source pages, kept crawlable (see schema for AI and llms.txt) |
None of these is a silver bullet, and the research backs that caution up: the RAG paper cited above cut a specific structured-output hallucination rate from 68% to 10% — a large improvement, not a cure. Grounding a model in retrieved content reduces reliance on stale memory, but the model can still retrieve and trust the wrong document, which is why entity disambiguation and retrieval quality have to be solved together rather than treated as substitutes for each other.
Practically, that means three things belong on any brand’s AI-visibility checklist: a canonical, frequently updated page stating current facts (pricing, HQ, leadership, product status) in plain, extractable text; Organization structured data carrying disambiguating identifiers rather than just a logo and URL; and periodic re-testing — asking ChatGPT, Perplexity, and Gemini the same factual questions on a schedule, because the Tow Center data shows engines don’t fail the same way or at the same rate, so a fix verified on one doesn’t confirm the others are fixed too.
FAQ
Is a brand hallucination the same thing as a benchmark hallucination?
Not exactly. Grounded-summarization benchmarks like Vectara’s (covered in our benchmarks breakdown) measure whether a model stays faithful to a document it was explicitly given. Brand hallucinations usually happen with no document in hand at all, or with the wrong document retrieved, which is a related but distinct failure surface.
Can a company be held legally liable for what a chatbot says about it?
Yes, and the precedent is no longer hypothetical. Moffatt v. Air Canada (2024) and Germany’s Hamm court ruling (May 2026) both held companies responsible for their own chatbots’ fabricated statements, and a public tracker maintained by Damien Charlotin puts total AI-hallucination-related court decisions in the thousands as of mid-2026.
Does adding structured data guarantee an AI engine gets my brand facts right?
No. Structured data helps disambiguate which entity a model is talking about, and grounding in retrieved content helps override stale memory, but the Columbia Tow Center study found error rates as high as 94% on some engines even when a specific, quotable source was handed directly to the model — so retrieval quality on the engine’s side matters as much as what a brand publishes.
For the broader mechanics of how assistants decide what to say about a brand at all, see our hub page on how AI assistants decide which brands to recommend.
Last updated August 15, 2026. This page is refreshed as benchmarks and scores move.