Related: AI Communications pillar · Citation Share Is the New KPI · The AI Communications Dictionary · The Citation Share Index
ChatGPT says your CEO went to Stanford. She went to Penn. Claude says the company was founded in 2008. It was 2011. Perplexity quotes a product feature you've never shipped.
These are hallucinations. The model didn't lie. It filled a gap.
Understanding the gap is how you tend to stop the hallucination.
What hallucination actually is
Large language models compress. Given a prompt, the model retrieves the most plausible answer from its training corpus and live retrieval. When the retrieved evidence is thin, conflicting, or missing, the model often fills the gap with statistically plausible detail.
A model that has seen ten authoritative sources confirm a founding year of 2011 will generally state 2011. A model that has seen two sources say 2011 and one say 2008 will sometimes confidently state 2008. A model that has seen no clear sources at all will often invent a year that sounds plausible.
The model is not hostile. It's compressing under uncertainty.
The three common causes
Most hallucinations about brands trace back to one of three structural gaps in the citation graph.
- Thin authority stack. Not enough trusted sources have stated the fact. The model has no anchor.
- Conflicting sources. Different outlets reported different versions of the same fact. The model picks one — sometimes the wrong one.
- Stale dominance. An out-of-date source has become the model's most-cited reference. The fact was correct then. It isn't now.
Why it appears to be getting worse, not better
More engines are using live retrieval, which surfaces current articles — including current low-quality articles. SEO content farms have learned to write about brands, and that content is now feedstock. As the volume of brand-adjacent text grows, the signal-to-noise ratio tends to fall.
The implication: brands need to invest in authoritative source density not just to compete — to defend baseline accuracy.
How to reduce it
Four moves matter.
- Lock the high-stakes facts in tier-1 earned media. Founding year, leadership, headquarters, key milestones. Not just on your own site. On Reuters, Bloomberg, Forbes, The Wall Street Journal. The model tends to weight these heavily.
- Get Wikipedia right. A well-sourced Wikipedia article is often the single most powerful anti-hallucination asset for a brand.
- Issue structured owned content. A leadership team page with consistent schema, a press kit with consistent facts, a fact sheet linked from every news release. Retrieval engines tend to pull these. The schema and entity-rich content frameworks are defined in The AI Communications Dictionary.
- Monitor and surface corrections. When a model hallucinates a fact, push correction content into the citation graph. The model often updates on the next retrieval cycle. The underlying KPI framework is in Citation Share Is the New KPI.
What does not stop hallucinations
Contacting OpenAI. Contacting Anthropic. Filing a complaint. The engines don't generally run a manual review queue for brand corrections at scale. The fix is typically upstream — in the sources the model retrieves from.
Hallucinations are a measurement problem before they're a content problem. Find them first. Fix the inputs. Re-measure.
No communications firm can guarantee specific outputs inside third-party AI systems. The discipline is shaping the inputs the engines retrieve from — not directing the engines themselves.
