When AI lies about your brand: why, and how to fix it
By Miguel Maio ·
TL;DR
- AI invents facts about brands often: a 2026 benchmark across 37 models measures hallucinations between 15% and 52%, depending on the task.
- There are two distinct failures and almost nobody separates them: parametric (the model 'remembers' wrong, from training data) and retrieval (it reads a wrong or outdated source in real time). The fix differs for each.
- It already costs real money: Air Canada was held liable for a chatbot that gave false information (2024) and Deloitte refunded ~US$290K for a report with fabricated citations (2025).
- The defense isn't 'fix your site': it's building a source of truth AI reads (entity, schema, Wikidata, high-authority third parties) and monitoring with fixed prompts. Solid RAG with guardrails cuts hallucinations by 71-89%.
Ask ChatGPT what your company does. Then ask the price, the address, whether you hold a certification. There's a good chance one of those answers is wrong - stated with the same confidence as the correct ones. It's not bad luck: it's how models work. And unlike a Google error, there's no page to fix - the error lives inside the model or in the source it read. The good news: understand the cause and you can fix it.
The two failures nobody tells apart - and should
Almost everyone treats "AI got me wrong" as one problem. It's two, with opposite fixes:
1. Parametric hallucination. The model "remembers" your brand from its training weights - and remembers it wrong, or out of date. It isn't reading anything in real time; it's generating from what it absorbed. Fixing your site today doesn't change this immediately: it only lands with new training, once your entity signals are strong and consistent enough for the model to "relearn" the right fact.
2. Retrieval hallucination. The model fetches information in real time (search grounding, RAG) and reads a wrong, old or low-quality source - then summarizes it with confidence. Here, what's published now does count: the fix is to control which sources it finds and to make the right one the most visible and unambiguous.
Telling the two apart is the work. If AI cites a source, it's retrieval - go to the source and fix it. If it asserts without citing, it's (probably) parametric - the game is entity and consistency over the medium term. Treating everything as "I'll fix my site" is the mistake that makes people give up, convinced nothing can be done.
What it actually costs
This stopped being an image problem. In 2024, a Canadian tribunal held Air Canada liable for false information its chatbot gave about bereavement fares - the airline argued the bot was "a separate legal entity" and the tribunal refused: the brand answers for what its AI says. In 2025, Deloitte agreed to refund part of a ~AU$440K (~US$290K) contract to the Australian government after a report surfaced with fabricated academic citations and a quote attributed to a court judgment that didn't exist - all AI-generated. If it happens to a Big Four, the tolerance for your brand is zero.
Why your brand is an easy target
Models hallucinate most where they have the least signal and the most noise. A brand with a thin entity - inconsistent name, no schema, no presence in independent sources - leaves the gap the model fills on its own. Worse: if you have contradictory data scattered around (old prices in one directory, a wrong address in another, an outdated bio on LinkedIn), you're feeding the hallucination yourself. The rates help size the risk: the average fell from ~38% (2021) to ~8.2% (2026), but open-ended generation runs at 15-25% and agent chains with tool calls reach 20-40%. The more the buying decision runs through an agent, the greater your exposure.
The source-of-truth playbook
The defense is building a single source of truth AI reads and reuses, in this order:
- Unambiguous entity: a coherent Person/Organization, with
sameAsto your official profiles and, if possible, Wikidata - that's how the model links the name to the right facts. - Factual schema on the page itself: price, address, founding, certifications in valid structured data, consistent with the visible text (retrieval reads this first).
- Align the third parties: fix NAP, bios and facts in the directories, press and profiles AI cites - your site isn't enough if the model reads another source.
- Content that answers the question: an FAQ stating the fact directly and citably gives the model the correct version to extract.
- llms.txt and crawler control: make the right information easy to access and decide which AI bots may read what (see block or allow AI).
Notice the logic: points 2 to 5 attack retrieval (fast effect); point 1 attacks the parametric side (medium-term effect). You need both.
How to monitor without spending a fortune
You can't fix what you don't measure. Fix a set of 10-20 prompts about your brand (what it does, prices, competitor comparisons, "is it trustworthy?") and run them periodically across the four engines (ChatGPT, Perplexity, Gemini, AI Mode), always the same, to compare over time. AI brand-monitoring tools (Ahrefs Brand Radar, for example) automate part of this, but a disciplined spreadsheet already catches the essentials. The goal isn't vanity: it's spotting the hallucination before your customer does and having the evidence to act. That's what I do in GEO consulting, tied to the AI visibility audit.
An honest read
There's no button to "fix ChatGPT", and anyone promising one is selling smoke. Hallucinations don't disappear - they shrink, with consistent signals and continuous monitoring. Rates improve with each model generation (the best are already below 1%), but your brand doesn't control anyone's training calendar. It controls what AI reads about you. Start there.
Key data
| Hallucination range (37 models, 2026) | 15-52% |
|---|---|
| Historical average (2021 → 2026) | ~38% → ~8,2% |
| Open generation / agent chains | 15-25% / 20-40% |
| Air Canada case (chatbot, 2024) | Company held liable |
| Deloitte case (fabricated citations, 2025) | ~US$290K refund |
| Reduction with RAG + guardrails | -71 a -89% |
Rates vary widely by task and model; read ranges, not absolutes.
Sources: OECD.AI - Deloitte refunds Australia (2025) · CBC News - Air Canada chatbot liability (2024) · SQ Magazine - LLM hallucination statistics 2026
Common mistakes and how to do it right
There are no tricks against hallucinations - there's a well-built source of truth. What separates those who control the narrative:
| Avoid | Do |
|---|---|
| ✗ Ignoring what AI says about you until a customer complains | ✓ Monitoring brand answers with fixed, recurring prompts |
| ✗ Fixing only your site and waiting | ✓ Fixing at the source AI reads: entity, schema, Wikidata and third-party pages |
| ✗ Assuming AI 'made it up' and nothing can be done | ✓ Working out if it's a parametric or retrieval failure - the fix differs |
| ✗ Leaving contradictory data across the web | ✓ Aligning NAP, prices and facts everywhere (one source of truth) |
Tools I use and recommend
- ChatGPTAI assistant for research, analysis and quick audits.
- PerplexityAnswer engine with sources, useful for research and GEO.
- GeminiGoogle's AI assistant, grounded in Search.
- ClaudeAI assistant for analysis and content.
- AhrefsKeyword research, Domain Rating (authority), backlinks and Brand Radar (AI citations).
- Google Search ConsoleOrganic performance, coverage and field Core Web Vitals.
Sources
- OECD.AI - Deloitte refunds Australia after AI-generated report errors (Oct 2025)
- CBC News - Air Canada liable for chatbot's bad advice (Moffatt v. Air Canada, 2024)
- SQ Magazine - LLM hallucination statistics 2026
Frequently asked questions
Does ChatGPT really invent facts about brands?
Yes, and confidently. A 2026 benchmark across 37 models measured hallucination rates between 15% and 52%, depending on the task. For brands with little online information, the risk is higher. Audit AI visibility →
Can I force AI to correct itself?
There's no button. You fix it by influencing what AI reads (retrieval) and reinforcing your entity for the next training (parametric). It's source-of-truth work, not a takedown request. GEO guide →
What's the difference between parametric and retrieval hallucination?
Parametric comes from the model's memory (training data); retrieval comes from a source it reads in real time. If AI cites the source, it's retrieval and you fix the source; if it asserts without citing, it's parametric and you work on entity and consistency. Get cited by AI →
Does fixing my site solve it?
Only half. It solves retrieval if AI reads your site - but if it reads a directory with wrong data, or 'remembers' training wrong, your site alone isn't enough. You also have to align third-party sources. Digital PR and citations →
Who's liable if AI gives wrong information about my company?
Your company, in practice. Air Canada tried to argue the chatbot was a 'separate entity' and the tribunal refused (2024). Assume you answer for what your AI - and AI about you - says. GEO consulting →
Which data should I keep consistent?
Name, address and contact (NAP), prices, founding date, certifications and the founder's bio - identical on the site, in directories, press and social. Contradictions feed hallucinations. Entity example →
Does schema help against hallucinations?
Yes, on retrieval. Valid structured data, consistent with the text, gives AI the factual version to extract. It's no guarantee, but it narrows the room to invent. llms.txt explained →
Are monitoring tools worth paying for?
If your brand depends on reputation, yes - they save time and catch what a person can't watch. But start with a fixed set of prompts in a spreadsheet: it already catches the essentials. Audit method →