Hallucination (AI)

An AI hallucination is a false statement that a language model presents as fact — invented, misattributed, or outdated — delivered with exactly the same confidence as its accurate answers. For site owners this stopped being an academic curiosity the moment assistants started answering questions about your brand, your prices, and your content to users who will never click through to verify.

Why hallucination is a site-owner problem

When ChatGPT, Gemini, or Perplexity answers a question about your business, you have become part of the supply chain of that answer whether you opted in or not. A model can state the wrong price for your product, describe a feature you removed two years ago, or "summarize" your article into a claim it never made — and the user experiences all of it as information from you. There is no comments section to correct, no SERP snippet to fix. The only lever you hold is what the systems can retrieve and how unambiguous it is.

Grounded systems that fetch live pages before answering — the RAG pattern — hallucinate less about you than pure model memory does, which is precisely why retrievability is a defensive play, not just a traffic play.

What it looks like in practice

A pattern every consultant who monitors AI answers has seen: an assistant is asked about a mid-size company and returns opening hours that were never published anywhere, a founding year off by a decade, and a citation to /pricing-2023/ — a URL that has never existed on that domain. Each piece is plausible. None of it is checkable by the user in the flow of the conversation. Server logs even show these phantom URLs getting real hits, because users and other bots follow the fabricated citations.

Hallucination types, how they hit your site, and the defense

How to check it

1. Build a brand question battery. 15–20 questions users actually ask about you: pricing, comparisons, "is X legit", how-tos tied to your product.
2. Run it monthly through ChatGPT, Gemini, Perplexity, and Google's AI answers. Log every answer.
3. Score against ground truth. Mark each claim right, wrong, or stale. Track the error rate over time — it moves when your content and retrievability change.
4. Verify the citations. Click every URL the assistants attribute to you. Note fabricated paths and misattributed claims separately; they need different fixes.
5. Grep your 404 logs for plausible-but-fake paths arriving with AI-assistant referrers or after spikes in bot traffic. That is hallucination showing up in your server data.

Common mistakes and fixes

• Trying to correct the model's memory. You cannot patch training data, and support tickets to AI labs about a wrong answer go nowhere. Fix: change what retrieval finds — that is the writable layer.
• Leaving money facts ambiguous on-page. "Contact us for pricing" guarantees the model guesses. Fix: publish the facts you want repeated, in plain text, near the top.
• Blocking AI crawlers, then complaining about wrong answers. If grounded systems cannot fetch your pages, they answer from stale training data — the worst of both worlds. Fix: decide the tradeoff deliberately; blocking is a real choice, but it has this cost.
• No monitoring at all. Most brands have never once asked an assistant about themselves. Fix: the battery above costs an hour a month.
• Reaching for lawyers before infrastructure. Legal escalation over a hallucinated claim is slow and usually moot once retrieval improves. Fix: exhaust the retrievability and markup levers first; they act in days, not quarters.

FAQ

Can I stop a model from hallucinating about my brand entirely?

No. You can shrink the surface. The more explicit, current, and retrievable your facts are, the less gap-filling the model does — and grounded answers with an AI citation back to you displace free-floating memory. Zero hallucination is not on the menu; low and correctable is.

Why do assistants cite pages on my site that don't exist?

Because URLs are just text to a language model: it has seen your domain's URL patterns and completes a plausible one, the same way it completes a plausible sentence. Treat recurring phantom paths as free intent data and redirect them somewhere useful.

Does schema markup prevent hallucination?

Prevent, no — models do not consult your JSON-LD at generation time. But structured data makes your facts easier for retrieval systems and knowledge graphs to extract correctly, which shrinks the ambiguity the model would otherwise fill. It is one defensive layer, not a shield.

Are hallucinations getting better or worse?

Both, depending on the surface. Raw LLM memory still fabricates freely, but the surfaces users actually touch are increasingly retrieval-backed, and grounding measurably cuts error rates on factual queries. The practical trend: what the systems can fetch about you matters more every quarter, and what they memorized about you matters less.