Why Models State Wrong Facts Confidently — The Source‑Conflict Mechanism

TL;DR

How source conflict works

An AI assistant does not have a single database of facts. It retrieves multiple sources that may disagree. Your website says one price; an old directory says another; a review mentions a third. The assistant ranks these sources and synthesises an answer from the top‑ranked one. If the outdated directory outranks your current site, the assistant will state the wrong price with confidence.

This is not a bug in the model — it is how retrieval‑augmented generation works. The model has no ground truth beyond the sources it sees. The source it trusts most becomes the answer.

Why your correct information loses

• Freshness signals missing — your site has no publication date, no lastmod, no structured data indicating the information is current.
• Directory authority — a third‑party directory that has not been updated in years may still rank higher because it has more backlinks or more citations across the web.
• Entity ambiguity — the assistant treats “Clinic A” and “Clinic A (old name)” as the same entity, conflating the two.
• No structured data — without schema.org markup, the assistant must infer your services, prices, and location from unstructured text, which is error‑prone.

What the audit does

The audit identifies every factual error the assistants state about your practice and traces it to the source that won the conflict. For each error it delivers:

• The wrong fact the assistant gave
• The source the assistant likely used
• The correct fact from your current site
• The fix: update the source, add structured data, correct the entity, or disambiguate

Why “hallucination” is the wrong label

Calling this a hallucination implies the model made something up. In reality, it reported a fact from a real source — just the wrong one. The remedy is not to ask the model to “stop hallucinating” but to make the correct source the one the model ranks highest.