Field note
AI cites sources that don't check out
In my own audit, AI answers pointed to pages that don't exist. A citation that looks real isn't proof.
When I audited what AI assistants say about a category, I did the obvious thing: I asked each one for its sources. Most handed back a tidy list of links. Some of those links went nowhere: placeholder addresses and pages that don't exist. The answer looked sourced; the sources were hollow.
That isn't just my bad luck. In 2025, researchers checked whether AI assistants' answers were actually supported by the sources they cited, and found a large share were not fully backed up, even with web search switched on. That study was in the medical domain, so treat the exact figures as a caution rather than a number for your category (Nature Communications, 2025).
Two things follow. As a reader: a citation-shaped string is not evidence. Click it before you trust it. The model sounds equally confident whether its source agrees, disagrees, or doesn't exist. As a company trying to be visible: this is the opening. The bar for being genuinely cited (by a real, checkable source that actually says what the AI claims) is lower than it looks, precisely because so much of what passes for citation today is hollow. Be the source that holds up.
See "citation" and "hallucination" in the glossary, or the overview for how this fits the method.
New here? Start with the overview → · By Denes Csaszar · written in the open