When Perplexity Hallucinates: Pattern-Spotting And Recovery

Retrieval reduces hallucination, but it doesn't eliminate it. The model can still misread a source, attribute the wrong claim to the right URL, or glue together passages from different pages into a single fluent paragraph that no individual page actually says. Perplexity's failure modes are subtler than pure-LLM hallucinations because the citations make them look authoritative.

The five failure patterns

1. 1Phantom citation: the URL exists but the claim is not on the page
2. 2Summary drift: the summary subtly overstates what the source actually says
3. 3Source soup: a single sentence cites three sources, only one of which actually contains the claim
4. 4Outdated as current: a 2022 article cited as if its claims still hold in 2026
5. 5Confident no-such-thing: a fabricated entity, person, or paper cited with what looks like a real URL

Recovery moves when you spot a hallucination

• Open the cited URL; if the claim isn't there, ask Perplexity 'find me a source that actually says X'
• If the second answer also fails, the claim probably isn't true; pivot the question
• Search the exact quoted phrase in Google with quotes; check if it appears anywhere
• Switch focus mode (Academic instead of All) and re-run
• If pattern recurs in the thread, start a fresh thread — context can poison subsequent answers

Build a hallucination journal

Keep a running list of hallucinations you've caught — what the prompt was, what the false claim was, what the real answer was. After 20 entries, patterns emerge: certain topics, certain question shapes, certain time windows fail more than others. Knowing your own failure surface beats trusting any benchmark.

The big idea: cited models still lie. Knowing the specific patterns Perplexity hallucinates against is the verification skill — and it does not transfer cleanly from how you check ChatGPT.