AI citationsdefinition and how it works in 2026
- AI citations
- AI outputs that include verifiable links or references to source documents — the trust primitive that separates research-grade AI from pure generative chat.
Citations attach source provenance to AI outputs. A cited answer says "the deflection rate is 50–80% (Sierra 2026 case study)" with a clickable link, not just "the deflection rate is 50–80%". Citations are how RAG-based products earn user trust — and how they're differentiable from raw chat models.
Quality of citations varies wildly. The honest hierarchy in 2026: (1) Real, working URLs that load the cited content — Perplexity, Claude Research, OpenAI Deep Research at their best. (2) Document IDs in your private corpus — most enterprise RAG. (3) Hallucinated citations that don't resolve — the worst-case bug that still happens occasionally and costs vendors deals.
Good citation UX surfaces them inline (clickable footnotes), shows the exact passage on hover, and explicitly distinguishes "from my training" answers from "from this retrieved document" answers. The best products go further — they highlight which sentences in the response are grounded vs inferred.
Frequently asked
Why do citations matter beyond looking nice?+
They make AI outputs verifiable. Without citations, users can't check claims, and trust collapses on the first wrong answer they spot. With citations, even wrong answers are recoverable — you click through and see what went wrong.
Can general LLMs cite reliably?+
No. General chat LLMs (without retrieval) hallucinate citations. Citation quality is a property of the retrieval pipeline + the prompting, not the underlying model.
