Citations and grounding

Grounding means generating an answer from specific source material, such as retrieved documents, a knowledge base or the user's own files, rather than from the model's general training alone. Citations are the visible half of grounding: links, footnotes or highlighted passages that show which source supports which claim. Together they let people check an answer instead of taking it on trust.

Grounding reduces hallucination but doesn't eliminate it. A model can still misread a source, combine two sources wrongly, or attach a citation that doesn't say what the sentence claims.

Why it matters

Citations change the user's job from "do I believe this?" to "does the source say this?", which is a far easier question. NN/g recommends presenting source materials so users are encouraged to verify and fact-check generated content (Laubheimer, 2025). Microsoft's Guidelines for Human-AI Interaction ask designers to "Make clear why the system did what it did" (G11), and showing sources is one of the most direct ways to do that.

Citations are only useful if they are accurate. A 2023 audit of four generative search engines found that "only 51.5% of generated sentences are fully supported by citations" and "only 74.5% of citations support their associated sentence" (Liu et al.). A citation that looks authoritative but doesn't support the claim can increase misplaced trust rather than reduce it.

How to apply it

• Do attach citations at the claim level, not as a list at the end. A research assistant should mark each sentence with the source it came from.
• Do make sources one click away and, where possible, deep-link or highlight the exact passage.
• Do show when an answer is not grounded, for example "No matching documents found, this answer is from general knowledge".
• Do evaluate citation accuracy as a product metric, not just answer quality.
• Don't cite sources the user cannot open, or sources that were retrieved but not actually used.
• Don't pad answers with many citations to look thorough. Fewer, precise sources are easier to check.
• Don't hide the source type. A forum post, an internal wiki page and a peer-reviewed paper deserve different levels of trust.

Common mistakes

• Treating the presence of citations as proof of accuracy. If people don't click through, the citation's presence alone lends credibility it hasn't earned.
• Citing a whole document when the supporting text is one paragraph on page 40.
• Mixing grounded and ungrounded content without telling the user which is which.
• Forgetting data transparency. PAIR's Explainability + Trust chapter recommends explaining what data the system uses, which for grounded answers includes which collections were searched.