Original practice question. The scoring rubric is our teaching guide. How we research and score answers
TL;DR: Attach citations to individual claims; make each one show the supporting passage on hover or tap without leaving the page; show source type and date; distinguish what sources say from the model's own synthesis; and never display a citation the system has not verified exists and supports the claim. Measure citation accuracy alongside click-through.
How to approach it
Start with what users need to verify and how much effort they will spend, which is a few seconds. Then decide granularity, presentation, source information, and the back-end checks that keep citations honest.
A strong answer
Granularity. A list of five links at the bottom of an answer tells users little: which source supports which sentence? Citations belong at the claim level, a marker after each factual sentence or clause, pointing to the passage that supports it.
Presentation.
- Hover or tap to preview. Show the supporting passage, with the relevant words highlighted, in place. One click opens the full source at that location.
- Source cards. Title, publisher or document owner, source type (official documentation, news, forum, internal wiki), and date. Users judge credibility quickly from these.
- Multiple sources for important claims. When two independent sources agree, show both; when they disagree, say so explicitly.
Separate evidence from synthesis. Some sentences restate sources; others are the model combining or concluding. Mark synthesis clearly ("Based on these sources…") or leave it uncited rather than attaching a citation that does not directly support it.
Back-end honesty checks.
| Check | Purpose |
|---|---|
| Citation exists | The cited document and passage are real and retrieved |
| Support check | The passage supports the attached claim (validated judge plus audits) |
| Freshness | Outdated sources flagged or replaced when newer ones exist |
| Access | Users only see citations to sources they are allowed to view |
Unsupported claims are either removed, rephrased with a caveat, or left without a citation. A wrong citation is worse than none: it lends false authority.
How often citations fail in shipped products. Liu, Zhang and Liang's audit of generative search engines (Findings of EMNLP 2023) checked Bing Chat, NeevaAI, perplexity.ai and YouChat: only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported their sentence. The Tow Center's March 2025 test of eight AI search tools on 1,600 queries about news articles found incorrect answers to more than 60% collectively; 154 of Grok 3's 200 citations led to error pages, and most tools rarely hedged. In law, a Stanford preprint by Magesh et al. (May 2024) found the AI research tools from LexisNexis and Thomson Reuters hallucinated between 17% and 33% of the time despite retrieval. Retrieval alone does not make citations honest. The support check in the table above is the feature, and its audited accuracy is the number to put on the launch review.
Domain variations. In legal products, citations must be exact and to authoritative sources, with pinpoint references. In internal knowledge tools, document owner and last-updated date matter more than publisher. In health, prefer authoritative sources and show a caution when evidence is limited.
Measurement. Citation accuracy (share of citations that exist and support their claims, from audits), preview open rate (are people verifying?), complaints about wrong citations, and whether users who check citations report higher trust.
What interviewers probe next
"Users rarely click citations. Why bother?" Citations change behaviour even unclicked: they discipline the system to answer from sources, they let the cautious users verify, and audits of citation accuracy are a strong quality metric.
"What if no good source exists?" Say so, rather than citing a weak source to look authoritative.
How it's scored
| Signal | Weak | Hire | Strong Hire |
|---|---|---|---|
| Granularity | Links at the end | Per-claim citations | Per-claim citations with passage-level previews |
| Credibility signals | None | Source titles | Source type, owner, date, agreement and disagreement across sources |
| Honesty | Displays whatever the model cites | Checks links exist | Existence and support checks, synthesis distinguished, wrong citations removed, accuracy audited, with published audit rates as the benchmark to beat |
Common mistakes
Decorative citation lists that do not map to claims.
Citations that exist but do not support the sentence.
Making users leave the page to verify.
See designing for uncertainty and evaluating RAG.