PTN—23
Citations
Attach retrievable sources to specific claims so people can verify the answer instead of trusting the vibe.
A fluent paragraph with no sources is an assertion; the same paragraph with inline citations is an argument. Citations turn AI output from something you believe into something you can check. That distinction is the whole game for retrieval-backed products: Perplexity built its identity on numbered, per-sentence citations, and it reads as trustworthy before you have clicked a single one — the presence of the affordance does most of the work.
The craft is in granularity. A pile of links at the bottom of a response is bibliography theater: nobody can tell which link supports which claim. Good citations are claim-level — a marker sits on the sentence it supports, and activating it shows the relevant passage, not just a URL. The person should get from “is this true?” to the supporting text in one interaction.
Citations also discipline the model-facing side of the system. When the interface commits to showing sources per claim, the retrieval pipeline has to actually ground each claim, and unsupported sentences become visible as what they are: uncited. Design the UI so an uncited factual claim looks slightly naked. That pressure improves the product more than any accuracy disclaimer.
The pattern emerged as the antidote to the hallucination era: once fabricated citations and invented studies made headlines, “where did this come from?” became the first question every serious buyer asked, and retrieval-backed products answered it in the interface itself. Maturity is measured behind the UI. Teams that treat citations as a rendering feature ship bibliography theater; teams that treat them as a contract run grounding audits — sampling claims and checking that the cited passage actually supports the sentence it decorates. Perplexity’s lesson is that the affordance sells the product, but only the audit keeps the promise it makes.
Anatomy
Assistant · answer
Invoice #1042 for Harrison Heating & Air is 45 days past due, and this customer has paid each of their last three invoices more than two weeks late. A firm reminder with a payment link is likely warranted.
Hover or tap a marker to inspect its source
A response with claim-level citation markers, each resolving to a source preview.
- 1Inline marker. A numbered superscript anchored to the specific claim it supports, not the paragraph.
- 2Source preview. Hover or tap reveals title, domain, date, and the exact quoted passage — verification without leaving the flow.
- 3Quoted span. The passage from the source that grounds the claim, so people compare text to text, not text to a homepage.
- 4Source list. A deduplicated footer of all sources, each linking back to the claims it supports.
When to use
- The response makes factual claims sourced from documents, the web, or a knowledge base.
- The domain punishes errors — medical, legal, financial, or anything a person will act on.
- You are summarizing or synthesizing across sources and provenance matters.
- Enterprise search or RAG products, where “which document said this” is the real question.
Design considerations
- 01Cite at the claim level, not the response level — one marker per supported statement.
- 02Show the quoted passage in the preview, not just the page title; a URL is not evidence.
- 03Include the source date and surface staleness — a 2019 pricing page cited today is a wrong answer with a footnote.
- 04Deduplicate repeated sources and keep markers compact; twelve superscripts per sentence is noise, not rigor.
- 05Make uncited factual claims visually distinguishable so gaps in grounding are legible.
- 06Handle dead or paywalled sources: cache the quoted passage so verification survives link rot.
- 07Keep markers out of the copy path — pasted text should not carry citation artifacts unless the person asks for them.
Pitfalls
- ✕Bibliography theater: a wall of links at the bottom that maps to nothing, giving the look of rigor without the function.
- ✕Citations that don’t support the claim — the model cites a real page that says something adjacent. Spot-check this in QA; it is the most corrosive failure because it punishes exactly the people who verify.
- ✕Fabricated references in ungrounded generation. If there was no retrieval, do not render anything that looks like a citation.
- ✕Over-citing trivia. Citing “Paris is the capital of France” trains people that markers are decoration and dulls their attention for claims that need checking.
In the wild
- Perplexity
- Numbered inline citations on nearly every sentence, with source cards showing favicon, title, and snippet — the citation system is the brand.
- Claude
- When answering from searched or attached sources, marks grounded spans with citations that reveal the underlying passage.
- Gemini
- Offers a “double-check” affordance that colors statements by whether Google Search found corroborating or conflicting results, linking out per claim.
- Notion AI
- Q&A answers over a workspace cite the specific pages the answer was drawn from, so “where did this come from” is one click.
- GitHub Copilot
- Can flag completions matching public code and show the referenced repositories and licenses — provenance applied to code.