Direct answer. No. A citation without a retrievable source is not evidence, whatever the model's Truth Score. Perplexity Pro is the dataset's specialist for cited answers, with Truth Score 90, but that score does not make an unopenable footnote safe to forward. If you cannot open the source, do not pass the citation on.
An AI citation you cannot open is a claim, not evidence. This page is about that failure mode: a model prints a title, a court, a paper or a URL, and the string is not attached to a source you can retrieve. It is a different question from which model has the lowest hallucination rate, and it is a different question from which mitigation technique cuts errors in a benchmark. Best AI Match scores factual reliability as a Truth Score. It does not score a model's habit of inventing bibliography entries as a separate factor. Perplexity Pro is the specialist record aimed at cited answer generation, with a Truth Score of 90, privacy 80 and safety 80. Its dataset pricing is not an API token rate. The input and output prices are null because the product is subscription-shaped, and the context window is recorded as 0. A high Truth Score on any other model does not convert an unsourced footnote into a source. The decision rule below is operational: if the citation has no retrievable source, do not pass it on.
A citation string is not a source
Models can emit a plausible title, a volume number, a court and a URL in one sentence. The sentence can be fluent and still point at nothing. Treat the string as unverified until a person opens the document. This is the decision on this page. The ranked list of which model fabricates most lives on which AI hallucinates most, and the technique list lives on how to reduce hallucinations. Neither of those pages answers whether a printed citation is usable.
What the dataset actually records
| Record | Truth Score | Why it is here |
|---|---|---|
| Perplexity Pro | 90 | Best-fit includes cited answer generation. Token prices are null. Context window is 0. |
| Claude Sonnet 4.6 | 96 | Highest truth mark. Not a citation-retrieval product in the best-fit list. |
| GPT-5.4 | 82 | General business and writing. A high task score does not attach sources. |
| Gemini 3 Flash | 80 | Best-fit includes search augmentation, which is not the same as a checked footnote. |
Figures from /data/ai-comparison-2026.json. Last verified 2026-06-26. Scoring version v1.0. Affiliate links are off.
A decision rule for unsourced citations
- If the citation has no URL or document id you can open, remove it.
- If it has a URL, open it and confirm the quoted claim is on the page.
- If the job is research that must carry sources, start from best AI for research and still apply the open-the-source rule.
- Do not let a Truth Score of 96 stand in for that click.
The weights behind those scores are on methodology. Perplexity's official site is perplexity.ai, and the link is not an affiliate link.
Need a model for the research job? Use the match engine, then keep the citation rule even if the engine prefers a high truth score.
FAQ
Does a Truth Score prove a citation is real?
No. Truth Score measures published reliability. It is not a check that a specific footnote resolves to a document.
Which model in the dataset is aimed at cited answers?
Perplexity Pro lists cited answer generation among its best-fit jobs and has a Truth Score of 90. Its token prices are null because the record is subscription-shaped.
What should a reviewer do with an unsourced citation?
Do not repeat it. Either retrieve the source independently or delete the citation before the text is used.