A third of Perplexity's citations point to pages that don't back the cited figure

Perplexity answers with footnotes. Every sentence gets a numbered [1]-style link, like an academic paper, with a citation list at the bottom of the answer. An audit of 1,826 of those footnotes, pulled from Perplexity's answers to 310 factual questions about 210 technology companies, found the linked page doesn't contain the cited figure about 34.7% of the time.

That's the headline number from Haus Research's audit of Perplexity's two search models, sonar and sonar-pro, plus a GPT-4.1 control with a web plugin. The audit tracked the inline [n] markers Perplexity attaches to its answers, fetching each cited URL and checking whether the page actually contained the number the sentence claimed it did. The 310 questions came from 10 templates: founding year, latest funding round, headcount, entry price, headquarters, revenue, disclosed breaches, current CEO, acquisitions, paid-tier uptime SLA.

The 34.7% figure is per citation. Each footnote was scored on whether the page it pointed to contained the number. Scored per claim instead, with 872 distinct figure-bearing claims across the run, 14.4% fail when any one of the cited pages carries the number. The per-claim number is closer to a user's lived experience of "how often does a number not check out," but the audit's headline result is the per-citation one, and the two numbers point at the same underlying problem: provenance mismatch, where the footnote names a real page but the page does not say what the sentence says it says.

Most of those mismatches are not dead links. Only 1.3% of cited URLs were dead. The dominant failure modes, per the audit, were gated pages that require a login or subscription, empty or unreachable fetches that would not open to an ordinary reader, and live pages that simply do not contain the cited number. The link works. The page just doesn't say it.

The audit fetched 2,915 unique URLs for sonar alone, classifying each as dead, gated, empty, unreachable, or live, and retried failed fetches up to two more times per excerpt. That scale is what makes the headline number auditable rather than anecdotal.

The control is partly missing from the available excerpt. The audit tested GPT-4.1 with a web plugin at temperature 0, but the report does not quantify the control's failure rate the way it does Perplexity's. Without that number, readers can't tell whether this is a Perplexity-specific problem or a general failure of search-augmented language models. GPT-4.1 was a control, not a co-subject.

A few caveats on the audit itself. Haus Research is a single research firm, and the methodology, classification scheme, and dataset are described in the report but not fully published as code or raw data in the excerpted version. A Hacker News thread on the report flagged a concern: commenters claim a lookalike research firm posted a sibling anti-Perplexity piece on the same day. The claim is unverified, and the audit's findings stand on their own methodology, but the audit is one firm's work, not a consensus result. Perplexity was not given a response in the hydrated material; a published response, if one comes, would change how the failure rate should be weighted.

The takeaway for a reader using Perplexity for sourced research: the footnote is a real link, often to a real page, but the page may not say what the sentence says it says. Verify the figure on the linked page before trusting the footnote, especially when the page is gated, the answer involves a specific number, and the source is a small or specialist site.