Study finds Perplexity’s citations often fail to back up numerical claims

A study by consulting firm Haus Research examined the references provided by AI-powered search engine Perplexity in response to 310 factual questions about 210 tech companies. Researchers found that in over a third of cases, the provided link did not support the information presented.

Evaluation methodology

The analysis used both Perplexity models, sonar and sonar-pro, accessed through an external interface outside the app. Questions were asked about objective data—such as year of foundation, investment round value, number of employees, plan price, headquarters, revenue, security incidents, current CEO, acquisitions, and availability level—and then all cited sources were visited to verify whether the content matched the stated number.

Key findings

Of the total of 1,826 citations linked to numerical claims, 34.7% led to pages that either did not load or did not contain the mentioned value. When the analysis considered each claim rather than each isolated citation, the failure rate dropped to 14.4% among 872 statements evaluated.

• Only 1.3% of links were actually broken.
• 16.1% of URLs required authentication, were protected by a paywall, or were blocked for robots.
• The remainder corresponded to accessible pages that simply did not contain the cited data—such as in the case of Vercel’s basic plan price, where the official page listed plans but omitted the $20 value mentioned by the AI.

Information the system struggles most to verify

Questions about a company’s current CEO had the lowest confirmation rate, at 44.3%. Data on the headquarters address was also problematic, with a 53% success rate. On the other hand, information about the number of employees was confirmed in 82% of cases, and the year of foundation in 75%.

Researchers attribute the discrepancy to the fact that employee numbers are usually found in structured fields on dedicated pages, while data such as the exact date of an executive’s appointment is rarely recorded explicitly.

Difference between versions and most commonly used sources

The study found no significant difference between sonar and sonar-pro: the former confirmed 65.9% of numerical citations, while the latter reached 64.7%, a variation within the margin of error.

The most cited sources were the companies’ own websites (23.4%) and business directories such as PitchBook, ZoomInfo, and Crunchbase (23.1%). These directory pages proved to be the least durable: only 66% of them remained accessible during verification, and nearly 40% had never been archived by the Internet Archive.

Limitations of the research

The authors caution that the work represents a snapshot of a single day (September 2), includes only English questions, and uses a technical access method different from Perplexity’s standard app. In addition, the study did not measure whether the answers themselves were correct, but only whether the cited sources served as evidence for the claims.

Implications for users

The promise of AI search engines to provide verifiable references has not yet materialized in practice. Although the technology can locate relevant information, the association with inadequate sources can generate false confidence. Users should continue to treat answers provided by tools like Perplexity as starting points rather than definitive proof.