Study finds 34.7% of Perplexity's numerical citations are inaccurate or inaccessible

Citation Accuracy Concerns

A recent analysis of Perplexity's search models found that a substantial portion of their numerical citations are unreliable. Out of 1,826 citations attached to sentences stating a figure, 34.7% pointed to pages that were either inaccessible to an ordinary reader or did not contain any of the numbers cited in the corresponding sentence.

Claim-Based Failure Rate

When evaluating claims rather than individual citations, 14.4% of 872 claims containing a figure failed. This metric considers a claim valid if at least one of its associated citations accurately supports the figure. The study focused on citations because each marker represents a specific assertion of provenance for a given sentence.

Nature of the Failures

The primary cause of these failures was not dead links, which accounted for only 1.3% of cited URLs. Instead, the issues stemmed from pages that were inaccessible (gated content) or pages that opened but did not contain the specific numerical information they were cited for. This suggests a problem with how the AI models verify and attribute information.

Methodology of the Study

Researchers posed 310 factual questions about 210 technology companies to Perplexity's two search models (perplexity/sonar and perplexity/sonar-pro), as well as GPT-4.1 with a web plugin for control. Questions covered common factual queries such as founding dates, funding rounds, headcount, and revenue. Every unique cited URL was fetched and classified to determine its status and content relevance.

Implications for AI Search

Perplexity's models provide inline citation markers, allowing for direct auditing of information sources. The findings highlight a critical challenge for AI-powered search and summarization tools: ensuring the accuracy and verifiability of cited information. Such inaccuracies can undermine user trust and the utility of AI systems for factual retrieval.