Perplexity Citation Verification: 34.7% of Sentences with Numbers Have Source Discrepancies

Verification of 1,826 numerical sentence citations generated by Perplexity's search models revealed that 34.7% pointed to pages that did not contain the cited figures or were inaccessible to general users. When evaluated at the claim level, the failure rate was 14.4%.

The verification was conducted using 310 factual questions about 210 technology companies, comparing Perplexity's sonar and sonar-pro models against GPT as a control. Direct collection and accessibility testing of cited URLs categorized the primary causes of failure as follows:

• Access Restrictions: 16.1% of cited URLs were inaccessible to general users due to login requirements, paywalls, 403 errors, etc.
• Content Mismatch: Pages were accessible but contained none of the numbers (amounts, percentages, years, etc.) present in the cited sentence.
• Link Errors: Actual dead links (404, etc.) accounted for only 1.3% of the total.

Notably, one-third of answers from Perplexity's sonar model cited at least one inaccessible URL, a factor significantly undermining the reliability of provenance verification. Some dead links were identified as deleted directory pages from sites such as komo.ai and apollo.io.