Audit: Perplexity's Citation Reliability

Haus Research (Report HR-2026-09, dated September 2, 2026) has uncovered a significant discrepancy in how Perplexity AI attributes its data. Specifically, one-third of the citations provided by Perplexity's search models fail to actually contain the numerical data they are credited for.

The Core Findings

Out of 1,826 citations linked to sentences containing specific figures, $\approx 34.7\%$ were invalid. A citation was marked as a failure if the page was inaccessible to a standard user or if the page opened but lacked the figure mentioned in the text.

However, if we shift the metric from per citation to per claim (where a claim passes if at least one of its multiple citations is correct), the failure rate drops to $14.4\%$ (based on 872 claims).

"We lead with the citation because a marker is an individual claim of provenance: this sentence came from that URL."

---

🛠 Methodology: How the Audit Was Conducted

The researchers tested two search models—perplexity/sonar and perplexity/sonar-pro—alongside GPT-4.1 (with a web plugin) as a control. They posed 310 factual queries regarding 210 different technology firms at temperature 0.

The Query Framework

The team used ten specific templates to ensure consistency:

• Founding date
• Most recent funding round
• Total headcount
• Entry-level pricing
• Corporate headquarters
• Annual revenue
• Disclosed security breaches
• Current CEO
• Company acquisitions
• Paid-tier uptime SLA

The Technical Workflow

To ensure accuracy, the researchers used a normalization process for numbers. For example: $185 million $\rightarrow$ $185M $\rightarrow$ 185000000

The $34.7\%$ failure rate is actually a floor not a ceiling, because it only counts pairs where zero numbers from the sentence appeared on the page.

---

🌐 Analysis of Inaccessible Citations

The study analyzed 2,915 unique URLs cited by perplexity/sonar. The results are summarized below:

While sites like Reuters or Crunchbase have a right to charge for access, providing a gated link as a citation is a failure of the system. It creates a claim of provenance that the reader cannot verify.

The "Disposable" Web

Interestingly, many "dead" links followed a pattern. The researchers found "minted" pages designed specifically to capture search queries, which were deleted shortly after. Examples include:

• portersfiveforce.com/blogs/brief-history/[company]
• matrixbcg.com/[path]
• canvasbusinessmodel.com/[path]

Specific 404/410 Examples:

• Elastic HQ: komo.ai/directory/elastic-offices $\rightarrow$ 404
• Reddit HQ: apollo.io/where-is/reddit $\rightarrow$ 410 Gone
• Discord Pricing: temperstack.com/plans/discord $\rightarrow$ 404

---

❌ Readable Pages with Wrong Content

Even when a page opened successfully, $16.1\%$ of those citations contained none of the figures mentioned in the claim.

Example 1: Pricing Hallucinations When asked about Vercel's pricing, sonar claimed the first paid tier starts at **$20/month** and cited `vercel.com/docs/plans`. While the page is live and lists plans, the string `$20` does not appear anywhere on that page.

Example 2: The Wikipedia Address Trap A recurring pattern emerged where models provided specific street addresses and attributed them to Wikipedia, despite the information being absent from the source.

In some instances, the model even admitted that "multiple sources list" the address, yet it still attached the citation marker to the Wikipedia URL. This suggests that the claim and the citation are generated by the same flawed process, rather than the citation being a true reflection of the retrieved text.