Audit finds a third of Perplexity's citations don't back the figures they cite

What the audit found

An audit published by Haus Research, which reached the front page of Hacker News, asked Perplexity's two search models 310 factual questions about 210 technology companies, then fetched every cited source and checked whether each page actually contained the number it was cited for. Out of 1,826 citations attached to sentences stating a figure, 34.7% pointed at a page that either would not open for an ordinary reader or contained none of the numbers in the sentence it supported. Scored per claim instead of per citation — a claim passes if any of its cited pages carries one of its figures — 14.4% of 872 figure-bearing claims failed.

The questions used ten templates covering facts people actually look up: founding year, latest funding round, headcount, entry-level pricing, headquarters, revenue, disclosed breaches, current CEO, acquisitions and paid-tier uptime guarantees. Both Perplexity models, sonar and sonar-pro, ran at temperature 0, with GPT-4.1 plus a web plugin as a control.

How each citation was tested

Perplexity's models attach inline markers to claims, each pointing at a specific URL, which is what makes sentence-level auditing possible at all. The auditors fetched every unique cited URL — 2,915 for sonar alone — and classified each as dead, gated, empty, unreachable or live, giving failed fetches extra attempts through longer timeouts and a rotating proxy, which rescued 192 URLs.

The pass test was deliberately lenient: extract the claim's money amounts, percentages, years and digit runs, and pass if the page's visible text contains at least one of them, with normalisation so $185 million, $185M and 185000000 all match. A single number, even a bare year, is enough. That makes the headline 34.7% a floor, since every failing pair is a page containing not one number from the sentence that cited it.

Gated pages and disposable SEO pages

Dead links are not the main problem: only 1.3% of cited URLs were dead. The larger category is gated sources. About one citation in six pointed at paywalled or login-walled sites such as PitchBook, ZoomInfo, Crunchbase and Reuters. Aggregated to the answer level, 84.2% of sonar's 310 answers cited at least one URL an ordinary reader could not open, and 10.6% cited at least one that was outright dead.

The dead links had a distinctive shape. About half of sonar's 38 dead URLs were programmatic pages generated per company and per question type — paths like komo.ai's office directories, apollo.io's where-is pages and temperstack.com's plan pages — built to rank for exactly these queries and removed as cheaply as they went up. Rechecks on the day of writing found the Elastic offices page returning 404 and apollo.io's Reddit head-office page returning 410 Gone.

Pages that open but don't contain the claim

Among pairs where the page did open and was readable, 16.1% contained none of the claim's figures. Asked for the entry price of Vercel's cheapest paid plan, sonar said the first paid tier starts at $20 per month and cited Vercel's own pricing documentation, a live page that names the plans but does not contain that price. Both models produced street addresses for headquarters questions and attributed them to Wikipedia articles that contain no street numbers or postal codes — several answers even hedged about multiple sources while attaching the Wikipedia marker anyway. Sonar named GitLab CEO Bill Staples with a 5 December 2024 start date while citing GitLab's executive page, which names Staples but carries no date.

According to the audit, failure tracks how a fact is published. Headcount and founding year live in structured fields on pages built for them, and hold up well; a CEO's start date or an office's street number is widely repeated but rarely published, so the model reproduces the consensus and points at a page that never carried it. The claim and the citation appear to come from the same generative process rather than genuine retrieval.

The two models test the same

A pilot had suggested sonar-pro grounded its claims worse than sonar, but at full scale that gap disappeared: sonar passed 65.9% of numeric pairs (95% interval 62.8–68.9) against sonar-pro's 64.7% (61.5–67.7), with both models citing roughly 9.8 sources per answer. GPT-4.1 with web access behaved differently, citing 2.0 sources per answer and leaning more on companies' own domains (36.4% versus Perplexity's 23.4%). It emits no inline markers, so sentence-level checking is impossible — itself a finding about end-of-answer bibliographies.

Sonar's citations spread across 989 distinct hosts, with 23.4% to the company's own domain and 23.1% to B2B directories and lead lists. LinkedIn was the largest single host at 5.6%, ahead of Wikipedia at 4.3%, and directory pages proved the least durable: only 66.0% opened, against 78.7% of citations overall.

Why it matters

Citations are Perplexity's core promise — the feature that separates a search-grounded answer from plausible-sounding generation. The audit suggests that promise is frequently unverifiable: a third of numeric citations either can't be opened or don't contain the cited figure, and gated footnotes offer provenance claims no reader can test. For teams building on Perplexity's sonar APIs, the practical lesson is that inline markers make grounding measurable, so citation accuracy should be tested per use case — and not assumed from the presence of a footnote.