Key findings
- Haus Research checked 1,826 citations that Perplexity's Sonar and Sonar Pro API models attached to sentences stating a figure. 34.7% pointed at a page that would not open or held none of the sentence's figures. Scored per claim, 14.4% of 872 claims failed.
- Gated pages caused more failures than dead links. Of Sonar's 2,915 cited URLs, 16.1% sat behind a login, paywall, or bot wall, and 1.3% were dead.
- Trellner found 59.8% of 7,534 Sonar citations for software recommendations pointed at domains outside the 100,000 most-visited sites, and three sites under apparent common control had published 215,128 generated 'best software' pages.
- In the EBU and BBC study of free consumer assistants, 30% of Perplexity's 681 responses had a significant issue and 15% a significant sourcing issue, against 36% and 24% for ChatGPT.
- In the Tow Center's test of eight AI search tools, Perplexity answered 37% of its 200 source-identification queries incorrectly. Grok 3 got 94% wrong.
Open the page behind every Perplexity figure before you use it. Independent audits found failure rates from about one claim in seven to one citation in three, depending on what they counted. The misses cluster on facts spread in pieces across several pages, such as a company's current CEO.
What the audits measured
Each study counted something different, so read their rates side by side and never add them up. Haus Research and Trellner called Sonar and Sonar Pro through the API on September 2, 2026 and checked every cited page. The EBU and BBC study, published in October of 2025, rated free consumer versions of four assistants on news questions. In March of 2025, the Tow Center asked eight tools to identify the sources of news excerpts.
All four tested the API or free versions, so measure Enterprise seats in your own pilot. Have the pilot team open the page behind every figure they use for two weeks and count the misses.
Haus passed a citation when its page held at least one figure from the sentence, so its 34.7% is the floor for unsupported citations. Most claims carry several citations, which brings the per-claim rate down.
| Study | Failure rate | What was counted | Product tested |
|---|---|---|---|
| Haus Research (Sep 2026) | Citations whose page did not open or held none of the sentence's figures, of 1,826 | Sonar and Sonar Pro API | |
| Haus Research (Sep 2026) | Claims with no supporting citation, of 872 | Sonar and Sonar Pro API | |
| EBU and BBC (Oct 2025) | Responses with a significant sourcing issue, of 681 | Free consumer version | |
| Tow Center (Mar 2025) | Incorrect answers to source-identification queries, of 200 | Free version |
Haus Research (Sep 2, 2026, n=1,826 citations and 872 claims); EBU and BBC (Oct 2025, n=681 Perplexity responses); Tow Center for Digital Journalism (Mar 6, 2025, n=200 queries per tool).
Where citations fail
The misses follow how well a fact is written down. Haus Research found pages carried the cited figure most often for headcount and founding year, which sit in structured fields. They carried it least often for the current CEO or an office street address, which are repeated everywhere and rarely published in full in one place. Several answers cited Wikipedia for addresses its articles did not contain.
- Headcount: 82.0% of 167 citation pairs passed.
- Founding: 75.0% of 128 passed.
- Revenue or ARR: 72.4% of 192 passed.
- Entry price: 62.6% of 171 passed.
- Headquarters address: 53.0% of 202 passed.
- Current CEO: 44.3% of 235 passed.
Which sources get cited
Trellner found that buyer-question sources favor small, new sites. Across 380 software categories, the median traffic rank of cited domains in the top million was 71,611. A blog from an interactive product-tour vendor was third most cited. Wikipedia appeared three times in 7,534 citations. EBU and BBC reviewers often rated Perplexity's sourcing well but also found unrelated articles in its source lists.
Of 13 public accounts that discuss Perplexity's accuracy, 10 raise concerns. One Hacker News commenter reported citations that said the opposite of the claims beside them.
What it means for a team
Perplexity's documentation says a source label covers a whole site and never replaces reading the source. The audits show where that check matters most.
- Open the cited page for every figure that leaves the thread, and prefer facts published in one canonical place.
- When a cited page sits behind a login, have someone with access read it before the figure goes out.
- Check buying recommendations against independent reviews, since generated 'best' lists rank among the most-cited sources.
- Use Model Council or a second reviewer when a decision rests on one answer.
Methodology
Every failure rate is the publisher's own, linked in Sources. GILD read the three independent audits and the EBU and BBC study in full on October 12, 2026. Haus Research and Trellner tested Perplexity's Sonar API models on September 2, 2026. The EBU and BBC published their study of free consumer assistants in October 2025, and the Tow Center for Digital Journalism published its study on March 6, 2025. Product details come from Perplexity's Help Center. Counts of public accounts come from 55 public sources published January 4, 2025 through September 28, 2026, including the Haus Research, Trellner, and Tow Center studies.
Products in this piece
Sources
- Official documentation
- 5
- Public source
- 5
- Perplexity Help Center: Understanding source labels
- Perplexity Help Center: What is Research mode?
- Perplexity Help Center: What's New in Advanced Deep Research
- Perplexity Help Center: What is Model Council?
- Perplexity Help Center: What advanced AI models are included in my subscription?
- Haus Research: A third of Perplexity's citations don't contain the number they're cited for (Sep 2, 2026)
- Trellner: Three sites made 215,128 "best software" pages for AI. Perplexity cites them (Sep 2, 2026)
- EBU and BBC: News Integrity in AI Assistants (Oct 2025)
- Columbia Journalism Review, Tow Center: AI Search Has a Citation Problem (Mar 6, 2025)
- Hacker News comment on citations that say the opposite of the claim (Feb 28, 2026)