A developer can build a source-cited software recommender on Perplexity Sonar and still surface an evidence trail partly supplied by 215,128 generated buying guides. In a September 2, 2026 audit, three sites behind those pages received 181 of 7,534 citations returned by Sonar and Sonar Pro. That was only 2.4 percent of the citations, yet it shows how a publisher can manufacture material for thousands of buyer questions and enter the retrieval layer of an AI answer engine, according to Trellner Research's report.
The finding matters because a citation changes how an answer feels. A list of five products accompanied by links looks inspectable and grounded. The new audit shows that the links can lead to marketing posts, low-ranked domains and pages that appear designed for machine retrieval. A citation records where the system looked; its presence alone says little about how the source researched a market or whether the named product still lives at the supplied domain.
One search stack sampled twice
Trellner wrote 380 buyer-intent categories before running the test, ranging from CRM software to museum collection management software. Each category went once to perplexity/sonar and once to perplexity/sonar-pro through OpenRouter, for 760 calls. The prompt required a ranked top five in JSON with each product's official homepage domain. All calls returned parseable answers, producing 3,800 recommendation slots for 1,807 distinct products and 7,534 source URLs across 2,055 domains. The researchers published the method, raw answers, citation table and analysis scripts.
Those two model runs do not provide independent confirmation. Their citation lists were byte-for-byte identical in 289 of the 380 categories, while the overlap between their URL sets had a Jaccard score of 0.898. They also chose the same first-place product in 290 categories. Trellner therefore describes the result as one retrieval stack sampled twice, a limitation made explicit in the report's methodology.
The scope is narrower than the headline might suggest. This was a one-day API experiment through OpenRouter, with one prompt wording and one run per category. It did not test Perplexity's consumer website. Google was excluded because its grounded model on OpenRouter would have used OpenRouter's search plugin, which would measure a different retrieval system. ChatGPT, Gemini, Copilot and Google AI Mode are outside the result, as the authors state.
Popularity was a poor predictor of retrieval
Nearly six in ten citations pointed to domains ranked below 100,000 on the September 1 Tranco list, and 23.4 percent pointed to domains absent from its top million. Among the 2,055 cited domains, 751 were outside that million. The median Tranco rank among citations to ranked domains was 71,611. Tranco measures popularity and does not score quality, so a low rank cannot establish that a page is unreliable. It does establish that Perplexity's source pool extended far beyond the sites most people visit, which is the limited use Trellner makes of the metric in its published analysis.
The retrieval was spread across that long tail. The ten most-cited domains accounted for 17.3 percent of all citations. G2 led with 291, followed by Reddit with 261. Guideflow placed third with 194, ahead of Gartner's 158, while Wikipedia appeared only three times in the full set. Those counts do not grade any source. They show which pages Perplexity reached for while constructing answers to this particular set of software-buying prompts, according to the domain table in the report.
A vendor's blog became a general buying guide
Guideflow sells interactive product demos and does not compete in any of the 380 categories. Its marketing blog nonetheless supplied 194 citations across 96 categories, including 3D rendering, IVR and RFID software. Every citation used a different Guideflow URL. Trellner counted 3,351 blog URLs in its sitemap, representing 2,176 distinct posts, and found that this single vendor blog became Perplexity's third-largest evidence source in the run. The report makes no allegation of deception by Guideflow; companies routinely publish search-oriented comparison posts.
The useful question sits downstream, inside retrieval. A vendor writing about markets it does not serve can rank above an established research firm in the evidence gathered for product selection. Perplexity's own Sonar documentation says the user message drives the web search and that Sonar returns source URLs in citations and search_results. Those fields make the retrieval trace visible to developers, but they do not contain an editorial audit of each publisher.
Three sites generated more lists than there are categories
WifiTalents, WorldMetrics and Gitnux contributed 181 citations across 41 categories. Trellner found that the three domains used the same page template and navigation, the same NameCheap registrar, and the same pair of Cloudflare nameservers. Each had a blog containing exactly six posts about the other brands in the set and a fourth related domain. That is strong circumstantial evidence of common control, which the report labels as an inference because none of the sites identifies an owner.
Their sitemaps held 70,731, 71,684 and 72,713 URLs under /best/{category}-software/, adding up to 215,128 buying guides. WorldMetrics and Gitnux gave their homepages HTML titles ending in "Facts & Grounding Page." Their matching meta descriptions called the page a machine-readable record. "Grounding" is the retrieval step that supplies documents to an answer model, and the sites' own wording indicates an intended machine audience, as documented in the saved evidence and sitemap counts.
Scale did not produce a consistent verdict. On the same "project estimation software" page, WorldMetrics and WifiTalents both ranked Float first, while Gitnux chose Saviom and placed Float fourth. The three pages credited nine different named staff members. Two displayed the unresolved phrase "Within the next 26 days" in a byline, and the third said 40 days. Gitnux still labeled its list "AI-verified · Expert reviewed," details captured in the September 2 snapshot.
Wrong domains survived the polished answer
The models supplied 1,502 distinct vendor homepages, and most worked. Ten names had no DNS address, four resolved but did not answer either of the researchers' two checks, and three returned a broken HTTP status. In all, 17 domains, or 1.1 percent, were gone or unreachable. Another 92 redirected to a different registrable domain. Most redirects reflected ordinary acquisitions or rebrands, so Trellner published the raw checks instead of treating every redirect as an error in its vendor-domain dataset.
Two tier disagreements show why applications should validate the destination as well as the product name. For research data management, Sonar Pro returned Dryad's real home at datadryad.org; Sonar supplied dryad.co, which redirected to an Indonesian gambling site when checked. For the data-observability company Monte Carlo, Sonar used montecarlodata.com, while Sonar Pro supplied montecarlo.com, the address of a Monaco hotel and casino group. These examples came from the models' homepage fields, and the audit does not claim that the 215,128-page network caused them, as the limitations section stresses.
A separate May 2026 preprint by Mowafak Allaham and Nicholas Diakopoulos offers wider context without confirming Trellner's exact result. It audited 712 human-written queries about politics, health and the environment across ChatGPT, Copilot, Gemini and Perplexity. The authors reported evidence of AI-generated material in roughly 16 percent of cited sources. Its questions and detection method differ from Trellner's software-buying test, but both studies locate a quality problem in the material retrieved before an answer is written.
Developers have controls, with limits
A system prompt cannot impose a hard source boundary on Sonar retrieval. Perplexity's official prompt guide says system messages influence generation only, while the user message seeds the search. The company directs developers to request parameters for domain, date and recency constraints. Its domain filter can allow or exclude as many as 20 domains, though an allowlist and denylist cannot be combined in one request.
That 20-domain ceiling makes a fixed allowlist awkward for open-ended software discovery, where the products are unknown at query time. A safer application design is a second verification pass: resolve each supplied vendor domain, follow redirects, match the destination to the named company and retrieve product claims from official documentation. Independent testing can then be kept separate from vendor pages. This workflow follows from the failure cases in the audit and the source controls Perplexity exposes in its API documentation.
The next useful test would repeat categories over several days, vary "best" against other buyer language, and remove the cited list network to measure whether recommendations change. Trellner performed none of those interventions. Its snapshot establishes which documents supplied evidence, while the causal effect on product rankings remains unmeasured. Until that work is done, a citation list records what a retrieval system found. Product identities, domains and comparative claims still need their own checks.