215,128 fake 'best software' pages are quietly grounding Perplexity's answers

A new audit of Perplexity’s retrieval layer has surfaced something uncomfortable for anyone who trusts grounded AI answers: 59.8% of the citations behind its product recommendations point at domains ranked worse than #100,000 in the Tranco top-1M list, and three sites operating under apparently common control have published 215,128 machine-generated “best software” pages between them. The report, published by Trellner Research on 2 September, put 380 buyer-intent categories to perplexity/sonar and perplexity/sonar-pro through OpenRouter, one prompt per category per model, 760 calls in all. Every call asked for a ranked top five as JSON with each product’s official homepage domain.

The result: 7,534 citations spanning 2,055 distinct domains, of which 751 do not appear in the Tranco top million at all. That is 36.5% of cited domains living entirely outside the widely-visited web. The median first Wayback capture for those unranked domains is 2020, against 2011 for ranked ones, and 16.6% of the archived unranked domains were first captured in 2025 or later. Wikipedia, for comparison, was cited three times in 7,534 citations.

The headline number is the scale of the manufactured evidence base. Three sites, wifitalents.com, worldmetrics.org and gitnux.org, list 103,578, 107,083 and 105,541 URLs in their sitemaps respectively, of which 70,731, 71,684 and 72,713 are /best/<something>-software/ pages. There are not 215,128 software categories. All three were registered through NameCheap between December 2023 and May 2024. All three delegate DNS to the same pair of Cloudflare nameservers, pam.ns.cloudflare.com and sean.ns.cloudflare.com. All three run the same page template with the same navigation: Services, Market Data, Software Advice, Editorial Process, Company.

What makes these sites unusual is not their volume but their self-description. Fetched on 2 September 2026, worldmetrics.org and gitnux.org both return an HTML title of the form “Brand — Facts & Grounding Page”. The meta description reads: “Verified facts about Gitnux: an independent market research company publishing industry statistics, custom research, and software Best Lists. Company, legal, methodology, and compliance details in one machine-readable record.”

Grounding is not a term buyers use. It is the name of the step in which a retrieval system fetches documents to condition an answer on. These pages are addressed, in their titles and descriptions, to the software that reads them. A fourth brand, zipdo.co, sits on the same nameserver pair and gives its own homepage the same “Facts & Grounding Page” title.

The three brands account for 181 citations across 41 of the 380 categories, 2.4% of the total. But the report’s more striking finding is the third-largest single source overall: guideflow.com, a vendor marketing blog. Guideflow sells interactive product demos. It is not a review site, a directory or a publisher, and it competes in none of the categories asked about. Its blog was nonetheless cited 194 times across 96 of the 380 categories, a quarter of them, placing it third overall and ahead of Gartner. Each citation is a different URL: 96 distinct guideflow.com blog URLs, one per category, six of them the Estonian-locale copy of a post. Its sitemap lists 3,351 blog URLs, 2,176 of them distinct posts. It supplied the grounding for “3D rendering software”, “IVR software”, “RFID software” and “architecture practice software” alike.

Nothing about Guideflow is deceptive. It publishes a large content-marketing blog, as thousands of companies do. The measurement is about what the retrieval layer does with it: a vendor’s own listicles about markets it does not operate in became the third-largest evidence base for a question about which product to buy.

The ten most-cited domains tell the story. G2 leads with 291 citations at 3.86%, followed by Reddit at 261. Guideflow sits third at 194. Gartner is fourth at 158. Then Zapier at 82, WifiTalents at 71, Capterra at 68, LinkedIn at 67, Worldmetrics at 60 and Gitnux at 50. The median Tranco rank of the 5,768 citations that point at a ranked domain is 71,611. Concentration at the top is unremarkable, the report notes: the ten most-cited domains take 17.3% of citations. The story is not that a cartel of famous sites supplies the answers. It is what fills the other four-fifths.

The report also tested whether the cited vendor homepages are real. Of the 1,502 vendor homepages the models supplied, 17 domains, 1.1%, are gone or unreachable. Ten resolve to no address at all, including graphiql.com (offered as the home of GraphiQL, which has no such site), todo.com (offered for Microsoft To Do) and aquasecurity.io (offered for Trivy). Two are worse. Asked for research data management platforms, both models named Dryad: sonar-pro gave the real repository at datadryad.org, sonar gave dryad.co, which redirects to an Indonesian online-gambling portal whose title begins “BIGSLOT288 | Portal Game Online”. Asked for data quality tools, both named Monte Carlo: sonar gave montecarlodata.com, sonar-pro gave montecarlo.com, which redirects to Monte-Carlo Société des Bains de Mer, the Monaco hotel and casino group.

The report is careful about its limits. The two Perplexity tiers are not two independent measurements: they returned a byte-identical citation list in 289 of the 380 categories and their URL sets overlap at a Jaccard of 0.898. The agreement on the top pick, the same product first in 290 of 380 categories, is a fact about a shared retrieval layer, not evidence that independent systems converge. The result covers Perplexity only. The categories are the researchers’ own construction, weighted towards niche verticals, which will surface more long-tail sources than common queries would. Every page fetch went out through a rotating datacentre proxy under a named research user-agent, so what these sites returned is not necessarily what they return to a retrieval crawler or to a browser.

The report also does not claim the sources change the answers. It did not test whether removing these sources would produce different recommendations. Guideflow and the three Best List brands may well name reasonable products. What was measured is which documents the evidence base is made of.

That distinction matters, because the evidence base is where the rot starts. A retrieval system that cites a page titled “Facts & Grounding Page” is not being tricked by a clever adversary. It is doing exactly what its designers built it to do: fetch documents that look authoritative, rank them by relevance signals, and condition the answer on them. The pages were built to be read by models, and the models read them. The failure is not in the content farm. The failure is in the retrieval layer’s inability to distinguish a document written for a human buyer from a document written to satisfy a grounding step.

The economics here are worth stating plainly. Worldmetrics advertises custom market research “from €5,000”, ready-made reports “from €499” and vendor selection “from €2,500”, above the same taxonomy of generated Best Lists that the models retrieve. The operation monetizes both sides: it sells research to vendors who want placement, and it sells the appearance of authority to the retrieval systems that now mediate software purchasing decisions. The report notes that none of the three brands names an owner, and common control is inferred from shared infrastructure and an identical template, not proven.

For AI builders, the takeaway is not that Perplexity is uniquely broken. It is that grounded generation inherits the pathologies of the open web, then amplifies them through scale. A human reader glancing at a Guideflow listicle can discount it as marketing. A retrieval system has no such reflex. It treats a vendor’s blog post about a market the vendor does not operate in as equivalent evidence to a Gartner report, and it treats a machine-readable self-description as a fact about the world.

The report’s own framing is the correct one. Tranco rank is a popularity measure, not a quality measure, and a low rank is not an accusation. The problem is not that obscure sites get cited. The problem is that the retrieval layer has no mechanism for asking why a site exists, who wrote it, or whether its incentives align with answering the question. A page titled “Facts & Grounding Page” is a confession of purpose, and the system that retrieves it cannot read the confession.

The last line of the report is worth sitting with. Asked for the same category page, “project estimation software”, from all three brands, each page ranks ten tools and states its ranking in JSON-LD. Worldmetrics puts Float first, Scoro second, Teamwork.com third. WifiTalents puts Float first, Scoro second, Teamwork.com third, then Buildertrend and Apropo. Gitnux puts Saviom first; its winner does not appear in Worldmetrics’ five at all. Each page credits three named staff, nine distinct people for one question, and each carries an unrendered template variable in the byline line reading “Within the next 26 days” on two of them and “Within the next 40 days” on the third. The templates were not even finished rendering before they became the evidence base for what software buyers should purchase.