Perplexity cited three manufactured websites holding 215,128 machine-generated “best software” pages among its top sources for software recommendations, according to a Trellner Research study (TR-2026-009) published on 2 September 2026. The study found that 59.8% of 7,534 Perplexity citations point to domains ranked worse than #100,000 globally.
Trellner Research queried Perplexity’s sonar and sonar-pro models across 380 software buyer-intent categories, such as CRM, help desk, and payroll software, in 760 API calls on 2 September 2026. The queries returned 7,534 citations from 2,055 distinct domains and 3,800 recommendation slots naming 1,807 distinct products. Trellner measured Perplexity because Perplexity exposes its source URLs; the study makes no claim about ChatGPT or Google Gemini.
What Trellner Found in Perplexity’s Sources
Low-authority domains supply the majority of Perplexity’s software-recommendation evidence. 59.8% of the 7,534 citations point to domains ranked worse than #100,000 on the Tranco list, and 23.4% point to domains outside the top 1 million entirely. Wikipedia appears 3 times in the full citation set.
The 10 most-cited domains were g2.com (291 citations), reddit.com (261), guideflow.com (194), gartner.com (158), zapier.com (82), wifitalents.com (71), capterra.com (68), linkedin.com (67), worldmetrics.org (60), and gitnux.org (50). Guideflow sells interactive product demos and operates no review business; Perplexity cited its content-marketing blog across 96 of the 380 categories, placing it ahead of Gartner. Trellner describes the Guideflow result as a retrieval-layer outcome, not deceptive conduct by Guideflow.
Who Runs the 215,128 “Best Software” Pages
Three near-identical sites — wifitalents.com, worldmetrics.org, and gitnux.org — list 215,128 generated /best/<category>-software/ pages in their sitemaps, according to Trellner. All three were registered through NameCheap between December 2023 and May 2024, share the same two Cloudflare nameservers, and use an identical page template. Trellner calls this strong circumstantial evidence of a common Cloudflare account and stops short of naming a single operator.
Two of the sites, worldmetrics.org and gitnux.org, title their homepages “Facts & Grounding Page.” Grounding is the retrieval step in which an AI system fetches documents to condition an answer. Trellner’s report states that these pages address the software that reads them, not human buyers, and summarises the risk in one line: “An AI cited it is not the same claim as a person vetted it.”
Why the Finding Matters for Software Buyers
AI answer engines are now a primary discovery channel for business software, and this study shows the evidence layer under those answers is dominated by low-authority, machine-addressed pages. A buyer who asks Perplexity for the best help-desk tool receives a ranked list whose supporting citations come, in 6 of 10 cases, from sites that almost no humans visit.
The pattern mirrors the early web-spam arc of search engine optimization. Generative engine optimization (GEO) is the practice of shaping pages so that AI search engines retrieve and cite them, and the three “Facts & Grounding” sites apply GEO at industrial scale: programmatic category pages, machine-readable titles, and no editorial layer. Buyers comparing tools should treat an AI recommendation as a starting list and verify each product against human-edited sources, such as our independently researched guide to the 15 best AI tools for business in 2026, where every tool was tested rather than retrieved.
For context: AI answer engines are turning into commercial surfaces on two fronts at once. Our report on ChatGPT ads going live in 7 countries covered paid placement inside AI answers; the Trellner study documents the unpaid route — manufacturing the sources those answers rely on. Our earlier coverage of researchers cracking hidden AI reasoning in OpenAI, Claude, and Gemini APIs showed how little of an AI answer’s evidence chain is visible to the person reading it.
Our Take: Perplexity’s retrieval layer does not know what a trustworthy software review looks like; it knows what looks like one to a machine. Three sites registered in 2023–2024 with 215,128 templated pages between them now hold 3 of the top 10 citation slots, one of them ahead of Capterra, and Wikipedia barely registers. Until AI search engines weight source authority the way a human editor does, software buying research still starts with people who have used the product — not with an answer box.
