Perplexity search cites network of sites with 215,000 AI-made buying guides

Perplexity’s search engine draws a large share of its shopping recommendations from an obscure corner of the web built specifically to be read by AI systems, according to an independent study published on trellner.com.

The researchers asked two Perplexity models, sonar and sonar-pro, to rank the best products in 380 software categories, from CRM tools to museum collection management software. Across 760 calls made through OpenRouter on 2 September 2026, the two models returned 7,534 citations pointing to 2,055 distinct domains. Nearly 60 percent of those citations led to sites ranked worse than 100,000th on the Tranco list, a measure of how much traffic a domain gets. Almost a quarter did not appear in the top million domains at all.

Three of the most obscure domains, worldmetrics.org, gitnux.org and wifitalents.com, turned out to share infrastructure. All three were registered through NameCheap between December 2023 and May 2024, all route through the same two Cloudflare nameservers, and all use an identical page template. Together they had generated 215,128 machine-written best-software buying guide pages, far more than there are real software categories to rank. Two of the three homepages carried the HTML title “Facts & Grounding Page,” language addressed to a retrieval algorithm rather than a human shopper. Worldmetrics also sells custom market research starting at 5,000 euros and “vendor selection” services from 2,500 euros, advertised on the same pages the researchers say the AI models are citing.

A vendor’s blog outranked Gartner

The third-biggest source of citations belonged to guideflow.com, a company that sells interactive product demos and does not compete in any of the 380 categories tested. Its blog was cited 194 times across a quarter of the categories, more often than the research firm Gartner. Wikipedia, by comparison, was cited three times in 7,534.

The study’s authors caution that they measured only Perplexity, not ChatGPT, Gemini or other AI search tools, and that their 380 categories were their own construction rather than a sample of real buyer searches. They also note that citing an obscure or newly registered site is not proof its advice is wrong, and that they did not test whether removing these sources would change the models’ final recommendations.

The researchers found smaller errors too. Asked for research data repositories, one Perplexity model correctly named datadryad.org while the other supplied dryad.co, which redirects to an online gambling site. The full dataset, including every citation and the scripts used to analyze it, has been released under a Creative Commons license. The authors say Perplexity’s retrieval layer changes daily, so the figures describe one snapshot rather than a fixed state, and note they have not yet checked whether other AI search tools cite the same network of sites.