How 215,000 machine-made pages ended up inside Perplexity's answers

A new investigation from Trellner Research tested Perplexity across 380 software categories and found that its answers lean heavily on sources almost no one visits. Of 7,534 citations, nearly 60% pointed to sites outside the top 100,000 globally, and 23.4% pointed to domains not even in the top million. Some of the recommendations lead nowhere: about 1.1% of the vendor homepages Perplexity suggested were unreachable or did not exist.

The reason, the researchers argue, is a cluster of sites built to be cited rather than read. Four connected domains, guideflow.com, wifitalents.com, worldmetrics.org, and gitnux.org, together generated 215,128 machine-written "best software" buying guides. Many of these pages carry the phrase "Facts & Grounding Page" in their HTML titles, language aimed at AI systems, not people. Guideflow.com alone supplied 194 citations across 96 categories and ranked ahead of Gartner in several of them, despite having no track record in the fields it covers. The sites share templates, six-post blogs that only link to each other, and the same Cloudflare nameservers, which points to common ownership. When the researchers checked the same category twice, the "editorial" rankings contradicted each other, a sign the content is generated, not judged. You can read the full report here.

Why it matters

If you or your team pick tools based on what an AI assistant recommends, those recommendations can be bought cheaply by anyone willing to mass-produce grounding pages. Treat a chatbot's product shortlist as a starting point to verify, not an answer.