Technology
Perplexity software recommendations rely heavily on low-traffic and machine-generated pages, test finds
An analysis of 760 software queries showed that Perplexity frequently cites marketing blogs and newly created template sites designed for AI retrieval.
The short version
- A study of 380 buyer queries across two Perplexity AI models revealed that nearly 60% of cited sources fall outside the top 100,000 most-visited websites.
- Perplexity frequently relied on machine-generated listicle sites and vendor content blogs rather than established software directories or review publishers.
- A network of three newly created domains containing over 215,000 AI-targeted pages and sharing identical infrastructure accounted for 181 citations.
- The analysis evaluated only Perplexity's shared retrieval layer on a single date, leaving open how other AI search engines handle similar queries.
Key facts
- In a 760-prompt test across 380 software categories, 59.8% of Perplexity's 7,534 retrieved citations pointed to domains ranked worse than #100,000 on the Tranco list, with 23.4% outside the top one million.[Hacker News]
- Three interconnected domains created after December 2023—worldmetrics.org, wifitalents.com, and gitnux.org—published 215,128 machine-generated buying guide pages explicitly labeled as facts and grounding records for AI models.[Hacker News]
- Guideflow.com, a product demo vendor's marketing blog, was the third most-cited source overall with 194 citations across 96 categories, ranking above Gartner.[Hacker News]
- Out of 1,502 vendor links provided by the models, 17 were unreachable or 404 errors, and two redirected to gambling and casino portals due to domain mismatches.[Hacker News]
What remains uncertain
- The true ownership and operators of the three automated guide networks remain unconfirmed, though common control is indicated by shared Cloudflare nameservers, identical templates, and mutual cross-linking.[Hacker News]
- The study did not test whether removing these machine-generated or marketing sources would alter the final product recommendations, nor did it evaluate competitors such as ChatGPT, Gemini, or Copilot.[Hacker News]