Three Sites Generate 215K Pages to Manipulate Perplexity

A recent analysis of AI-driven software recommendations reveals that retrieval-augmented generation models heavily rely on low-authority, machine-generated, and vendor-operated websites to condition their outputs. The study evaluated two Perplexity AI models, Sonar and Sonar Pro, across 380 buyer-intent software categories, retrieving 7,534 citations. Distribution analysis shows that 59.8 percent of these citations point to domains ranked outside the top 100,000 in global web traffic, with 23.4 percent falling entirely outside the top million. The median rank of cited domains sits at 71,611, indicating a decentralized but predominantly long-tail evidence base rather than a concentration of established publishers. The most significant finding involves a coordinated network of three domains that share identical Cloudflare infrastructure, page templates, and navigation structures. Registered between late 2023 and mid-2024, these sites collectively host over 215,000 machine-generated buying guides formatted as best category pages. Their homepage titles and meta descriptions are explicitly optimized for AI retrieval systems, referencing internal fact-checking records and editorial methodologies tailored to machine consumption rather than human readers. Despite minimal human-focused content, the trio accounts for 181 citations across multiple software verticals, demonstrating how structured, algorithmically produced pages can dominate AI grounding layers. Vendor marketing material also emerged as a primary citation source. A company selling interactive software demos contributed 194 citations across 96 distinct categories, outpacing major industry analysts. Each citation linked to a uniquely generated blog post optimized for specific software niches the vendor does not directly operate in. The analysis confirms that AI retrieval systems frequently treat commercial blogs and directory-style listings as authoritative evidence when synthesizing purchasing recommendations. Minor link rot and misrouting were noted among cited vendor homepages. Of 1,502 direct product URLs supplied by the models, 1.1 percent resolve to unreachable domains, while 6.1 percent redirect to unrelated registrable domains. In two instances, AI outputs directed users toward gambling and casino websites due to identical naming conflicts with legitimate software providers. The study emphasizes that these findings reflect citation distribution rather than answer accuracy, and the models share a common retrieval backend, meaning results should be interpreted as a single measurement rather than independent convergence. Researchers note that the evaluation covers Perplexity exclusively and utilizes a niche-heavy category list that amplifies long-tail sourcing. No conclusions are drawn about competing platforms or the actual quality of the software recommendations. The underlying dataset, including domain verification logs and methodological scripts, has been published openly. The broader implication is clear: as AI systems increasingly depend on external web grounding, their outputs remain highly susceptible to the structure and visibility of algorithmically optimized, low-traffic content ecosystems.