Content farms built for AI, not humans, are grounding Perplexity's software picks

Researchers queried Perplexity’s Sonar and Sonar-Pro models for the top five products across 380 software categories—760 calls yielding 7,534 citations—and traced where the answers came from. The retrieval layer leans heavily on the long tail: 59.8% of citations point to domains ranked below #100,000 on the Tranco list, and 23.4% sit outside the top million entirely. The median cited domain ranks around 71,600. For contrast, Wikipedia surfaced just three times.

Much of that evidence base turns out to be manufactured for machine consumption. Three sites—wifitalents.com, worldmetrics.org, and gitnux.org—appear to be a single operation: registered through NameCheap between late 2023 and mid-2024, sharing the same pair of Cloudflare nameservers and an identical page template. Between them they’ve published 215,128 auto-generated “best ” buying guides while maintaining just six real blog posts each. Two literally title their homepages “Facts & Grounding Page” and describe themselves as machine-readable records—text addressed to retrieval systems rather than people—while selling custom market research starting at €5,000. The same category page yields different winners across the three brands, credits nine fictitious-looking staffers, and leaks unrendered template variables like “Within the next 26 days.” Separately, guideflow.com, a vendor’s marketing blog for products in none of the surveyed categories, was cited 194 times, ranking third overall—ahead of Gartner.

The significance is a new incentive structure: grounded AI recommendations reward whoever floods the index with plausible, structured, machine-optimized content, not whoever has authority or traffic. Nothing here is technically deceptive—these are ordinary content-marketing tactics—but they reveal that “grounding” can be gamed cheaply, and that buyers trusting AI product advice may be reading synthetic listicles engineered to be retrieved. A sanity check on the actual recommendations found 1.1% of vendor homepages dead or unreachable, including wrong domains offered for GraphiQL, Microsoft To Do, and Trivy.