AI retrieval is starting to look less like search and more like a new procurement attack surface.
Trellner Research tested Perplexity's web-grounded models across 380 software buying categories, with 760 total model calls and 7,534 citations. The result was ugly: 59.8 percent of citations pointed to domains ranked worse than 100,000 in Tranco, and 23.4 percent pointed outside the Tranco top million entirely.
The more important finding is structural. Three apparently related sites had generated 215,128 “best category” software pages, and Perplexity cited those sites 181 times across 41 of the 380 categories. Two of the sites even used the homepage title “Facts & Grounding Page,” language aimed at retrieval systems more than human buyers.
Grey Haven's read: grounded answers are only as good as the retrieval substrate, and that substrate is now economically worth manipulating. This is not just an SEO story. It is a source-governance problem for AI copilots used in vendor selection, competitive research, analyst workflows, and vertical SaaS discovery.
Operators should stop treating a citation list as proof of quality. The practical control is a source-quality layer: domain age, known publisher status, vendor homepage confirmation, citation diversity, and explicit rejection of generated listicle farms. If a buying assistant cannot explain why its sources deserve trust, it should not rank vendors.
Watch whether enterprise AI tools start exposing retrieval provenance controls instead of hiding them behind a polished answer. The story becomes more important if manufactured-source networks begin targeting narrow industrial, healthcare, government, or compliance categories where buyers have fewer trusted references.
Source: Trellner Research, “Three sites made 215,128 ‘best software’ pages for AI. Perplexity cites them.”
