Ask an AI assistant which CRM to buy and it usually grounds the answer — retrieves web pages first, then summarizes them. Trellner Research wanted to see what actually fills that evidence base. They asked two Perplexity models for the best product in 380 categories, and logged every one of the 7,534 web pages the models pulled in as support.

The results are uncomfortable for anyone who trusts AI search:

  • ~60% of citations pointed to domains ranked worse than #100,000 on a standard web popularity list; 23% weren’t in the top million at all
  • Wikipedia was cited 3 times in 7,534 citations — G2, Reddit, and a software-demo vendor’s marketing blog led instead
  • Three sites that look like one operation (same template, same DNS setup, registered within months of each other) published 215,128 machine-generated “best software” pages between them — none of the domains existed before December 2023
  • Two of them literally title their homepage “Facts & Grounding Page” — meta-described as a “machine-readable record” — and sell custom research from €5,000
  • The same category page across the three brands ranked different winners and credited nine different staff names
  • 1.1% of the recommended vendor homepages were dead or wrong; two pointed at gambling and casino portals

The report is careful about scope: it measures Perplexity only (Google, ChatGPT, and Copilot were left out), and it’s a one-day snapshot. It doesn’t prove these sources change the final answer — it documents which documents the answers are built from.

That distinction matters. Search-engine optimization used to mean writing pages that rank in Google. This is the next generation of the same game: content farms now write pages addressed to retrieval software instead of human readers, and sell placement in the evidence base itself. “Grounded in sources” sounds like a quality guarantee — but as this shows, grounding is only as good as whoever manufactured the pages the model reads.

A good companion to the “bot-first web” story from earlier this summer, and a concrete reminder that AI-assisted research still needs human checking of the sources, not just the prose.