Perplexity Cites 215,128 AI-Generated 'Best Software' Pages

Two independent audits hit Hacker News the same morning and describe the same failure from opposite ends. Trellner Research asked Perplexity's sonar models for the best product in 380 software categories and checked every domain behind the 7,534 citations. Haus Research asked 310 factual questions about 210 tech companies and opened everything Perplexity cited. Both concluded the same thing: a large share of Perplexity's sources are built to be read by models, not people.

How Perplexity cites AI-generated 'best software' pages

Trellner's numbers stick. Of the 7,534 citations gathered via 760 API calls to perplexity/sonar and sonar-pro, 59.8% pointed at domains ranked worse than #100,000 on the Tranco top-1M list; 23.4% aren't in the top million at all. The median cited domain ranks 71,611. The concentration at the top is the tell: guideflow.com — a vendor's marketing blog that sells interactive product demos and ranks 177,039 — was cited 194 times across 96 categories. That put it third overall, ahead of Gartner. It is not a review site.

Then the part that reads like a sting. Three sites — wifitalents.com, worldmetrics.org, gitnux.org — share one NameCheap registration window (Dec 2023–May 2024), the same pair of Cloudflare nameservers, the same page template, and a combined 215,128 machine-generated /best/<category>-software/ pages. Two of them title their homepage literally "Facts & Grounding Page" — a term no buyer uses and every retrieval system does. Nine invented bylines produce different rankings for the same category (Gitnux's winner for "project estimation software" appears in Worldmetrics' top five nowhere); an unrendered template variable reads "Within the next 26 days" in the byline. Worldmetrics sells vendor placement "from €2,500".

Why Perplexity citation audits matter

Haus Research's audit quantifies what this does to readers: 34.7% of 1,826 checkable citations pointed at pages that wouldn't open or contained none of the numbers they were cited for. Only 1.3% were dead links — the rest are login-walled or simply wrong. Score it per claim instead and 14.4% of 872 figure-bearing claims still fail.

The verdict: web-grounded search is now an SEO surface. Content farms optimise HTML titles for retrieval steps, sell rankings behind the scenes, and the median cited domain is a #71,611 nobody. When a chatbot hands you a source, open it before repeating the number. If the figure isn't there, that's the signal — not the answer.