AI Search Cites 215K Machine-Generated Software Pages

The Hidden Architecture of AI Recommendations

When you ask an AI assistant for the best CRM software or the top project management tool, where does the answer actually come from? A new investigation by Trellner Research suggests the evidence base behind AI recommendations may be far less authoritative than users assume.

The study, conducted on September 2, 2026, analyzed 7,534 citations from Perplexity's Sonar and Sonar Pro models across 380 software categories. The results paint a concerning picture: 59.8% of citations point to domains ranked worse than #100,000 on the Tranco top-1M list, and 23.4% fall outside the top million entirely.

More troubling still, three sites under apparent common control—WorldMetrics, WifiTalents, and Gitnux—have published a staggering 215,128 machine-generated 'best software' pages between them. None of these domains existed before December 2023, yet Perplexity cited them 181 times across 41 categories.

The Scale of the Problem

The research team put 380 buyer-intent categories—from "CRM software" to "museum collection management software"—to Perplexity's models through OpenRouter. Each call requested a ranked top five with official product homepages. The results: 3,800 recommendation slots, 1,807 distinct products, and 7,534 citations spanning 2,055 domains.

The median Tranco rank of cited domains was 71,611, but the real story lies in the long tail. 751 of the 2,055 cited domains (36.5%) don't appear in the top million at all. These unranked domains are also newer—median first Wayback capture of 2020 versus 2011 for ranked sites—and 16.6% were first captured in 2025 or later, compared to just 1.6% for ranked domains.

Guideflow: A Marketing Blog Outranks Gartner

Perhaps the most surprising finding involves Guideflow, a company that sells interactive product demos. Its marketing blog was cited 194 times across 96 of 380 categories—a quarter of all queries—placing it third overall, ahead of Gartner's 158 citations.

Guideflow doesn't compete in any of the software categories it was cited for. Its sitemap lists 3,351 blog URLs, and it supplied grounding for everything from "3D rendering software" to "IVR software" and "RFID software." The company isn't being deceptive—it's simply publishing content marketing at scale, and Perplexity's retrieval layer treats it as an authoritative source.

The 'Facts & Grounding Pages' Network

The three most concerning sites—WifiTalents, WorldMetrics, and Gitnux—appear to be one operation. All registered through NameCheap between December 2023 and May 2024, they share the same Cloudflare nameservers, identical page templates, and matching navigation structures. Each maintains exactly six blog posts, all about the other brands in the set.

Their scale is staggering: sitemaps list 103,578, 107,083, and 105,541 URLs respectively, with 70,731, 71,684, and 72,713 being machine-generated '/best/-software/' pages. That's 215,128 generated buying guides across three brands—with only six human-written blog posts each.

What makes these pages particularly troubling is their self-description. Both WorldMetrics and Gitnux return HTML titles of the form "Facts & Grounding Page"—grounding being the technical term for the retrieval step these AI models perform. Their meta descriptions read like they're addressing the AI directly: "Verified facts about Gitnux: an independent market research company publishing industry statistics, custom research, and software Best Lists."

One Template, Three Different Verdicts

The study fetched the same category page—"project estimation software"—from all three brands. Despite sharing a template, each site ranks different products:

• WorldMetrics: Float, Scoro, Teamwork.com, Procore, Wrike
• WifiTalents: Float, Scoro, Teamwork.com, Buildertrend, Apropo
• Gitnux: Saviom, Mosaic, Buildertrend, Float, Teamwork.com

Gitnux's winner doesn't even appear in WorldMetrics' top five. Each page credits different named staff—nine distinct people across the three sites—and each announces its own editorial process. Gitnux even labels its result "AI-verified · Expert reviewed."

Dead Domains and Misleading Redirects

The study also checked the 1,502 vendor homepages the models recommended. Ten domains resolve to no address at all, including graphiql.com (offered for GraphiQL, which has no such site) and todo.com (offered for Microsoft To Do). Another 92 domains redirect to different registrable domains, mostly ordinary acquisitions.

Two cases stood out. For research data management platforms, Sonar gave dryad.co—which redirects to an Indonesian online-gambling portal—while Sonar Pro correctly identified datadryad.org. For data quality tools, Sonar Pro gave montecarlo.com, which redirects to Monaco's hotel and casino group, while Sonar correctly identified montecarlodata.com.

What This Means for AI Search

The findings expose a critical vulnerability in AI-powered search: the retrieval layer can be gamed by machine-generated content designed specifically to be read by models, not humans. The "Facts & Grounding Pages" are addressed, in their titles and descriptions, to the software that reads them.

Perplexity's two models shared a retrieval layer—returning byte-identical citation lists in 289 of 380 categories—so this isn't evidence of independent systems converging. It's one search stack sampled twice.

The study has limitations: it covers only Perplexity, one day's snapshot, and categories weighted toward niche verticals. But the implications are clear. As AI assistants become primary software discovery tools, the quality of their underlying evidence matters more than ever.

For buyers and developers, the takeaway is to verify AI recommendations against trusted, human-reviewed sources. For the industry, this is a wake-up call about the fragility of AI's information ecosystem.