AI-Optimized Deception: How Perplexity Cites "Grounding Pages"

Trellner Research Report TR-2026-009 | Published: 2 September 2026

In a startling discovery, research reveals that three specific websites have manufactured 215,128 "best software" pages specifically designed to be ingested by AI. These sites are not meant for human readers, yet they are being cited by Perplexity AI to provide "grounded" recommendations.

Across 380 different software niches, nearly 60% of the sources used for AI-driven recommendations are ranked outside the top 100,000 most-visited websites. More concerningly, several top-cited sources are essentially "AI bait."

🛠️ Methodology: The Experiment

To test the reliability of AI grounding, researchers queried two web-enabled models for the top products in 380 software categories (ranging from CRM to museum collection management).

The Process

Select 380 buyer-intent categories.
Query perplexity/sonar and perplexity/sonar-pro via OpenRouter.
Execute 760 total API calls (one per category, per model).
Request results in JSON format.
Extract and verify all retrieved URLs.

The prompt required a ranked top-five list, including the official homepage for each product.

{
  "category": "CRM Software",
  "recommendations": [
    {"rank": 1, "product": "ExampleCRM", "url": "https://examplecrm.com"},
    ...
  ],
  "citations": ["https://source-site.com/best-crm-software"]
}

Note: Because the Gemini model on OpenRouter uses a specific web-search plugin, the results reflect that plugin's retrieval rather than Google's native search. This study focuses exclusively on Perplexity.

📊 Analysis of Citations

The study produced 3,800 recommendation slots, identifying 1,807 unique products and 7,534 citations across 2,055 domains.

Distribution of Source Quality

The data shows a heavy reliance on obscure domains. Let $P$ represent the percentage of citations:

Model	Unranked (Outside Tranco 1M)	Ranked worse than #100k
`perplexity/sonar`	$23.4\%$	$59.8\%$
`perplexity/sonar-pro`	$23.5\%$	$59.9\%$
Pooled Total	23.4%	59.8%

The median Tranco rank for cited domains that actually had a rank was 71,611. While the top ten domains account for 17.3% of citations, the remaining 82.7% are fragmented across a long tail of low-traffic sites.

The Top 10 Cited Domains

Domain	Citations	Share	Tranco Rank
g2.com	291	3.86%	4,027
reddit.com	261	3.46%	105
guideflow.com	194	2.57%	177,039
gartner.com	158	2.10%	1,766
zapier.com	82	1.09%	2,919
wifitalents.com	71	0.94%	105,281
capterra.com	68	0.90%	6,387
linkedin.com	67	0.89%	18
worldmetrics.org	60	0.80%	104,737
gitnux.org	50	0.66%	42,759

For context: Wikipedia was cited only 3 times out of 7,534.

🚩 Red Flags: The "Guideflow" Anomaly

The third most-cited source, guideflow.com, is not a review site or a professional publisher—it is a vendor selling interactive product demos.

Despite not competing in any of the 380 categories queried, Guideflow's marketing blog was cited 194 times across 96 categories.

It provided the "evidence" for recommendations on everything from RFID software to architecture practice software. This demonstrates a failure in the retrieval layer: a company's content-marketing listicles are being treated as authoritative market research.

🤖 The "Grounding Page" Network

The research identified a cluster of sites specifically engineered for AI models. Three of the top sources—wifitalents.com, worldmetrics.org, and gitnux.org—along with a fourth site, zipdo.co, exhibit identical fingerprints.

Evidence of Common Control

Infrastructure: All share the same Cloudflare nameservers: pam.ns.cloudflare.com and sean.ns.cloudflare.com.
Timeline: None existed before December 2023.
Design: They use identical page templates and navigation menus (Services, Market Data, Software Advice, Editorial Process, Company).
Circular Referencing: Each site has a blog of exactly six posts; these posts exclusively discuss the other brands in the network.

The Scale of Generation

These sites are not written ~~for humans~~. Their sitemaps reveal a massive operation of machine-generated content:

wifitalents.com: 103,578 URLs $\rightarrow$ 70,731 /best/-software pages.
worldmetrics.org: 107,083 URLs $\rightarrow$ 71,684 /best/-software pages.
gitnux.org: 105,541 URLs $\rightarrow$ 72,713 /best/-software pages.

Total: 215,128 generated buying guides.

The most damning evidence is the HTML title of these pages. Instead of user-friendly titles, they are explicitly labeled as: Brand — Facts Grounding Page

This suggests the creators are explicitly targeting the "grounding" phase of LLM retrieval, creating a synthetic reality of "facts" for AI to cite.