Which AI models did Trellner test?

Trellner Research tested two Perplexity models, sonar and sonar-pro, called through OpenRouter. The report does not cover ChatGPT, Gemini, Claude or Copilot, so the findings describe Perplexity's grounding behaviour on these queries rather than AI search in general. The published scripts can be pointed at another engine to repeat the test.

Can I download the data behind the report?

Yes. Trellner Research published the full dataset under CC BY 4.0: answers.csv and answers_raw.jsonl with the raw model responses, plus citations.csv, cited_domains.csv and vendor_domains.csv. The release also includes seven Python analysis scripts, around 20 saved evidence pages, a README and a METHOD.md documenting the method.

Which sites does the report name?

The Trellner report names worldmetrics.org, gitnux.org and wifitalents.com as three apparently related sites that together published 215,128 machine-generated 'best category' pages. It also flags guideflow.com, a vendor marketing blog that placed third overall with 194 citations even though it operates in none of the 380 surveyed categories.

How obscure were the cited domains?

Among cited domains that Tranco ranks at all, the median rank in Trellner's data was 71,611, and 23.4% of citations went to domains missing from Tranco's top million entirely. Combined with the 59.8% ranked worse than #100,000, most of the sources behind these Perplexity answers sit far outside the popular web.

Trellner Research · 2026-09-02 · major

215,128 machine-made 'best software' pages — and Perplexity cites them

Trellner Research ran 380 software categories through Perplexity's sonar models and found 59.8% of 7,534 citations went to domains ranked worse than #100,000. Three linked sites had published 215,128 'best software' pages.

Trellner Research card for the report on manufactured sources behind AI recommendations

A citation audit of Perplexity's software recommendations, published with the full dataset and scripts.

Quick facts

Publisher	Trellner Research (report TR-2026-009)
Models tested	perplexity/sonar and sonar-pro
Scope	380 software categories, 760 API calls
Citations analysed	7,534
Obscure sources	59.8% ranked worse than #100,000 on Tranco
Machine-made pages	215,128 across three linked sites
Data licence	CC BY 4.0

What is it?

Trellner Research audited where an AI answer engine gets its software recommendations. Report TR-2026-009 put 380 software categories through Perplexity's sonar and sonar-pro models and logged every source cited. 59.8% of the 7,534 citations pointed at domains ranked worse than #100,000 on the Tranco popularity list, and 23.4% at domains outside the top million entirely.

How does it work?

Each category was queried twice, once per model, through OpenRouter — 760 calls in total. Every cited domain was then looked up in Tranco and the Wayback Machine, and the vendors named were checked to see if they were still live. Three of the most-cited sites turned out to share Cloudflare nameservers, identical templates and registration dates between December 2023 and May 2024: worldmetrics.org, gitnux.org and wifitalents.com.

Why does it matter?

Answer engines are becoming a discovery channel for software, and this report puts numbers on what they read. The pages behind many of those citations describe themselves in their HTML titles as a 'Facts & Grounding Page' — wording written for machines, not people. Trellner released the dataset, the analysis scripts and a METHOD.md under CC BY 4.0, so anyone can re-run the same check against another engine.

Who is it for?

AI search researchers and marketers

Frequently asked questions

Which AI models did Trellner test?: Trellner Research tested two Perplexity models, sonar and sonar-pro, called through OpenRouter. The report does not cover ChatGPT, Gemini, Claude or Copilot, so the findings describe Perplexity's grounding behaviour on these queries rather than AI search in general. The published scripts can be pointed at another engine to repeat the test.
Can I download the data behind the report?: Yes. Trellner Research published the full dataset under CC BY 4.0: answers.csv and answers_raw.jsonl with the raw model responses, plus citations.csv, cited_domains.csv and vendor_domains.csv. The release also includes seven Python analysis scripts, around 20 saved evidence pages, a README and a METHOD.md documenting the method.
Which sites does the report name?: The Trellner report names worldmetrics.org, gitnux.org and wifitalents.com as three apparently related sites that together published 215,128 machine-generated 'best category' pages. It also flags guideflow.com, a vendor marketing blog that placed third overall with 194 citations even though it operates in none of the 380 surveyed categories.
How obscure were the cited domains?: Among cited domains that Tranco ranks at all, the median rank in Trellner's data was 71,611, and 23.4% of citations went to domains missing from Tranco's top million entirely. Combined with the 59.8% ranked worse than #100,000, most of the sources behind these Perplexity answers sit far outside the popular web.

Try it

https://trellner.com/data/manufactured-sources-behind-ai-recommendations/