We Counted the Pages in Three AI Content Farms and Found 293,918 of Them

On September 2, Trellner Research published a report (TR-2026-009) with a number in it that is easy to repeat and hard to check: three coordinated websites have published 215,128 machine-generated "best software" pages, and Perplexity cites them.

We do not take numbers like that on faith, so we spent an afternoon checking the checkable parts. Most of it holds. The page count is, if anything, understated: we counted 293,918 URLs across the three sitemaps.

What the report actually measured

The methodology is narrower than the headline suggests, and the authors say so. Researchers queried two web-grounded Perplexity models, perplexity/sonar and perplexity/sonar-pro, through OpenRouter with 380 software category questions, asking for a ranked top five as JSON. They kept every URL that came back: 7,534 citations in total. Then they cross-referenced those domains against the Tranco top-1M popularity list.

The findings, as reported:

• 59.8% of citations pointed at domains ranked worse than #100,000 globally
• 23.4% pointed at domains not in the top million at all
• The median Tranco rank among cited domains was 71,611
• The most-cited single domain was g2.com at 291 citations (3.86%), followed by reddit.com at 261

The limitations are stated plainly in the report, and they matter. This is a single-day snapshot. The 380 categories were constructed by the researchers rather than drawn from real buyer queries. Requests came from datacenter proxies. Only Perplexity was tested, so nothing here says anything about ChatGPT, Gemini, or Copilot. And as the authors note, "Tranco rank is a popularity measure, not a quality measure", which is the honest caveat: an obscure domain is not automatically a bad source.

What we verified ourselves

The claim we could test independently is the coordination one. The report names three domains: wifitalents.com, worldmetrics.org, and gitnux.org. We started with DNS.

$ dig +short NS wifitalents.com
pam.ns.cloudflare.com.
sean.ns.cloudflare.com.

$ dig +short NS worldmetrics.org
sean.ns.cloudflare.com.
pam.ns.cloudflare.com.

$ dig +short NS gitnux.org
pam.ns.cloudflare.com.
sean.ns.cloudflare.com.

The identical Cloudflare nameserver pair is suggestive rather than conclusive on its own, since Cloudflare assigns pairs from a pool. The sitemaps are the part that is not ambiguous. All three sites shard their sitemap into exactly the same eight files, with the same names, in nearly the same order:

/sitemap/shards/best_list.xml
/sitemap/shards/best_list-2.xml
/sitemap/shards/blog.xml
/sitemap/shards/statistic.xml
/sitemap/shards/alternative.xml
/sitemap/shards/best_list_service.xml
/sitemap/shards/alternatives.xml
/sitemap/shards/authors.xml

Then we counted the <loc> entries in every shard on all three domains.

Shard	wifitalents.com	worldmetrics.org	gitnux.org
best_list	50,000	25,443	50,000
best_list-2	22,713	22,698	21,988
best_list_service	15,872	15,798	15,853
statistic	16,429	16,526	19,020
alternative	464	461	462
blog	6	6	6
alternatives	4	4	6
authors	53	53	53
Total URLs	105,541	80,989	107,388

293,918 URLs in total. The best-list family alone (the three shards that hold buying guides) comes to 240,365. The report's 215,128 sits between two reasonable definitions of what counts as a "best software" page, so it is the right order of magnitude and probably now stale on the low side.

Six blog posts each. Fifty-three authors each. Those two rows do more work than the DNS lookup does.

The 159 authors who do not exist

Every one of the three sites advertises 53 author profiles in its sitemap, under a path spelled /autoren/. That is German for authors, on three English-language sites, which tells you something about the template and probably about the operator.

We fetched them. All three sites return HTTP 404 for the author pages listed in their own sitemaps:

$ curl -sLo /dev/null -w "%{http_code}" https://worldmetrics.org/autoren/amara-osei/
404
$ curl -sLo /dev/null -w "%{http_code}" https://gitnux.org/autoren/abigail-foster/
404
$ curl -sLo /dev/null -w "%{http_code}" https://wifitalents.com/autoren/ahmed-hassan/
404

159 advertised bylines, none of which resolve. The names are still printed on the articles. We opened gitnux.org/best/soccer-scheduling-software/, one of the long-tail pages, and it carries a full editorial apparatus: Henrik Dahl as author, Catherine Wu as editor, Jonathan Hale as fact-checker, published February 11 2026, updated August 23 2026. Catherine Wu is one of the 53 gitnux author slugs. Her page 404s.

Two of the three sites go further and say who the audience is directly in the HTML title. worldmetrics.org and gitnux.org both render as "Facts & Grounding Page". Grounding is the retrieval step a web-connected model performs before answering. The title is addressed to the crawler.

One correction to the report, in its favour and against ours: it describes all three homepages carrying that title, and wifitalents.com now titles itself "Original data, independently audited | WifiTalents". Either the site changed since the report ran or the description generalised. We cannot tell which from the outside.

What the thread got right that the report did not test

The Hacker News discussion reached 461 points and 220 comments, and the top thread went somewhere the report explicitly declined to go: whether models prefer this stuff on purpose.

xpct opened with the mechanism: "there were some papers which suggested that LLMs favor LLM-generated passages over human written ones", adding that Claude consistently picks its own earlier code over a human refactor of it. abhgh supplied the term and the citation, pointing at self-preference bias and the G-Eval paper. dspillett gave the least mystical explanation: a model asked to judge the best response to an input will favour whatever most closely matches what it would have produced.

keeda drew the conclusion the report's methodology cannot support but everyone in the thread reached anyway: "The easiest way to do AI SEO is to generate all your content with AI". That is a hypothesis, not a finding. The report tested which domains get cited, not why.

The best comment nobody upvoted came from dgellow, describing a failure mode that needs no content farm at all: a model searches, finds a Reddit comment containing something casually wrong that any human would have dismissed, and expands it into confident prose. Content farms are the industrialised version of a weakness that already existed.

The most useful dissent was thephyber, arguing that one bad recommendation does not indict a whole category, and that users have to cross-reference. Fair. It also does not scale: the entire pitch of a grounded answer engine is that it does the cross-referencing.

Our read

The report's Perplexity numbers are one lab's single-day snapshot of 380 constructed queries, and should be treated that way. The coordination finding is different, because it does not depend on the queries at all. Three domains, one nameserver pair, one sitemap template, 293,918 URLs, 159 bylines that 404, and a page that admits its testing was done by AI personas. Anyone can reproduce that in twenty minutes with curl and dig, which is the strongest thing about it.

A page titled "Facts & Grounding Page" is not trying to rank. It is trying to be quoted.

What would change our mind: the report does not test whether removing these three domains changes Perplexity's actual recommendations, and says so. It is possible they are cited constantly and influence nothing, because the products they name are the obvious ones anyway. That would be the reassuring finding, and nobody has run it.

We run a site whose entire credibility argument rests on doing the work and saying when we did not, so we will say the uncomfortable part out loud. The 240,365 buying guides and the 53 fabricated bylines per site are the mature version of what a content operation looks like when nobody checks. The only durable defence is the thing that cannot be generated: a number somebody actually measured, and a name attached to it that resolves.