AI Product Recommendations Are Being Farmed — 215,128 Pages Deep
TL;DR: New research found that when Perplexity is asked to recommend software, nearly 60% of its citations point to sites ranked below the top 100,000 on the web — and that three related sites have published 215,128 machine-generated “best software” guides between them. We checked those sites ourselves. Two of them have homepages literally titled “Facts & Grounding Page” — a phrase aimed at AI systems, not people. If you lean on AI product recommendations to pick your tools, this is worth ten minutes of your attention.
What the research found
On 2 September 2026, Trellner Research published a report on where AI product recommendations actually come from. The method was simple: ask Perplexity’s `sonar` and `sonar-pro` models to name the top five products across 380 software categories, then examine every source they cited.
That produced 7,534 citations across 2,055 domains. The findings:
| What they measured | Result |
|---|---|
| Citations to domains ranked below #100,000 | 59.8% |
| Citations to domains outside the top million entirely | 23.4% |
| Median traffic rank of cited domains | #71,611 |
| Recommended homepages that were dead or redirected | 109 of 1,502 (7.2%) |
Three sites — gitnux.org, worldmetrics.org and wifitalents.com — accounted for 181 of those citations across 41 categories. Between them, the report counts 215,128 generated buying guides. All three were registered between December 2023 and May 2024 and share templates and infrastructure.
Two redirects the researchers flagged are worth repeating, because they show what “cited source” can mean in practice: one recommended homepage redirected to an Indonesian gambling site, another to a Monaco casino.
What we found when we checked ourselves
A single research report is one source, and we don’t publish claims about named companies on one source. So rather than look for someone else’s write-up, we went and checked the central claims directly. Everything in this section we verified first-hand on 3 September.
The homepage titles are real, and they’re remarkable. The HTML title of gitnux.org is “Gitnux — Facts & Grounding Page”. Worldmetrics.org: “Worldmetrics — Facts & Grounding Page”. Grounding is the technical term for the documents a language model retrieves to base an answer on. These are homepages named after their function in an AI pipeline.
The templates really are shared. Both sites serve sitemaps with identical structures — the same shard filenames in the same arrangement. Their meta descriptions are near-identical boilerplate describing “an independent market research company publishing industry statistics, custom research, and software Best Lists”. Wifitalents.com uses a different title but almost exactly the same description.
The scale checks out. The first sitemap shard on gitnux.org alone lists 50,000 `/best/` URLs, with a second shard beside it. Of those 50,000, more than 1,200 are AI or software-building categories.
And one of them is our category. We opened “Best Drag And Drop Website Builder Software” — a live page, updated 18 August 2026, ranking Hostinger first, GoDaddy second, Weebly third. It carries three human bylines: written by one person, edited by a second, fact-checked by a third. It is badged “AI-verified · Expert reviewed” and labelled a 30-minute read.
It also publishes its own methodology, and step three is this:
Synthetic User Modeling — AI persona simulations modeled how different user types would experience each tool across common use cases and workflows.
That is a page carrying three named humans and an “Expert reviewed” badge, openly stating that its user experiences are simulated. One more tell: the byline block ends “Published February 11, 2026 · Updated August 18, 2026 · Within the next 30 days” — an unfilled template variable, left in the page.
Why it matters if you don’t code
If you can’t evaluate ten AI coding tools yourself, you ask something that can. That is not a failure on your part — it is the sensible move, and it is exactly how most of our readers work.
The problem is that AI product recommendations are only as good as what the model retrieved, and retrieval rewards a very specific thing: pages that look comprehensive, structured and authoritative to a machine. Generating 70,000 of those is cheap. Earning the same coverage honestly is not.
So the failure mode behind bad AI product recommendations isn’t an AI inventing a product. It’s an AI faithfully summarising a page that was manufactured to be summarised — and handing you a ranked list with no visible difference between a genuinely tested recommendation and a synthesised one.
Two things follow that are worth internalising:
- A confident ranked list is not evidence. “Top 10 Best X of 2026” is a format, not a finding. It costs nothing to produce at scale.
- Human bylines and “expert reviewed” badges are not evidence either. We just read a page with three named humans and a fact-checker that describes its own user research as persona simulation.
How to sanity-check AI product recommendations
You don’t need to stop using AI product recommendations. You need a cheap habit for checking them.
- Ask for the sources, then actually open one. Every AI search tool will show you what it cited. If the citations are sites you’ve never heard of with names like worldmetrics or gitnux, treat the ranking as unverified.
- Look for specifics only a user would know. Real experience produces oddly precise details — what breaks at 400 rows, which export is broken, what the support reply was. Generated guides produce balanced pros and cons that could apply to anything.
- Check whether the writer says what they actually did. “We tested this for a month” is a claim you can weigh. “AI persona simulations modeled how different user types would experience each tool” is a disclosure that no one used it.
- Prefer sources with something to lose. A named person with a reputation, a forum where users will contradict a bad claim, or a site that publishes its mistakes.
- Ask a differently-shaped question. Instead of “what’s the best X”, try “what do people complain about with X after three months”. Complaint-shaped questions retrieve different, harder-to-fake sources.
Who should care (and who shouldn’t)
- Choosing your first AI building tool: the group most exposed to bad AI product recommendations, because you have no baseline to notice a wrong answer. Use the checks above before committing money.
- Building a revenue app: you’re likely past tool selection, but the same problem hits every “best X for Y” question you’ll ask along the way.
- Already settled on a tool that works: honestly, ignore this. Don’t let it talk you out of something that’s working.
- Publishing content yourself: read the Trellner report properly. It is a clear picture of what you’re now competing against for citation.
Our take, including the awkward part
We should be straight about our own position here, because a site that publishes tool recommendations telling you to distrust tool recommendations owes you that.
We publish “best tool” content, we use affiliate links, and we benefit when AI product recommendations point here. The page we pulled apart above ranks Hostinger first; Hostinger hosts this site and is a partner. We are, structurally, in the same category as the thing this research is about. Pretending otherwise would be worse than the pages we’re criticising.
What we think actually separates the two is checkable, and you should hold us to it: whether a specific person is accountable for the claim, whether the site says plainly what it did and didn’t test, whether the recommendation ever says “don’t buy this” or “this isn’t for you”, and whether it discloses what it earns. Those are the tests we’d apply to us. This piece has no affiliate links in it at all, because there was nothing honest to sell you.
The wider point is one we keep arriving at from different directions. The model isn’t the thing that decides your answer — the material around it is. We’ve made that argument about which models your tools quietly route you to and about capabilities hiding inside products you already pay for. This is the same lesson pointed at research: the quality of what an AI tells you is set by what it could find, and someone has worked out that what it can find is purchasable.
Want a tool recommendation with the reasoning shown? Our 60-second Vibe Coding Tool Finder quiz asks what you’re building and tells you why it picked what it picked →
FAQ
Are AI product recommendations trustworthy?
Treat them as a starting point, not an answer. Trellner’s research found 59.8% of Perplexity’s software citations came from domains ranked below #100,000, and identified 215,128 machine-generated buying guides across three related sites. The model isn’t lying — it is faithfully summarising sources that were built to be summarised.
How can I tell if a “best software” page is machine-generated?
Look for specifics only a real user would know, check whether the methodology says what was actually tested, and be sceptical of enormous category coverage. One site we checked lists 50,000 “best” pages in a single sitemap shard, and its own stated method includes “Synthetic User Modeling” — AI persona simulations rather than real users.
Does this affect ChatGPT and Gemini too?
Unknown, and the report is explicit about that limit: it measured Perplexity only, on a single day, across 380 categories that may not match real buyer questions. The incentive to manufacture citable pages applies to any AI that retrieves from the open web, but nobody has yet published the equivalent measurement for other assistants.
