Perplexity Cites the Sites That Made 215,128 Machine-Written Buying Guides
An independent audit of the sources behind AI product recommendations found six in ten pointing outside the 100,000 most-visited websites, and several of the most-cited sites describing their own homepages as machine-readable records.
Three websites end their homepage HTML title with the words "Facts & Grounding Page."
Grounding is not a word buyers use. It names a step inside a retrieval system, the part where the machine fetches documents to condition an answer on before it writes anything. A page titled that way is not addressed to a person. It is addressed to whatever software is reading it, and the meta description drops the pretense entirely: "Company, legal, methodology, and compliance details in one machine-readable record."
Two of those three sites, plus one more that shares their DNS and their page template, have published 215,128 generated best <category> software pages between them. All three of that second group land in the top ten sources Perplexity cited across a 380-category test.
Those counts come from TR-2026-009, published September 2 by Trellner Research, an independent firm that says it takes no money from the subjects it covers and accepts no sponsorship. The whole dataset ships under CC BY 4.0 with the scripts that produced every figure, which is the part that made me read it twice instead of once.
The question nobody asks the machine
When you ask an AI assistant which project management tool to buy, you evaluate the answer. You look at the five names, recognize two of them, and form an opinion.
What you do not evaluate is the pile of documents the model read to produce those five names. Web-grounded models retrieve first and write second, so the recommendation is a compression of an evidence base. Perplexity reports the URLs it retrieved, which is why Trellner chose it. Worth knowing that the choice was partly forced: the method notes say no native grounded Google route returning citations was available on the account, so Gemini got dropped rather than ignored.
The useful question is not whether the answer is good. It is what the answer is made of.
What the run actually did
On September 2, Trellner put 380 buyer-intent software categories to perplexity/sonar and perplexity/sonar-pro through OpenRouter, one prompt per category per model, 760 calls total. Every call asked for a ranked top five as JSON with each product's official homepage domain. All 760 returned a parseable answer. The categories were written before anyone saw a result and never revised, which matters more than it sounds like it does.
That produced 3,800 recommendation slots naming 1,807 distinct products, and 7,534 citations spanning 2,055 distinct domains. Every cited domain then got looked up in the Tranco daily list for September 1 and in the Wayback Machine, and all 1,502 vendor homepages the models supplied got fetched to see whether they still exist.
Here is the distribution that made the report travel. Of the 7,534 citations, 59.8% point at domains ranked worse than #100,000, and that figure includes the 23.4% pointing at domains not in Tranco's top million at all. The other way to read it: roughly 36 percent of citations land in the ranked-but-obscure band between #100,000 and #1,000,000, and another 23 percent land below the map entirely. Among the 5,768 citations that do point at a ranked domain, the median rank is #71,611.
The obvious guess is that a handful of famous review sites supply everything and the long tail is noise. That guess is wrong. The ten most-cited domains take only 17.3% of all citations. The long tail is not the leftovers. It is the evidence base.
For scale, Wikipedia was cited three times out of 7,534.
The third-largest source is a vendor's own blog
The domain ranked third by citation count, ahead of Gartner, is guideflow.com. It sells interactive product demos. It is not a review site, not a directory, not a publisher, and it competes in none of the 380 categories anyone asked about.
Its blog was cited 194 times across 96 of the 380 categories, a quarter of the whole set. Gartner got 158. Each citation is a different URL, roughly one per category, though six of the 96 are the Estonian-locale copy of a post. The sitemap lists 3,351 blog URLs covering 2,176 distinct posts. Guideflow supplied the grounding for "3D rendering software," "IVR software," "RFID software" and "architecture practice software" alike.
Nothing here is a scam. Guideflow runs a large content-marketing blog, the way thousands of companies do, and probably nobody there decided to become the third-largest evidence base for software purchasing decisions. The retrieval layer decided that. A vendor writing listicles about markets it does not operate in got promoted into evidence because the pages existed, matched the query shape, and had nothing louder sitting on top of them.
Three more sites in the top ten, one template between them
Ranked sixth, ninth and tenth are wifitalents.com, worldmetrics.org and gitnux.org. They supplied 181 citations between them, 2.4% of the total, across 41 of the 380 categories.
They appear to be one operation, and Trellner calls the evidence strong but circumstantial. All three registered through NameCheap between December 2023 and May 2024. All three delegate DNS to the same pair of Cloudflare nameservers. All three run the same template with the same navigation: Services, Market Data, Software Advice, Editorial Process, Company. Each keeps a blog of exactly six posts, and all eighteen posts are about other brands in the set, including a fourth, zipdo.co, which sits on the same nameservers and gives its own homepage that same "Facts & Grounding Page" title. What the report does not have is an owner. None of the four names one.
Their sitemaps list 103,578, 107,083 and 105,541 URLs, of which 70,731, 71,684 and 72,713 are /best/<something>-software/ pages. Trellner's line about it is the best sentence in the report: there are not 215,128 software categories.
Then the researchers did the thing I wish more people did. They pulled the same category page, "project estimation software," from all three brands. Worldmetrics ranks Float, Scoro, Teamwork.com, Procore, Wrike. WifiTalents ranks Float, Scoro, Teamwork.com, Buildertrend, Apropo. Gitnux ranks Saviom, Mosaic, Buildertrend, Float, Teamwork.com. Gitnux's winner does not appear in Worldmetrics' top five at all.
Each page credits three named staff, so nine distinct human beings answered one question three different ways. Each page announces an editorial process. Gitnux labels its result "AI-verified · Expert reviewed." And all three carry an unrendered template variable sitting in the byline, reading "Within the next 26 days" on two of them and "Within the next 40 days" on the third.
One more finding worth carrying around. Of the 1,502 vendor homepages the models handed back, 17 came back unreachable, though Trellner counts four of those as alive and merely unresponsive to an automated request, nasdaq.com and solidworks.com among them. That leaves 13 genuinely dead or gone. Two of the redirects are the memorable part: dryad.co, offered as the home of the Dryad research repository, lands on an Indonesian online-gambling portal, and montecarlo.com, offered for the data-quality company Monte Carlo, lands on the Monaco hotel and casino group. In both cases the two Perplexity tiers disagreed with each other, and one of them was right.
Putting this into practice
Start with the cheapest possible habit. When an AI assistant recommends a product, open three of its sources. Not all of them, three. Perplexity puts them right there. You will know within about ninety seconds whether you are reading G2 and a vendor changelog or a /best/ page with an editorial-process link and a six-post blog.
The second habit takes about as long. Paste a suspicious domain into the Tranco list. A rank in the hundreds of thousands is not proof of anything, and Trellner is careful to say popularity is not quality. It does tell you that whatever you are reading was not vetted by traffic.
If you build agents rather than just use them, the operational version of this is one logging change. Store the citation set your agent retrieved next to the answer it produced, and put both in front of whoever approves the action. Most agent frameworks shipping this year have an approval step, and most of those steps show a human the proposed action and nothing about the reasoning that produced it. The reviewer approves a conclusion and never sees the premises. That is half a review.
And if you publish, read the report as a description of what the retrieval layer currently rewards, then decide for yourself what you want to do about it. I am not going to tell you to go make 70,000 pages.
What this does not show
The report's limitations section is unusually good, and repeating it is not hedging.
It covers Perplexity only. ChatGPT, Gemini, Copilot and Google's AI Mode were not measured, and there is no reason to assume their retrieval mixes match.
The two Perplexity tiers are not two independent measurements. They returned byte-identical citation lists in 289 of the 380 categories and their URL sets overlap at a Jaccard of 0.898, so this is one search stack sampled twice. Their agreement on the top pick in 290 of 380 categories is a fact about shared retrieval, not evidence that independent systems converge.
The 380 categories are the authors' own construction, weighted toward niche verticals, and a niche-heavy list surfaces more long-tail sources than a list of common queries would. One prompt wording, one run per category, no repeat sampling, one day's snapshot of an index that changes. An earlier pilot suggested the product shortlist moves noticeably when "best" is swapped for "most popular" while the citation mix moves much less, but this run did not measure that.
Citations are grouped by registrable domain, so a link to a vendor's documentation subdomain counts toward the vendor's domain. That affects any concentration argument, including the 17.3% figure.
Every page fetch went out through a rotating datacentre proxy under a named research user-agent, so what those sites returned to Trellner is not necessarily what they return to a retrieval crawler or to your browser. The title tags quoted above were read once, on September 2.
The largest gap is one the authors state plainly: they did not test whether removing these sources changes the answers. Guideflow and the three Best List brands may well name reasonable products. What got measured is which documents the evidence base is made of, not whether a better evidence base would produce a different top five. Somebody should run that experiment, and until they do, the honest version of this story is about provenance and not about accuracy.
The part you can act on
I do not think the takeaway is that AI shopping advice is worthless. It is that the confidence you feel reading a grounded answer is borrowed from sources you have never looked at, and looking at them costs about ninety seconds.
Several sites decided the most useful thing they could put in a title tag was the name of a retrieval step. That is a real signal about who the modern web is being written for. Go open three citations on the next answer you get and see who wrote them.
Sources: Trellner Research, TR-2026-009: "Three sites made 215,128 'best software' pages for AI. Perplexity cites them", published 2 September 2026, with the dataset, method notes and scripts released under CC BY 4.0 at /data/manufactured-sources-behind-ai-recommendations/; Tranco list; Trellner Research, About.