On 2 September 2026, Trellner Research asked two of Perplexity's search models to name the best products in 380 software categories, and kept every source the answers cited. Its report counts 7,534 citations, and the third most-cited domain in them is guideflow.com. The report is specific about what got cited. Guideflow's blog was, in the report's words, "cited 194 times across 96 of our 380 categories", and Trellner counts 96 distinct blog pages behind those citations, one per category. The domain gartner.com came fourth, on 158.

Guideflow's homepage describes it as an interactive demo platform, and Trellner notes it "competes in none of the categories we asked about". That includes yacht management software and RFID tracking. Its blog ranks those anyway.

What the report measured

The study is about what Trellner calls grounding, the outside evidence a search-backed system pulls in to support an answer. What it measures are the citations the answers carried: where those domains rank, how old they are, and who publishes them. Trellner says it sent one prompt per category to perplexity/sonar and perplexity/sonar-pro through OpenRouter, a service that routes requests to model providers, for 760 calls in all. It chose those two because, in its words, "both models report the URLs they retrieved". Those citation lists are what the study measures. They are not a full trace of everything the retriever looked at. Every cited domain was then looked up in Tranco, a research ranking of the top million domains built by combining several popularity lists.

The finding that got picked up: Trellner reports 59.8 per cent of the 7,534 citations point at domains Tranco ranks worse than 100,000, and 23.4 per cent at domains it does not rank at all. Wikipedia, the report says, was cited three times.

The report's own title points somewhere else. Trellner names three sites, worldmetrics.org, gitnux.org and wifitalents.com, registered through NameCheap between December 2023 and May 2024, delegating to the same pair of Cloudflare nameservers, running the same template, with 215,128 generated "best category" pages between them against six blog posts each. Two of them, the report says, titled their homepage "Facts and Grounding Page" and described themselves in their meta description as "one machine-readable record". That is a meta description written in the vocabulary of retrieval rather than the vocabulary of buyers.

The parts I could check myself hold up. On 8 September 2026 I opened gitnux.org's sitemap, followed the two best-list sitemaps it points to, and counted 50,000 and 21,988 listed pages. Its blog sitemap lists six. I fetched the "project estimation software" page from both Gitnux and WifiTalents. Gitnux ranks Saviom, Mosaic, Buildertrend, Float and Teamwork.com; WifiTalents ranks Float, Scoro, Teamwork.com, Buildertrend and Apropo, which is what Trellner reported six days earlier. Both pages carry the same stray placeholder in the byline, which Trellner calls an unrendered template variable: "Within the next 40 days" on one, "Within the next 26 days" on the other. I found both stamped "Updated August 22, 2026". Two sites that each describe themselves as an independent market research company, carrying the same update stamp for the same obscure category, disagreeing about the winner.

Why the generated pages are the easy part

Those three sites are the satisfying version, because whatever the relationship between them, they are three domains you can name. Trellner is careful about how far the evidence goes: it says common control is inferred from shared infrastructure and an identical template, and that "we do not know who operates them; none of the three names an owner". However they are related, they amount to three names that can be written on a list.

Guideflow is the harder case, because the report finds nothing deceptive in what it publishes.

Guideflow sells a real product, and the report is fair to it: "Guideflow publishes a large content-marketing blog, as thousands of companies do." On 8 September 2026 I opened its sitemap, which the Internet Archive captured the same day, and counted 3,348 listed blog pages, 2,174 of them outside the Estonian-language section, close to the 3,351 and 2,176 Trellner reports for 2 September. I read four of those posts on 8 September 2026. The yacht management page opens on Seahub, YACHTWAVE and Floatist, with a paragraph on each about who it suits; the RFID page opens on Avancir, TagMatiks and Zebra MotionWorks. They read like buying guides, because that is what they are. The report's verdict on them is one sentence: "Nothing here is deceptive."

So a vendor's marketing blog became the third most-cited domain in that run, ahead of Gartner, and neither the report nor my own reading of those pages turns up any deception on those pages. That is the uncomfortable version, because the generated pages were not the only route into that citation list. Guideflow is in it with a content-marketing blog of the sort thousands of companies already run.

What retrieval can score, and what it cannot

Perplexity's public API reference documents the search options a caller can set, not the ranking that runs behind them, so I cannot tell you why any particular page was picked. What is on the record is that they were picked: Trellner reports 194 citations to Guideflow's blog. The one I found on 8 September 2026 is titled "6 best yacht management software for 2026".

Independence is harder to compute. Sometimes it is legible in the document, in a disclosure line or an affiliate tag. Often it is not, because it is really a relation between the author and the question, and the same publisher can sit on either side of it depending on what was asked. The caller-facing controls are coarse. The options Perplexity documents are search_mode, which selects the corpus the search runs over (web, academic or SEC filings), and search_domain_filter, which limits results to a list of domains the caller supplies. Both are useful. Neither expresses that distinction. Guideflow has no product in yacht software and a product in product adoption software, on the same domain, on the same day.

I went and looked at that second case. On its "product adoption software" page, a category where Guideflow competes, Guideflow ranks itself first, ahead of Pendo. The page says its order "reflects a combination of market presence, feature breadth, and editorial judgment", lists Guideflow's pricing tiers and a G2 rating of 5.0 out of 5, offers a "Start your journey with Guideflow today" link, and carries no sentence disclosing that the publisher and the top-ranked product are the same company. It does include a section headed "Honest trade-offs".

Same site, same blog template. In one of those categories the publisher is also a contender. Someone reading that page on guideflow.com sees the logo, and can work out whose house they are in. A citation carries a link and a domain. It does not carry the fact that the publisher sells one of the things on the list, and a reader who does not already know that would have to go and find out.

The case against me

The strongest objection is Trellner's own, and to its credit it is in the report: "We have not shown that any of this changes the answers." The prompts were not re-run with these sources removed. Guideflow's yacht page does name real yacht software. If the shortlist would be the same either way, then the provenance of the evidence is an aesthetic complaint about a system that works.

Two more, also conceded in the report. The 380 categories are Trellner's own construction rather than a sample of what buyers ask, and they lean towards niche verticals such as museum collection management software, where long-tail sources are most of what exists. And the two models are not two independent measurements. The report puts it bluntly: the Perplexity tiers "share a retrieval layer and should be read as one search stack sampled twice". One stack, one day, one product.

Each of those is fair. They cap what the study proves, and they are why nothing here should be read as a claim about ChatGPT or Google's AI Mode, neither of which Trellner measured.

What they do not touch is the argument I am actually making, which is about what the visible surface can express. In the surface Perplexity documents, and in the pages themselves, I can find nothing that represents whether an author has a stake in the answer. That is an argument from what is visible, and Perplexity may well run a classifier I cannot see; if it does, this is the moment to say so. So when a shortlist comes back looking reasonable, I cannot point to the part of the pipeline that made it so. Guideflow has no product in yacht software, and its yacht page reads like it. It has a product in product adoption software, and its page there puts Guideflow on top. From the outside I cannot see what would make a retriever score those two pages differently, and the second is the page a vendor has an obvious incentive to publish about its own category.

One thing web ranking has is the link graph. It is gameable and it was never a measure of independence, but a link to you has to sit on some other page, and often that page is not yours. Nothing in Perplexity's public API reference exposes an equivalent for grounded answers. Whether something inside does, I have no way to tell, and I would like to be told.

The bet

Something moved this week and I cannot tell you what. On 8 September 2026, worldmetrics.org and zipdo.co answered every request I made to their homepages with an HTTP 500 error reading "Internal Server Error". So did gitnux.org's homepage, while its buying guide at gitnux.org/best/project-estimation-software/ served a full page. WifiTalents was healthy, and its homepage title had changed to "Original data, independently audited" from the wording preserved in the Internet Archive's capture of 3 September 2026. Whether that is a deployment failure or a rebrand, I have no way to tell from outside.

None of that changes the argument. So, the bet, and it is a bet about what gets documented in public rather than about hidden machinery. Within twelve months, one of Perplexity, OpenAI or Google will publicly document a retrieval signal that treats a page ranking its own publisher differently from one that does not: a conflict flag, a publisher register, or a stated demotion. If none of the three does, I will want to hear what is doing that job instead. A vendor can publish the ranked listicle for its own category using staff it already pays. The page will be branded and obvious to anyone who lands on it. The citation is what travels.

Trellner published the dataset and the scripts under CC BY 4.0. The cheapest test of the objection above is to re-run Trellner's 380 questions with guideflow.com excluded and see whether the shortlists move. I would like to see that number. Trellner did not measure it and neither did I, and until somebody does, "the recommendations are probably fine" and "the recommendations are skewed" are both guesses.