Most teams building for AI vendor recommendations are optimising against an imagined shortlist. They picture a buyer asking an assistant for the best software in a category, and they picture the assistant reaching for G2, Gartner, a few respected trade publications, and the vendors' own sites. That picture is worth testing, because somebody finally did, and the answer is stranger than the assumption.
Trellner Research ran 380 buyer-intent software categories through two Perplexity models on 2 September 2026, asking each for a ranked top five. The 760 calls produced 3,800 recommendation slots naming 1,807 distinct products, grounded in 7,534 citations across 2,055 distinct domains. The citation list is the interesting artefact, not the product rankings, because it is a direct read on which pages an answer engine considered load bearing for a commercial question.
That last pairing is the one to sit with. A software vendor's marketing blog was cited more often than the best known analyst firm in enterprise technology, in a test about which software to buy. Nothing about that is a scandal. It is a clean signal about what the retrieval layer is actually rewarding, and it is not brand equity.
What actually grounds AI vendor recommendations
The top of the cited-domain list looks reassuringly familiar for about four rows, and then it stops. G2 leads, Reddit sits behind it, and both are where you would expect given how much of B2B software discourse lives in review profiles and community threads. Then a comparison blog from a vendor in the sales enablement space slots in above Gartner, and below all of that sits a very long tail of sites the average practitioner could not name.
| DOMAIN | CITATIONS | SHARE OF ALL CITATIONS | WHAT IT IS |
|---|---|---|---|
| g2.com | 291 | 3.86% | Review marketplace with structured category pages and comparison inventory at scale |
| reddit.com | 261 | 3.46% | Community threads, still a heavy source for Perplexity even after a turbulent year for its crawl access |
| guideflow.com | 194 | 2.57% | A demo automation vendor whose blog publishes comparison and best-of posts across adjacent categories |
| gartner.com | 158 | 2.10% | Analyst research, much of it behind registration or paywalls that limit what a crawler can use |
Look at the share column rather than the raw counts. The single most cited domain in the entire run accounted for under 4% of citations. There is no dominant source. The corpus behind AI vendor recommendations in software is flat and wide, which means no individual placement moves the needle much and the aggregate shape of a category's published inventory matters more than any one page in it.
The Gartner result has an obvious mechanical explanation worth stating plainly, because it is the useful lesson rather than a knock on the firm. Analyst research is frequently gated. A crawler that cannot read the argument cannot ground an answer in it, however authoritative the argument is. Guideflow's comparison posts are free, indexed, and structured like the question being asked. That is the whole of the advantage. We made the same case when we argued that AI citations behave like a bibliography rather than a brainstorm: the engine assembles from what it can retrieve and verify, and reputation that lives behind a form is invisible to it.
Three domains, 215,128 buying guides
The part of the Trellner report that will get quoted is the farm finding, and it deserves the attention. Three apparently connected domains, none of which existed before late 2023, have between them published 215,128 pages shaped as software buying guides. Two of the three carry the HTML title Facts and Grounding Page on their homepage, which is about as close to stating the business model out loud as a website gets.
| DOMAIN | TOTAL URLS | BUYING GUIDE PAGES | FIRST REGISTERED |
|---|---|---|---|
| worldmetrics.org | 103,578 | 70,731 under a best-of software path | Between December 2023 and May 2024 |
| gitnux.org | 107,083 | 71,684 under a best-of software path | Between December 2023 and May 2024 |
| wifitalents.com | 105,541 | 72,713 under a best-of software path | Between December 2023 and May 2024 |
| Combined | 316,202 | 215,128 generated buying guides | All three inside the same six month window |
Trellner notes that the set includes guides for software categories that do not exist, which is the tell. Nobody researched 215,128 markets. A template was pointed at a category list and run until it stopped. The pages were not built to be read by people, and their traffic profile says nobody reads them. They were built to be retrieved.
“Two of the three domains title their homepage as a grounding page. The content was never for a reader. It was inventory positioned at the retrieval layer, and it worked well enough to show up in a commercial answer.”
The honest caveat is that showing up in a citation list is not the same as changing an answer. Trellner says explicitly that it did not test whether removing these sources would produce different recommendations, and that limitation matters. A source can be cited as supporting material after a recommendation is already formed from stronger signals. What the finding does prove is that the retrieval pass in a commercial query is not filtering hard on provenance, and that is a different problem from the ranking pass most practitioners are used to arguing about.
Obscurity is not the same as irrelevance
The Tranco distribution is the number that should change how you scope the work. If the median cited domain sits around rank 71,611, the sources grounding your category are, in the main, sites with no meaningful direct audience. Your competitive set in search and your competitive set in AI vendor recommendations are overlapping but different lists, and most teams have only built the first one.
Where Perplexity's 7,534 citations sat in the Tranco popularity ranking, split into non-overlapping bands, from the Trellner Research run of 2 September 2026
There is a reading of this that is too pessimistic, and it is worth heading off. Low popularity rank does not automatically mean low quality. Specialist trade sites, standards bodies, niche documentation and regional publications all sit outside the top hundred thousand and are exactly the sources a careful analyst would want. The distribution alone cannot separate a useful obscure source from a manufactured one. That separation is manual work, and it is the actual deliverable.
It also explains something practitioners have been complaining about for a year: why visibility in AI answers moves without any corresponding movement in classic rankings. If two thirds of the grounding corpus for your category lives outside the sites you track, your rank tracker was never going to see the change coming. We walked through the aggregator side of the same dynamic in why SEO aggregators started losing AI citation share, and the source-level view here is the other half of that picture.
Ranking and getting cited are different jobs
A second data point from early September lines up neatly and makes the split concrete. Cleanlist ran 200 usable go-to-market software queries and found a reddit.com URL on 144 of them, which is 72% of queries and 8.0% of every page one slot. When it looked at AI citations for the same queries, community pages made up 11.5% of organic results but only 6.4% of citations. Reddit ranks more than it is cited.
None of this argues that classic search work has stopped mattering. It argues that the two jobs have separated far enough that measuring one and reporting it as the other is now a real error. If your board deck shows position changes and calls them AI visibility, the numbers above are the reason those two lines will keep diverging.
How to audit the sources behind your AI vendor recommendations
This is a half day of work for one person and it is the highest value half day available in GEO right now, because almost nobody has done it. The output is a named list of the domains grounding your category, sorted into sources you can earn, sources you can correct, and sources that are noise.
Run that list and you will end up with something most competitors do not have: a written account of which sources teach an answer engine about your market, which of them are wrong, and which of them you could realistically influence inside two quarters. For B2B software companies in particular, that document tends to reorder the content roadmap within a week of being read.
The move that looks obvious and is not
The tempting conclusion from 215,128 generated pages ranking as a top source is to generate 215,129. Do not. The economics look attractive for about one quarter and then stop, for reasons that have nothing to do with taste.
Search platforms already police exactly this pattern under site reputation and scaled content abuse policies, and the enforcement risk lands on your main domain rather than on a disposable one, which we went through at length in the multi-region site reputation abuse playbook. Beyond the policy question there is a simpler commercial one: the three farm domains earn citations because the corpus is thin, and a thin corpus is a temporary condition. Engines are already tightening provenance handling on commercial queries, and when that tightens further, a brand that built its AI presence on generated inventory is holding a liability, not an asset.
The durable position is the boring one. Be the source with a name on it, a date on it, stated criteria, real product knowledge and nothing behind a form. That combination is scarce enough in most software categories to be a genuine advantage, and it is the only version of this work that survives the engines getting better at telling manufactured evidence from the real thing.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.