We asked 2 web-grounded models for the champion products successful 380 package categories and kept each URL they retrieved. Of the 7,534 citations that came back, 59.8% constituent astatine domains classed worse than #100,000 successful the Tranco top-1M list and 23.4% astatine domains that are not successful the apical cardinal astatine all. Two of the sites doing the grounding person fixed their homepage the HTML title “Facts & Grounding Page” — grounding being the retrieval measurement these models execute — and they and a 3rd tract nether apparently communal power person published 215,128 machine-generated champion <category> pages betwixt them; nary of the 3 domains existed earlier December 2023.
What we ran
On 2 September 2026 we put 380 buyer-intent categories — from “CRM software” to “museum postulation guidance software” — to perplexity/sonar and perplexity/sonar-pro done OpenRouter, 1 punctual per class per model, 760 calls successful all. Each telephone asked for a classed apical 5 arsenic JSON, pinch each product’s charismatic homepage domain. All 760 returned a parseable answer, and some models study the URLs they retrieved, which is why they were chosen. The categories were written earlier immoderate results were seen and ne'er revised.
That produced 3,800 proposal slots naming 1,807 chopped products, and 7,534 citations spanning 2,055 chopped domains. We past looked up each cited domain successful the Tranco regular database for 2026-09-01 and successful the Wayback Machine, and fetched each 1 of the 1,502 vendor homepages the models supplied to spot whether it still exists.
Google was near out. Grounding a Gemini exemplary connected OpenRouter intends routing it done OpenRouter’s ain web-search plugin, truthful the citations would picture that plugin alternatively than Google’s retrieval. Only Perplexity was measured, and thing present should beryllium publication arsenic a declare astir immoderate different engine.
Where the citations land
The median Tranco rank of the 5,768 citations that constituent astatine a classed domain is 71,611. Concentration astatine the apical is unremarkable — the 10 most-cited domains return 17.3% of citations — truthful the communicative is not that a cartel of celebrated sites supplies the answers. It is what fills the different four-fifths: 751 of the 2,055 cited domains, 36.5% of them, do not look successful the apical million.
Those domains are besides newer. The median first Wayback seizure is 2020 for the unranked cited domains against 2011 for the classed ones, and 16.6% of the archived unranked domains were first captured successful 2025 aliases later, against 1.6% of the archived classed ones.
The 10 most-cited domains:
Wikipedia, for comparison, was cited 3 times successful 7,534.
The third-largest root is 1 vendor’s trading blog
guideflow.com sells interactive merchandise demos. It is not a reappraisal site, a directory aliases a publisher, and it competes successful nary of the categories we asked about. Its blog was nevertheless cited 194 times crossed 96 of our 380 categories — a 4th of them — placing it 3rd wide and up of Gartner. Each citation is simply a different URL: 96 chopped guideflow.com blog URLs, 1 per category, six of them the Estonian-locale transcript of a post. Its sitemap lists 3,351 blog URLs, 2,176 of them chopped posts. It supplied the grounding for “3D rendering software”, “IVR software”, “RFID software” and “architecture believe software” alike.
Nothing present is deceptive. Guideflow publishes a ample content-marketing blog, arsenic thousands of companies do. The measurement is astir what the retrieval furniture does pinch it: a vendor’s ain listicles astir markets it does not run successful became the third-largest grounds guidelines for a mobility astir which merchandise to buy.
Facts and grounding pages
Three different sites successful the apical 10 and conscionable beneath it are wifitalents.com (71 citations, 27 categories), worldmetrics.org (60, 22) and gitnux.org (50, 23). Together they relationship for 181 citations, 2.4% of the total, and look successful 41 of the 380 categories.
They look to beryllium 1 operation. All 3 were registered done NameCheap betwixt December 2023 and May 2024, each 3 delegate DNS to the aforesaid brace of Cloudflare nameservers, pam.ns.cloudflare.com and sean.ns.cloudflare.com, and each 3 tally the aforesaid page template pinch the aforesaid navigation — Services, Market Data, Software Advice, Editorial Process, Company. Each besides keeps a blog of precisely six posts, and each eighteen are astir the different brands successful the set: 2 posts each connected the different two, and 2 connected a 4th brand, zipdo.co, which sits connected the aforesaid nameserver brace and gives its ain homepage the aforesaid “Facts & Grounding Page” title. Sharing a nameserver brace is beardown circumstantial grounds of a communal Cloudflare relationship alternatively than impervious of ownership, but the template, the taxonomy and the blogs lucifer point for item.
Their standard is the point. Their sitemaps database 103,578, 107,083 and 105,541 URLs, of which 70,731, 71,684 and 72,713 are /best/<something>-software/ pages: 215,128 generated buying guides crossed 3 brands, against six blog posts each (the seventh /blog/ URL successful each sitemap is the blog index). There are not 215,128 package categories.
The self-description is what makes them unusual. Fetched connected 2 September 2026, worldmetrics.org and gitnux.org some return an HTML title of the shape <Brand> — Facts & Grounding Page, and an identical meta explanation isolated from the marque name:
Verified facts astir Gitnux: an independent marketplace investigation institution publishing manufacture statistics, civilization research, and package Best Lists. Company, legal, methodology, and compliance specifications successful 1 machine-readable record.
Grounding is not a word buyers use. It is the sanction of the measurement successful which a retrieval strategy fetches documents to information an reply on. A machine-readable grounds of verified facts astir oneself is not a work to a quality scholar either. These pages are addressed, successful their titles and descriptions, to the package that sounds them.
That reference is being purchased successful the mean measurement arsenic well. worldmetrics.org advertises civilization marketplace investigation “from €5,000”, ready-made reports “from €499” and vendor action “from €2,500”, supra the aforesaid taxonomy of generated Best Lists that the models retrieve.
One template, 3 verdicts
We fetched the aforesaid class page from each 3 brands: “project estimation software”. Each page states its ranking successful JSON-LD, truthful it tin beryllium publication without interpretation. Each ranks 10 tools; the apical 5 are shown.
Gitnux’s victor does not look successful Worldmetrics’ 5 astatine all. Each page carries 3 named unit — Worldmetrics credits Kathryn Blake, Alexander Schmidt and Victoria Marsh; Gitnux credits Diana Reeves, Helena Kowalczyk and Olivia Thornton; WifiTalents credits Ryan Gallagher, Isabella Rossi and Natasha Ivanova — 9 chopped group for 1 question. Each page announces an editorial process; Gitnux labels its consequence “AI-verified · Expert reviewed”. All 3 transportation an unrendered template adaptable successful the byline line, reference “Within the adjacent 26 days” connected 2 of them and “Within the adjacent 40 days” connected the third.
Where the recommendations point
The 1,502 vendor homepages the models supplied are mostly fine. We checked each twice, erstwhile straight and erstwhile done a rotating proxy, counting a tract arsenic reachable if either effort reached it, truthful that a big blocking 1 of our IPs is not recorded arsenic a dormant company.
Ten of the 1,502 resoluteness to nary reside astatine each — 8 of them are not delegated to immoderate nameserver — including graphiql.com (offered arsenic the location of GraphiQL, which has nary specified site), todo.com (offered for Microsoft To Do) and aquasecurity.io (offered for Trivy). Four much resoluteness but ne'er answer. With the 404s, 17 domains — 1.1% — are gone aliases unreachable. Another 92, 6.1%, redirect to a different registrable domain; astir of those are mean acquisitions and rebrands, and we people the afloat database alternatively than conjecture astatine each.
Two are not, and successful some the 2 tiers disagreed. Asked for investigation information guidance platforms, some named Dryad: sonar-pro gave the existent repository astatine datadryad.org, sonar gave dryad.co, which redirects to an Indonesian online-gambling portal whose title originates “BIGSLOT288 | Portal Game Online”. Asked for information value tools, some named Monte Carlo: sonar gave montecarlodata.com, sonar-pro gave montecarlo.com, which redirects to Monte-Carlo Société des Bains de Mer, the Monaco edifice and casino group.
What this does not show
The 2 models are not 2 independent measurements. They returned a byte-identical citation database successful 289 of the 380 categories and their URL sets overlap astatine a Jaccard of 0.898, truthful the Perplexity tiers stock a retrieval furniture and should beryllium publication arsenic 1 hunt stack sampled twice. Their statement connected the apical prime — the aforesaid merchandise first successful 290 of 380 categories — is simply a truth astir that shared retrieval, not grounds that independent systems converge.
The consequence covers Perplexity only. We person not measured ChatGPT, Gemini, Copilot aliases Google’s AI Mode, and location is nary logic to presume their retrieval mixes match.
The 380 categories are our ain construction, not a sample of what buyers really ask, and a database weighted towards niche verticals will aboveground much long-tail sources than a database of communal queries would.
Every page fetch went retired done a rotating datacentre proxy nether a named investigation user-agent, truthful what these sites returned to america is not needfully what they return to a retrieval crawler aliases to a browser.
Four of the seventeen unreachable vendor domains are ample sites, nasdaq.com and solidworks.com among them, that are plainly live and simply ne'er answered an automated request; they are counted arsenic unreachable, not arsenic dead. The full tally is 1 day’s snapshot of a retrieval scale that changes.
One punctual wording, 1 tally per category, nary repetition sampling. An earlier aviator suggested the merchandise shortlist moves noticeably erstwhile “best” is swapped for “most popular” while the citation operation moves overmuch less, but this tally does not measurement it.
Tranco rank is simply a fame measure, not a value measure, and a debased rank is not an accusation. It is utilized present only to abstracted the widely-visited web from everything else; each declare astir a circumstantial tract rests connected that site’s ain pages, which are linked and archived successful the dataset.
We person not shown that immoderate of this changes the answers. We did not trial whether removing these sources would nutrient different recommendations, and Guideflow and the 3 Best List brands whitethorn good sanction reasonable products. What we measured is which documents the grounds guidelines is made of.
Finally, communal power of the 3 brands is inferred from shared infrastructure and an identical template. We do not cognize who operates them; nary of the 3 names an owner.
Data and method
The afloat dataset — each citation, each recommendation, the Tranco and Wayback lookups, the vendor liveness checks — and the scripts that produced each fig supra are astatine /data/manufactured-sources-behind-ai-recommendations/, pinch the method and column documentation alongside. Released nether CC BY 4.0. A PDF type of this study is disposable astatine trellner.com/data/manufactured-sources-behind-ai-recommendations/manufactured-sources-behind-ai-recommendations.pdf.
English (US) · 
Indonesian (ID) · 
