Who publishes the pages behind AI buying advice?

A recommendation-backing citation is a link between one cited page and one brand that the same AI answer directly ranks as a recommendation. We classified the website behind each of 3,805,095 such links by its relationship to the recommended brand. The recommended brand's own website supplied 1,072,309 links, or 28.2%. Websites of other vendors — companies selling products or services, just not the recommended brand — supplied 776,102, or 20.4%. Together that is 1,848,411 links, or 48.6%: about half.

Independent publishers — commercial media, communities, directories, and reference sites together — supplied 1,338,477 links, or 35.2%. Marketplace listings supplied 3.1%, and 13.2% could not be classified: the cited website's business model was unknown and it did not match the recommended brand's verified website. The common worry, sharpened by the HouseFresh case in classic search, is that big media publishers control buying advice. In AI answers, the larger evidence channel is the sellers themselves.

We grouped cited websites into business-model buckets. Commercial media — professional publishers monetized by ads, affiliate fees, or subscriptions — supplied 761,536 links, or 20.0%. Vendor sites supplied 1,447,750, or 38.0%. Community and social platforms supplied 11.6%, reference and institutional sites 6.2%, search and app marketplaces 3.1%, and directories and review platforms 2.9%. The remaining 18.1% stayed unclassified.

For every commercial-media link there are 1.9 vendor-site links. The 2.9% directory share independently repeats what the directory-tax study found at the category grain: listing platforms are a small part of the evidence behind AI recommendations.

On ChatGPT Search, the recommended brand's own website supplied 557,496 of its recommendation-backing citations, or 50.4%. On Google AI Mode the own-site share was 19.1%. Independent publishers moved the other way: 29.4% on ChatGPT Search and 37.6% on Google AI Mode.

The two engines answer the same buying questions with different kinds of proof. A brand with a weak website is exposed on ChatGPT Search; a brand with no independent coverage is exposed on Google AI Mode.

The bucket mix by engine shows the same divergence. Community and social platforms supplied 14.3% of Google AI Mode's recommendation-backing citations and 5.2% of ChatGPT Search's. Search and app marketplaces — google.com, Google Play, and the App Store — supplied 4.3% on Google AI Mode and 0.26% on ChatGPT Search. Directories ran the other way: 4.0% on ChatGPT Search versus 2.5% on Google AI Mode.

Commercial media is the steadiest slice: 22.7% on ChatGPT Search and 18.9% on Google AI Mode.

Commercial media supplied 51.1% of the recommendation-backing citations in Financial Services prompts — 106,494 of 208,429. Gaming (39.1%), Health Care (36.9%), Sports (36.8%), and Food and Beverage (36.6%) followed. At the other end sit the B2B categories: Sales and Marketing 12.7%, Information Technology 12.5%, Software 11.0%, Artificial Intelligence 9.8%, and Data and Analytics 8.3%.

The affiliate-review economy that Google's site-reputation-abuse policy targets — card pickers, insurance comparisons, product roundups — is exactly where AI answers still lean on commercial media. In software categories the same engines mostly read vendors' own pages.

The ownership split by industry makes the same point from the brand's side. In Professional Services, 40.2% of recommendation-backing citations came from the recommended brand's own website. Data and Analytics reached 37.1% and Artificial Intelligence 31.7%. In Consumer Electronics the own-site share was 10.3%, in Clothing and Apparel 12.5%, and in Sports 13.9%, where independent publishers supplied about half the evidence.

Where buyers can read a vendor's documentation and pricing, AI engines do too. Where products are physical and tested by third parties, the engines reach for independent publishers.

Forbes supplied 24,701 recommendation-backing citations from 3,902 pages across 3,526 prompts and 4,137 brands — the most of any commercial-media publisher. TechRadar followed with 23,116 links, then NerdWallet with 11,915, The New York Times with 7,165, and Business Insider with 6,301.

The list mixes general news publishers with review specialists. NerdWallet's links concentrate on 948 prompts — finance questions — while Forbes spreads across 3,526 prompts. The count identifies which publishers already back recommendations; it does not show that coverage there caused the recommendation.

At the single-domain grain, community platforms lead: Reddit supplied 148,676 recommendation-backing citations and YouTube 143,758. google.com — search results, Maps listings, and other Google surfaces, cited almost entirely by Google AI Mode — supplied 108,299. The largest commercial-media domain, Forbes, ranks sixth with 24,701.

Domain leaderboards and business-model mixes tell different stories. Communities have the biggest single domains, but vendor sites win on volume because they are hundreds of thousands of small domains. The live version of the all-citations domain leaderboard is on the public Parse index.

The 761,536 commercial-media citations came from 73,928 distinct media domains. The top 10 supplied 99,050 links, or 13.0% of the media slice. The top 50 supplied 26.3%.

In classic search, a handful of large publishers were accused of crowding out small sites. In the evidence behind AI recommendations, no short list of media brands holds the slice: roughly three quarters of media-backed evidence comes from outside the top 50 publishers.

We flagged cited pages whose titles match roundup and comparison patterns — best, top N, review, versus, comparison, deals, or discount codes. Among distinct cited pages with an observed title, 40.7% of commercial-media pages matched, against 30.6% of directory pages, 23.6% of community pages, and 21.4% of vendor pages. Overall, 26.8% of the 1,530,328 titled pages matched.

The vendor number is the notable one: one in five vendor pages backing a recommendation is itself a roundup-style page, often a vendor's own "best tools" listicle that includes the vendor. Title patterns are a monetization signal, not proof — a matching title does not confirm affiliate links on the page.

For the brand ranked first in an answer, vendor sites supplied 38.3% of the backing citations and commercial media 20.6%. For brands ranked second or lower, the shares were 38.0% and 19.8%. No bucket differed by more than 1.6 percentage points between the two groups.

The publisher mix behind the top pick is the same one behind the rest of the shortlist. What differs for the top pick is whether it gets cited at all, which a separate study measures.

Buckets come from the index's per-website source-type label plus a manual overlay: we hand-classified the 118 highest-volume websites the index labels only as general-purpose, which covered 11.7% of all links. After the overlay, 18.1% of links stayed unclassified, down from 29.8% before it. The ownership split marks a link as the brand's own site only when the cited website matches the recommended brand's verified website; 13.2% of links stayed unclassified there because the cited website's business model was unknown and no own-site match was found. Industry cuts cover the 58.1% of links whose prompt maps to an industry.

Monetization is inferred from publisher type and title patterns; we did not verify affiliate links, ads, or paywalls on individual pages. Only ranked recommendations with at least one attached citation are measured — 42.9% of ranked brand appearances — so this study describes the evidence AI shows, not the evidence it consulted. Two self-referential domains were excluded.

Vendor sites supplied 48.6% of the evidence behind AI recommendations, and the recommended brand's own site was the single biggest channel at 28.2%. Commercial media supplied 20.0%, concentrated in finance, health, and consumer categories. Directories supplied 2.9%.

Treat your own website as your first citation channel: it carries half of ChatGPT Search's recommendation evidence. In finance, health, and consumer categories, add commercial-media outreach — that is where publishers still control a third to a half of the evidence. Spend on directory listings last. Rerun the same study over a window of the same length next quarter to see whether the vendor share grows.