How much of AI citation goes to editorial publications depends on how a domain gets called a publication, and the difference is large enough to change which source class leads. In the Machine Relations Index release of 2026-10-14 (25,323 monitored answer runs, six engines, 30,757 cited domains, window 2026-05-10 to 2026-10-14), editorial publications hold 10.5% of all cited runs when a word in the domain name may assign the label, and 7.1% when only site evidence or a reviewed decision may. Under the first rule editorial publications narrowly lead vendor-owned sites, which hold 10.3%. Under the second, vendor-owned sites lead. In five of the 20 categories with published data, the largest classified source class changes with the rule.
No citation count moves between the two readings. Only the label does.
Last updated: October 14, 2026
What was measured #
Every figure below is computed from the public Index file, machine-relations-index.json, methodology version mri_score_v2.1, release mri_score_v2.1+2026-10-14+7a624b28cd41. A monitored run is one buyer prompt issued to one engine. For each domain the file reports the runs that cited it, a source role, and the basis for that role in the field source_role_basis:
- reviewed: a recorded decision about that specific domain.
- site_evidence: the domain's own homepage identifies what it is.
- domain_suffix: a registry suffix such as .gov or .edu.
- name_pattern: a word inside the domain name, such as "news", "press", "review", "journal" or "research", and nothing else.
- none: no rule applies, and the domain is reported as "Other observed source".
The same file is read under two rules:
- As published. Every label stands.
- Evidence or review only. A label whose only basis is a name pattern becomes "Other observed source". Every other label stands.
Neither rule is the truth. The first admits any domain whose name contains a keyword. The second removes real publications that have not yet been reviewed or matched to site evidence. Read together, they bound the figure. Shares are of 167,832 domain citations: each domain counted once for every monitored run that cited it.
Which labels rest on a name alone #
A name pattern is the basis for labels in exactly two classes, the two whose names describe a kind of publisher.
| Source class | Domains | Cited runs | Domains labelled by name alone | Their share of the class's cited runs |
|---|---|---|---|---|
| Editorial publication | 1,751 | 17,637 | 1,150 (65.7%) | 32.8% |
| Market and company database | 832 | 7,696 | 749 (90.0%) | 42.4% |
| Vendor-owned source | 1,291 | 17,245 | 0 | 0% |
| Academic and government source | 626 | 5,883 | 0 (315 by registry suffix) | 0% |
| Community and social platform | 57 | 6,577 | 0 | 0% |
| Analyst and consulting research | 416 | 4,242 | 0 | 0% |
| Search or media platform | 27 | 1,988 | 0 | 0% |
| Wire and press-release distribution | 13 | 567 | 0 | 0% |
Two thirds of the domains called editorial publications carry the label because of a word in their name, but those domains hold one third of the class's citations. Widely cited publications are far more likely to have been reviewed or matched to their own site, so the name rule decides most of the long tail and a minority of the citations. A count of "publications cited" and a share of "citations to publications" therefore carry very different exposure to the rule.
The rule errs in both directions. It reads "press" inside metalpressmachinery.com, an equipment seller's domain, the same way it reads "news" inside usnews.com. It finds no keyword in thefabricator.com, a metal-fabrication trade magazine, which stays "Other observed source".
The whole-release source mix under both rules #
| Source class | As published | Evidence or review only |
|---|---|---|
| Other observed source | 63.2% | 68.6% |
| Editorial publication | 10.5% | 7.1% |
| Vendor-owned source | 10.3% | 10.3% |
| Market and company database | 4.6% | 2.6% |
| Community and social platform | 3.9% | 3.9% |
| Academic and government source | 3.5% | 3.5% |
| Analyst and consulting research | 2.5% | 2.5% |
| Search or media platform | 1.2% | 1.2% |
| Wire and press-release distribution | 0.3% | 0.3% |
As published, editorial publications lead vendor-owned sites by 392 cited runs out of 167,832 (17,637 against 17,245). Removing the name-only labels takes 5,791 runs out of the editorial class and leaves vendor-owned sites ahead by 5,399. A statement that answer engines cite publications more than vendors, or the reverse, is a statement about the classification rule as much as about the engines.
Five of 20 categories change their leading source class #
Category shares below follow the Index's vertical analysis: each category's 100 most-cited domains within its published question shapes, weighted by cited runs. The first two columns name the largest classified class (excluding "Other observed source") under each rule.
| Category | Leading class, as published | Leading class, evidence or review only | Editorial share, evidence only to as published | Unclassified share, as published to evidence only |
|---|---|---|---|---|
| Consumer Products | Editorial 33.1% | Editorial 27.0% | 27.0% to 33.1% | 50.0% to 57.3% |
| Consumer Finance | Editorial 28.5% | Editorial 24.7% | 24.7% to 28.5% | 52.9% to 57.5% |
| Consumer Health | Editorial 26.3% | Editorial 21.1% | 21.1% to 26.3% | 50.3% to 56.7% |
| Family Software | Editorial 23.1% | Editorial 19.2% | 19.2% to 23.1% | 66.7% to 71.4% |
| Emergent Prosumer | Editorial 20.5% | Editorial 18.8% | 18.8% to 20.5% | 63.6% to 65.2% |
| Healthcare Services | Editorial 17.2% | Academic and government 13.1% | 12.4% to 17.2% | 37.5% to 42.3% |
| iGaming & Betting | Editorial 9.3% | Analyst research 3.9% | 3.8% to 9.3% | 79.0% to 84.5% |
| Local Services | Editorial 8.2% | Community and social 5.1% | 4.6% to 8.2% | 82.4% to 86.4% |
| Industrial | Editorial 7.9% | Community and social 5.4% | 0.6% to 7.9% | 80.0% to 88.9% |
| Logistics & Freight | Market database 6.4% | Analyst research 6.0% | 3.0% to 5.2% | 75.7% to 80.2% |
| Cybersecurity | Vendor-owned 36.7% | Vendor-owned 36.7% | 7.0% to 11.3% | 31.0% to 37.9% |
| Enterprise Software | Vendor-owned 33.9% | Vendor-owned 33.9% | 11.6% to 11.6% | 25.5% to 30.5% |
| Martech & Advertising | Vendor-owned 33.8% | Vendor-owned 33.8% | 10.0% to 13.7% | 20.2% to 27.1% |
| Fintech | Vendor-owned 24.7% | Vendor-owned 24.7% | 8.4% to 9.3% | 43.6% to 46.8% |
| AI Visibility & GEO | Vendor-owned 21.8% | Vendor-owned 21.8% | 4.5% to 5.0% | 57.7% to 58.2% |
| HR & Talent | Vendor-owned 21.5% | Vendor-owned 21.5% | 14.5% to 16.8% | 36.8% to 39.2% |
| AI Security & Privacy | Vendor-owned 21.4% | Vendor-owned 21.4% | 2.9% to 3.3% | 44.0% to 46.1% |
| AI Infrastructure | Community and social 18.0% | Community and social 18.0% | 1.9% to 2.2% | 50.4% to 52.6% |
| Deep Tech & Hardware | Academic and government 12.9% | Academic and government 12.9% | 2.2% to 6.3% | 63.6% to 68.5% |
| Education & Training | Community and social 8.5% | Community and social 8.5% | 2.7% to 4.0% | 70.2% to 72.9% |
The flips do not all point the same way, which is why the range matters more than either end of it. Reading the name-only domains in each changed category shows where the truth sits:
- Healthcare services flips because of real publications. Its name-only editorial domains are Becker's Hospital Review, MedCity News, STAT, HealthTech Magazine and HIPAA Journal. Here the strict rule is the one that misleads, and editorial publications plausibly lead at close to the published 17.2%.
- iGaming & betting is similar. RotoWire, Legal Sports Report, VegasInsider, iGaming Today, Sportsbook Review and Traders Magazine account for most of its name-only editorial citations, so its editorial share sits near the top of its range.
- Industrial flips the other way. Most of its name-only editorial citations go to equipment sellers whose domains contain "press" or "monitor" (industrialmonitordirect.com, metalpressmachinery.com, beckwoodpress.com, macrodynepress.com, magnumpress.com), alongside two genuine titles, MetalForming Magazine and WOOD Magazine. A direct count of trade titles, The Fabricator included, gives 5.5% of the category's top-100 cited runs, inside the 0.6% to 7.9% range and well below its top.
- Local services sits in between: its name-only editorial citations go mostly to home-improvement content sites (Best Pick Reports, Plumbing Tips Today, Today's Homeowner).
- Logistics & freight changes leader because erpresearch.com is labelled a market database for the word "research" in its name. With only two of six question shapes published, its leader is the least settled reading in the table under either rule.
The consumer categories keep editorial publications in the lead under both rules but lose two to six points. In consumer health the name-only editorial set mixes publications (Medical News Today, Consumer Reports, US News) with affiliate review sites (Garage Gym Reviews, ReviewPooch), so the true share is neither end of the range.
What holds under either rule #
Three readings survive the change of rule intact.
Vendor-owned leadership in B2B categories. No vendor-owned label rests on a name pattern; each comes from review or site evidence. Vendor-owned sites lead cybersecurity, enterprise software, martech, fintech, AI visibility, HR and AI security under both rules, with identical shares.
Domain-level citation rates. Every domain's numerator, denominator, rate, confidence grade and rank are untouched by either reading. The classification rule changes how citations are summed into classes, never how often a domain is cited.
The size of the unclassified majority. "Other observed source" is 63.2% of all cited runs as published and 68.6% under the strict rule. Any source-class pie chart, from any instrument, describes a minority of what the engines cite; the rule only decides whether that minority is roughly 37% or 31%.
How to report a source-class share #
- Publish the basis with the label. A class share is reproducible only if each domain's label and the reason for it are available. The Index file carries both, so either rule here can be re-run from the public data.
- Report a range where name rules are in play. State the share as published and under evidence or review only. A single figure hides the sensitivity this analysis measures.
- Read the domains before naming a category's leader. In thin categories a few dozen citations to mislabelled or unlabelled domains decide the leader. Industrial and healthcare services fail in opposite directions.
- Keep domain counts apart from citation shares. The name rule decides 65.7% of editorial domains but 32.8% of editorial citations. A count of publications cited is far more exposed than a citation-weighted share.
- Do not compare class shares across studies built on different rules. A difference of three or four points between two studies' editorial shares can be entirely a difference of taxonomy.
What not to conclude #
This analysis does not show any change in what answer engines cite; every observed citation is identical under both rules. It does not show that the strict rule is correct: in healthcare services and iGaming it removes real publications. It does not rank source classes by quality or influence. A source role is metadata used to summarise observations, and the evidence is the domain-level count it summarises.
The earlier version of this note illustrated the same point with a hypothetical regrouping of three domains from the 2026-09-12 release. It is replaced here by the measured effect on the full 2026-10-14 file.
FAQ #
What share of AI citations go to editorial publications? #
In the Machine Relations Index release of 2026-10-14, between 7.1% and 10.5% of 167,832 domain citations across 25,323 monitored runs on six engines. The lower figure counts only domains labelled editorial by review or by their own site's evidence; the higher one also counts domains labelled by a word in their name.
Do AI engines cite vendor sites more than publications? #
Across the whole release the two are close enough that the answer depends on the classification rule: editorial publications hold 10.5% as published and 7.1% under the strict rule, against 10.3% for vendor-owned sites under both. Within categories the answer is clearer. Vendor-owned sites lead seven B2B categories under either rule, and editorial publications lead consumer products, consumer finance, consumer health, family software and emergent prosumer under either rule.
Why not simply use the strict rule? #
Because it removes real publications. In healthcare services the strict rule drops Becker's Hospital Review, STAT and MedCity News out of the editorial class and hands the category's lead to academic and government sources. A name-only label is a hypothesis about a domain, and the strict rule treats every such hypothesis as false.
Can the figures be reproduced? #
Yes. Download machine-relations-index.json, sum mri_score_v2.overall.runs_cited by source_role, then repeat after treating every domain whose source_role_basis is name_pattern as uncategorized_source. Category figures use each domain's strata entries with status published, summed per category, over the category's 100 most-cited domains. The release manifest records the release id and engine coverage.
Sources #
- Machine Relations Research. "Machine Relations Index public data," release mri_score_v2.1+2026-10-14+7a624b28cd41. https://machinerelations.ai/data/machine-relations-index.json
- Machine Relations Research. "MRI release manifest." https://machinerelations.ai/data/mri-release-manifest.json
- Machine Relations Research. "AI Citation Rates by Industry: Why Your Vertical Determines Which Sources Answer Engines Trust." https://machinerelations.ai/research/ai-citation-rates-by-industry-vertical-source-data-2026
- Machine Relations Research. "Why 'Collecting' Is Not a Zero Citation Rate." https://machinerelations.ai/research/collecting-is-not-zero-citation-rate