AI answer engines do not concentrate their citations equally across industries, and the difference is structural rather than marginal. In the Machine Relations Index release of 2026-10-14 (25,323 monitored answer runs, six engines, 1,206 buyer prompts, 30,757 cited domains over 151 days), the single most-cited source in consumer health appears in 26.3% of that category's answers, while the most-cited source in industrial appears in 7.2%. The kind of source that leads changes too: consumer categories are led by one dominant editorial or community destination, B2B software categories by vendor-owned domains, and the three categories that cleared the evidence floor in October (local services, iGaming & betting and industrial) by sources that fit no named class at all.
A brand running one AI visibility playbook across verticals is optimising for an average that describes no actual industry.
Last updated: October 14, 2026
What this page measures, and on whose instrument #
Every figure below is a direct observation from the Machine Relations Index, methodology version mri_score_v2.1, release mri_score_v2.1+2026-10-14+7a624b28cd41, window 2026-05-10 to 2026-10-14. A monitored run is one buyer prompt issued to one engine, and a domain's citation rate is the share of monitored runs in that scope which cited it.
The Index publishes a rate only above its evidence floor: at least 10 observed runs across at least 7 distinct dates. Below that line a segment is marked collecting, which is not a zero rate but the instrument declining to publish a number the evidence cannot yet carry. Of 157 strata (25 taxonomy nodes paired with buyer question shapes), 117 are published and 34 are still collecting in this release, up from 96 published on 2026-09-28. Every rate below carries its confidence grade in parentheses: A, B or C by how much evidence stands behind it.
Google AI Overviews and Google AI Mode stopped returning an observable answer surface for part of this window, so they are observed on 92 and 109 of 151 days. What they observed is real and stays in the Index; the dates, the effect on the denominator and the provider replacement are logged in the Index's correction record, on the model of Google's own Search Console data anomalies log.
Citation concentration ranges more than 3x across categories #
| Category | Monitored runs | Leading source | Its citation rate | 10th-place rate | Largest classified source class | Unclassified share |
|---|---|---|---|---|---|---|
| Consumer Health | 1,321 | healthline.com | 26.3% (B) | 7.3% | Editorial publication 26% | 50% |
| Consumer Finance | 1,328 | nerdwallet.com | 24.9% (A) | 7.5% | Editorial publication 28% | 53% |
| AI Security & Privacy | 1,252 | arxiv.org | 23.0% (A) | 5.8% | Vendor-owned 21% | 44% |
| Emergent Prosumer | 871 | reddit.com | 17.6% (A) | 4.5% | Editorial publication 20% | 64% |
| AI Infrastructure | 1,147 | medium.com | 17.5% (A) | 6.5% | Community platform 18% | 50% |
| Family Software | 923 | reddit.com | 17.0% (A) | 10.1% | Editorial publication 23% | 67% |
| AI Visibility & GEO | 1,226 | youtube.com | 17.0% (A) | 6.4% | Vendor-owned 22% | 58% |
| Consumer Products | 1,325 | reddit.com | 16.8% (A) | 4.7% | Editorial publication 33% | 50% |
| Local Services | 631 | angi.com | 16.8% (B) | 5.7% | Editorial publication 8% | 82% |
| iGaming & Betting | 800 | track360.io | 15.9% (B) | 4.8% | Editorial publication 9% | 79% |
| Martech & Advertising * | 578 | martech.org | 15.7% (B) | 8.3% | Vendor-owned 33% | 21% |
| Healthcare Services | 1,490 | nih.gov | 14.9% (A) | 3.4% | Editorial publication 17% | 38% |
| Education & Training | 1,087 | reddit.com | 14.0% (A) | 4.7% | Community platform 9% | 70% |
| Cybersecurity | 1,771 | sentinelone.com | 13.0% (B) | 5.1% | Vendor-owned 37% | 31% |
| Deep Tech & Hardware | 1,135 | reddit.com | 13.0% (A) | 5.8% | Academic and government 13% | 64% |
| Logistics & Freight * | 177 | wearewarp.com | 13.0% (C) | 7.3% | Market database 6% | 76% |
| HR & Talent | 1,473 | peoplemanagingpeople.com | 12.2% (B) | 5.6% | Vendor-owned 21% | 37% |
| Enterprise Software | 1,463 | erpresearch.com | 11.3% (B) | 4.3% | Vendor-owned 34% | 25% |
| Fintech | 1,512 | stripe.com | 11.0% (B) | 3.8% | Vendor-owned 25% | 44% |
| Industrial | 626 | linkedin.com | 7.2% (A) | 3.0% | Trade press 5.5%‡ | 81% |
* Not every question shape in this category clears the evidence floor yet (martech 1 of 7, logistics 2 of 6); the rates shown are computed only over its published strata. Source-class shares are computed over each category's 100 most-cited domains, weighted by cited runs; "unclassified" is the Index's "other observed source" class.
Part of the editorial and market-database labels rest only on a word in the domain name. With those labels removed, editorial shares fall by up to 7.3 points and the leading classified class changes in healthcare services, iGaming & betting, local services, industrial and logistics & freight; the measured range for every category is in the Index's source-role classification sensitivity analysis.
‡ Industrial is read by hand. The Index's automated classifier labels 95 of industrial's top-100 cited runs editorial (8%), but it assigns that label partly from words in a domain name, and in this category 71 of those 95 runs go to equipment makers whose domains contain "press" or "monitor" (metalpressmachinery.com, beckwoodpress.com, industrialmonitordirect.com and others), while the most-cited trade magazine, thefabricator.com, is left unclassified. Counting the trade titles directly (The Fabricator, Modern Machine Shop, Metalforming, Assembly, Wood Magazine and Cutting Tool Engineering) gives 66 of 1,196 cited runs, 5.5%.
Read the first and last rows together. In consumer health, one domain answers more than a quarter of the category's monitored questions, and the tenth-ranked domain still reaches 7.3%. In industrial, the leader reaches 7.2% and the tenth-ranked domain 3.0%: a flat distribution with no gatekeeper. These are two different competitive games. Displacing a leader is a coherent strategy in the first and an incoherent one in the second, because there is nothing at the top to displace.
Consumer categories cluster at the top of that table and B2B categories at the bottom, with healthcare services sitting in between: institutional, led by nih.gov at 14.9% (the public index behind it, PubMed, is the gatekeeping surface a healthcare brand has to reach), but nothing like as concentrated as consumer health.
Consumer health's concentration matters more than its rank suggests, because of who is asking. Pew Research Center's February 2026 survey of 5,119 U.S. adults found that 49% now use AI chatbots, 42% use them to search for information, and 20% use them for medical advice (Pew Research Center, 2026). One domain answering 26.3% of monitored consumer-health questions is a concentration with a fifth of the adult population downstream of it.
The newly measured categories run on sources no class names #
Three categories crossed the evidence floor between the 2026-09-28 and 2026-10-14 releases and now publish all six question shapes: local services, iGaming & betting and industrial. They share one property no earlier category had to this degree. Among each one's 100 most-cited domains, 82%, 79% and 81% of cited runs go to domains outside every named source class, against 25% in enterprise software and 31% in cybersecurity.
- Local services is led by two service marketplaces, angi.com at 16.8% (B) and thumbtack.com at 15.8% (B), then reddit.com at 11.1% (A). Editorial publications hold 8% of its top-100 citations.
- iGaming & betting is led by software suppliers to operators: track360.io at 15.9% (B), gammastack.com 11.1% (C), smartico.ai 9.6% (C). Editorial publications hold 9%.
- Industrial is the flattest category the Index has published: linkedin.com leads at 7.2% (A), followed by a supplier directory (iqsdirectory.com, 5.4%, C) and a manufacturer's own site (stamtec.com, 5.0%, C), and the tenth-ranked source reaches 3.0%. Trade magazines hold about 5.5% of its top-100 citations, led by thefabricator.com at 3.7% (still collecting); most of the rest go to manufacturers' and suppliers' own sites and industrial directories such as thomasnet.com and directindustry.com.
For a brand in these categories the usual earned-media list (national business press, trade titles) is a small part of what the engines draw on. The citation surface is marketplaces, supplier directories and specialist vendors' own pages. Logistics & freight points the same way on its two published shapes (76% unclassified, freightwaves.com the top editorial source at 6.8%), with four shapes still collecting.
Which kind of source leads flips between consumer and B2B #
The source class that holds the largest share is not a constant to be optimised against. It is a property of the vertical.
In enterprise software, vendor-owned domains hold 34% of the cited runs among the category's 100 most-cited domains, and market-and-company databases a further 14%; its leader is now erpresearch.com at 11.3% (B), ahead of linkedin.com at 9.8% (A). In cybersecurity vendor-owned holds 37% and editorial publications 11%, and the top five sources are four vendors and Reddit: sentinelone.com 13.0% (B), paloaltonetworks.com 13.0% (A), reddit.com 10.3% (A), microsoft.com 10.1% (A), huntress.com 8.4% (B). In fintech, vendor-owned holds 25% and the leader is stripe.com at 11.0% (B).
In consumer health a single vendor-owned domain appears among the category's 100 most-cited sources, a hospital system's site with 21 cited runs. The leaders are healthline.com at 26.3% (B), nih.gov at 17.0% (A), youtube.com at 17.0% (A) and reddit.com at 10.6% (A). Consumer finance leads with nerdwallet.com 24.9% (A), reddit.com 18.6% (A), youtube.com 18.4% (A), forbes.com 15.9% (A).
The practical consequence is that "invest in owned comparison content" and "earn institutional coverage" are not competing philosophies. They are correct answers to different verticals, and the Index says which is which per category.
63% of AI citations go to domains that fit no named source class #
This is the finding that should make every tidy source-mix pie chart suspect, including the ones in this article's earlier versions.
Across the whole release, the nine source classes the Index assigns split as follows, by cited runs:
| Source class | Cited runs | Share | Distinct domains |
|---|---|---|---|
| Other observed source | 105,997 | 63.2% | 25,744 |
| Editorial publication | 17,637 | 10.5% | 1,751 |
| Vendor-owned source | 17,245 | 10.3% | 1,291 |
| Market and company database | 7,696 | 4.6% | 832 |
| Community and social platform | 6,577 | 3.9% | 57 |
| Academic and government source | 5,883 | 3.5% | 626 |
| Analyst and consulting research | 4,242 | 2.5% | 416 |
| Search or media platform | 1,988 | 1.2% | 27 |
| Wire and press-release distribution | 567 | 0.3% | 13 |
25,744 domains, 84% of every domain observed being cited, belong to no recognised class of publisher, platform, vendor, database, journal or wire. They collectively take 63% of the citations, up from 62% on 2026-09-28 as the newly measured categories entered. Any model of the AI citation market built from six or eight clean buckets is describing the 37% that is easy to label.
Two rows in that table deserve reading as pairs. Community and social is 57 domains holding 3.9%, and reddit.com alone is cited in 9.6% of all 25,323 monitored runs: rank 1 in the entire Index, grade A, cited on 145 of 151 observed days, by four engines (Gemini, Google AI Mode, Google AI Overviews and Perplexity) and not by ChatGPT or Claude. Reddit's influence is real and it is one domain on some engines, not a channel; "community strategy" that is not Reddit-specific is not the same investment. Wire and press-release distribution is 13 domains holding 0.3% of cited runs, with prnewswire.com at 0.9% (B) of all monitored runs. A press release reaches the engines. It does not reach them the way a wire vendor's own marketing implies.
Question shape moves citation density more than most verticals do #
Averaging across categories hides a second dimension. Restricted to published strata, cited-source observations per monitored run vary by the shape of the buyer's question:
| Question shape | Monitored runs | Cited-source observations per run |
|---|---|---|
| Top list | 3,009 | 2.15 |
| News topic | 5,581 | 2.15 |
| Problem first | 3,137 | 1.93 |
| How to choose | 3,111 | 1.90 |
| Best X | 3,197 | 1.86 |
| Is X worth it | 3,105 | 1.75 |
| X vs Y | 3,062 | 1.45 |
A ranked-list question draws 48% more cited sources per answer than a head-to-head comparison. For a brand, that means the same category contains high-slot and low-slot questions, and a comparison page competes for materially fewer citation slots than a category ranking does, regardless of vertical.
What the Index does not reach #
Four of the 24 subject categories have no published stratum in this release: physical consumer products (177 monitored runs), professional services (179) and sales, GTM & CRM (150) are accumulating evidence below the floor, and VC & private equity has none. Martech & advertising publishes one of its seven shapes and logistics & freight two of six. For the unpublished segments the Index publishes nothing, and nothing here should be read as a low citation rate.
Several verticals commonly quoted in AI citation commentary (travel, legal, insurance, real estate, apparel, food and beverage) are not in this basket at all. The Index does not measure them, so this page does not report them. What it does report about travel comes from outside. Similarweb's 2026 Generative AI Landscape Report, covering U.S. desktop usage, found that the share of ChatGPT answers carrying a citation reached 22.6% in travel, 13.5% in retail and 10.7% in sports against a 6.8% overall average, and that reviews and user-generated content supply 54.1% of ChatGPT's top travel citations (Similarweb, 2026; Search Engine Journal, 2026). That same report found 65% of cited URLs sit two or three folders deep while 58.8% of AI referral traffic lands on homepages: cited pages and entry pages are different pages, and blending them is a third way to get a vertical read wrong. That is consistent with the pattern above (consumer decision categories concentrate on a small set of destination sources), but it is a third party's measurement, not ours.
Being cited is not the same as being named #
Citation counts overstate visibility in some engines and understate it in others, and the gap is measurable on a second instrument.
The Answer-Source Fidelity read of 14,447 readable citations across four engines found that the cited source's own name reached the answer text 8.6% of the time. Perplexity is the lowest at 440 of 7,501 readable citations, 5.9%, because it credits sources by numbered marker rather than by name. On Perplexity, being cited and being named coincide about one time in seventeen, so any blended "mention count" across engines is mostly noise.
Semrush and Kevin Indig reached a structurally similar conclusion from outside our data, measuring 3,981 domain appearances across 115 prompts in 14 countries: a majority of AI citations are what they call ghost citations, where the source is linked but the brand name never appears in the answer text, and the citation-to-mention ratio inverts between engines (Semrush, 2026). Two instruments, different prompt sets, same finding: citation and mention are separate events and must be measured separately.
There is a retrieval-side reason to expect citation selection to be systematically biased rather than merit-ordered. Google states that its AI features surface links to supporting web sources and that there are no additional technical requirements to be eligible beyond ordinary indexing (Google Search Central) — which locates the selection decision at retrieval rather than at anything a publisher can configure. Answer engines retrieve and rerank passages before they generate, which means citation is awarded at the retrieval layer to whatever the index can surface and verify (Karpukhin et al., 2020; Nogueira and Cho, 2019). Liu, Zhang and Liang's audit of generative search engines found that only about half of generated statements were fully supported by their own cited sources (Liu et al., 2023). And research presented at ICML found that in Google AI Overviews, AI-generated documents are cited more often than human-authored ones even after controlling for retrieval rank, with the effect strongest at high ranks (Kakimov et al., 2026). In a vertical where institutional credibility already decides selection, that bias compounds the incumbent's advantage.
What vertical citation patterns mean for Machine Relations #
Five structural conclusions, each resting on a figure above.
1. Concentration decides whether displacement is even a strategy. With a 26.3% leader and a 7.3% tenth place, consumer health has a gatekeeper set worth contesting. With a 7.2% leader and a 3.0% tenth place, industrial does not; presence across many surfaces beats pursuit of any one.
2. Source class follows the vertical, not best practice. Vendor-owned is the largest classified class in cybersecurity (37%), enterprise software (34%) and fintech (25%). In consumer categories the leaders are editorial destinations, community platforms and institutions; in local services, iGaming and industrial they are marketplaces, specialist suppliers and directories. Allocating channel budget from a generic list rather than from the category's own distribution misallocates it.
3. The long tail is the market, not the exception. 63% of citations go to 25,744 unclassified domains, and in the three newest categories the share is about four in five. Strategy built only around named publishers, named platforms and named vendors addresses a minority of the citation surface.
4. Question shape is a second allocation axis. Ranked-list questions carry 2.15 cited sources per answer against 1.45 for head-to-head comparisons. Two pages in the same category are competing for different numbers of slots.
5. Measure citation and naming separately or the number is meaningless. At 8.6% name-reach overall and 5.9% on Perplexity, a brand can be heavily cited and almost never named. Any single "AI visibility score" that blends the two hides which one is happening.
Methodology #
Primary instrument. Machine Relations Index, contract machine_relations_index_public_view_v2.0, methodology mri_score_v2.1, release mri_score_v2.1+2026-10-14+7a624b28cd41, window 2026-05-10 to 2026-10-14, 151 observed days. 25,323 monitored answer runs from 1,206 eligible buyer prompts across six engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity), producing 190,582 source events and 49,341 cited-domain observations over 30,757 distinct cited domains. Basket: 25 taxonomy nodes (24 subject categories plus a legacy news-topic bucket) paired with buyer question shapes, giving 157 strata, of which 117 are published and 34 are collecting. Evidence floor: 10 observations across 7 distinct run dates. Confidence grades A, B and C reflect accumulated evidence; membership in any external catalog is recorded as a label and never used to decide what is scored. Per-category source-class shares are computed over each category's 100 most-cited domains weighted by cited runs, and are stated on that denominator rather than on the category's full domain set.
Known data issues. Google AI Overviews and Google AI Mode stopped returning observable answer surfaces for part of the window and are observed on 92 and 109 of 151 days; the affected runs, the effect on the denominator and the provider replacement are logged in the Index correction record.
Second instrument. Answer-Source Fidelity, Machine Relations Research's companion measure, 14,447 readable citations across four engines, for the citation-versus-naming figures.
External sources. Pew Research Center, "Americans and AI 2026", 5,119 U.S. adults surveyed 17-23 February 2026; Similarweb 2026 Generative AI Landscape Report, direct and as reported by Search Engine Journal; Google Search Central documentation on AI features; Semrush and Kevin Indig, 3,981 domain appearances across 115 prompts in 14 countries; Karpukhin et al. 2020 and Nogueira and Cho 2019 on dense retrieval and passage reranking; Liu, Zhang and Liang 2023 on verifiability in generative search engines; Kakimov et al. 2026, ICML, on citation bias in AI Overviews.
What changed in this version. 2026-10-14 (later the same day): industrial's source-class figure is now a direct count of trade titles (5.5%) rather than the classifier's editorial label (8%), which in this category rested mostly on equipment makers' domain names. 2026-10-14: every Index figure re-derived from release 2026-10-14. Local services, iGaming & betting and industrial now publish all six shapes and enter the tables; industrial replaces enterprise software as the least concentrated category; enterprise software's leader changed from linkedin.com to erpresearch.com; HR & talent now publishes every shape. Earlier, this page rested its vertical tables on measurements published by AI-visibility and SEO vendors; those were replaced with first-party Index observations and primary sources. Where the Index does not reach a vertical, this page says so instead of reporting someone else's number for it.
FAQ #
Which industry has the most concentrated AI citations? #
Consumer health, of the 20 subject categories with published data in the Machine Relations Index release of 2026-10-14. Its most-cited source, healthline.com, appears in 26.3% of the category's 1,321 monitored answer runs (confidence B), and the tenth-ranked source still reaches 7.3%. Industrial is the least concentrated: its leader, linkedin.com, reaches 7.2% of 626 runs, and the tenth-ranked source 3.0%.
Do AI engines cite vendor websites or third-party sources more? #
It depends entirely on the category. Among the 100 most-cited domains in cybersecurity, vendor-owned domains hold 37% of cited runs; in enterprise software 34% and fintech 25%. In consumer categories the leaders are editorial destinations, community platforms and institutions, and vendor-owned domains are minor. Across the whole release, vendor-owned domains hold 10.3% of cited runs against 10.5% for editorial publications.
Which sources do AI engines cite for local services, iGaming and industrial? #
Mostly sources outside every named class. Among each category's 100 most-cited domains, 82% (local services), 79% (iGaming & betting) and 81% (industrial) of cited runs go to unclassified domains. Local services is led by angi.com (16.8%, B) and thumbtack.com (15.8%, B); iGaming by track360.io (15.9%, B); industrial by linkedin.com (7.2%, A). Editorial publications hold 8% in local services and 9% in iGaming; in industrial, trade magazines hold about 5.5%, and most citations go to manufacturers' own sites and supplier directories.
How much of the AI citation market do wire services hold? #
0.3% of cited runs in this release, across 13 wire and press-release distribution domains. prnewswire.com, the largest, is cited in 0.9% of all 25,323 monitored runs (confidence B). Press releases do reach answer engines; they are not the centre of the citation market.
Does the type of question change how many sources get cited? #
Yes, and by more than many verticals differ. Across published strata, ranked-list questions draw 2.15 cited-source observations per monitored run and head-to-head comparison questions 1.45, a 48% gap. The same category therefore contains high-slot and low-slot questions, which matters when choosing what to publish.
Is Reddit as important for AI citations as people claim? #
More important as a single domain and less important as a category than the framing suggests. reddit.com is rank 1 in the entire Index at a 9.6% citation rate across all 25,323 monitored runs (confidence A), cited on 145 of 151 observed days, by Gemini, Google AI Mode, Google AI Overviews and Perplexity but not by ChatGPT or Claude. The whole community-and-social class is 57 domains holding 3.9% of cited runs. Reddit leads five of the 20 published categories and does not lead in healthcare services, cybersecurity, fintech, enterprise software or any of the three newly measured categories.
Why doesn't this page cover travel, legal or insurance? #
The Machine Relations Index basket does not include them, so the Index has no observation to publish. Four of its 24 subject categories also have no published stratum in this release because they sit below the evidence floor, which is not a low citation rate but an absence of settled evidence. The one external figure this page reports for travel is attributed to Similarweb's report rather than to our instrument.