Summary: The Machine Relations Index puts every cited domain into one of nine source roles. Among the classified citations, one role leads all six buyer-question shapes (best tools, how to choose, is it worth it, problem-first research, top lists, head-to-head comparisons) in 8 of the 23 categories that clear the evidence floor in the 7 November 2026 release. Editorial publications sweep consumer-finance, family-software and physical-consumer. Vendor-owned sources sweep cybersecurity, fintech, martech-advertising, sales-gtm-crm and ai-visibility-geo, the category this site tracks. The other 15 categories change leader with the question. This replicates a September result on a larger release: all four categories that swept on 23 September still sweep.
An answer engine choosing sources for "best budget monitors" is not necessarily drawing on the pool it uses for "are budget monitors worth it" or "budget monitor vs premium monitor". The Index scores each domain per category and per question shape for that reason. This study holds the category fixed and asks one thing: does the type of source that leads change as the question changes?
What was measured #
Machine Relations Index public view, the November 7, 2026 Machine Relations Index release, methodology v2.1, generated 2026-11-07. Window 2026-05-10 to 2026-11-07, 175 observed days. 36,423 cited source domains, 252,998 source events, 37,140 observed answer runs across six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.
For each category and each of the six buyer-question shapes, every domain's citations were summed by that domain's own source-role label, leaving out the "other observed source" label. A citation here means one domain cited in one answer run, so a run that cites three domains adds three. A category entered the comparison only if all six shapes were published, had at least seven run dates, and had at least 20 classified citations. The release holds 24 subject categories plus a "legacy-unmapped" bucket, which is not a subject category and is excluded. 23 of the 24 clear the bar. vc-private-equity does not: four of its six shapes are still collecting.
The result #
A category "sweeps" when the same class leads all six shapes.
| Category | Sweeping class | Share of classified citations, weakest to strongest shape | Run dates in thinnest shape |
|---|---|---|---|
| ai-visibility-geo | Vendor-owned | 36.8% to 65.7% | 13 |
| consumer-finance | Editorial publication | 41.7% to 79.2% | 13 |
| cybersecurity | Vendor-owned | 42.6% to 67.0% | 13 |
| family-software | Editorial publication | 58.4% to 83.7% | 12 |
| fintech | Vendor-owned | 44.0% to 64.3% | 12 |
| martech-advertising | Vendor-owned | 53.5% to 82.2% | 8 |
| physical-consumer | Editorial publication | 46.4% to 91.9% | 8 |
| sales-gtm-crm | Vendor-owned | 63.7% to 81.6% | 7 |
Physical-consumer, martech-advertising and sales-gtm-crm clear the floor with only 7 to 8 run dates in their thinnest shapes. Read those three as early.
The 15 categories that do not sweep are the larger group. Their leader changes with the question, and several are one shape short of a sweep:
| Category | Leader by shape |
|---|---|
| consumer-health | Editorial publication in five shapes; academic and government in problem-first (42.9% against editorial's 41.5%) |
| consumer-products | Editorial publication in five shapes; academic and government in problem-first |
| emergent-prosumer | Editorial publication in five shapes; community and social in problem-first |
| igaming-betting | Editorial publication in five shapes; vendor-owned in how-to-choose |
| logistics-freight | Editorial publication in five shapes; analyst and consulting in top lists |
| hr-talent | Vendor-owned in five shapes; editorial publication in one |
| professional-services | Analyst and consulting in five shapes; vendor-owned in is-it-worth-it |
| ai-security-privacy | Vendor-owned in four; editorial publication in problem-first; academic and government in head-to-head |
| enterprise-software | Vendor-owned in four; market and company database in two |
| industrial | Editorial publication in four; academic and government in two |
| local-services | Editorial publication in four; community and social in two |
| ai-infrastructure | Community and social in four; vendor-owned in top lists and head-to-head |
| deep-tech-hardware | Academic and government in four; community and social in two |
| education-training | Academic and government in five; community and social in is-it-worth-it |
| healthcare-services | Academic and government in three; vendor-owned in two; editorial publication in one |
Does the September result hold? #
The 23 September release found four sweeps among the 13 categories that cleared the bar then. All four are sweeps in this release, on a window that has grown from 130 to 175 days:
| Category | Sweeping class | Range on 23 Sept release (as published) | Range on 7 Nov release |
|---|---|---|---|
| consumer-finance | Editorial publication | 39.8% to 73.9% | 41.7% to 79.2% |
| family-software | Editorial publication | 54.1% to 79.3% | 58.4% to 83.7% |
| cybersecurity | Vendor-owned | 30.3% to 51.7% | 42.6% to 67.0% |
| fintech | Vendor-owned | 32.5% to 52.2% | 44.0% to 64.3% |
The September ranges are quoted from the version of this page published then, not recomputed: the 23 September release is not in the repository's release history, which starts on 27 October. The same method was rerun on the 29 October release, which is. Seven of its eight sweeps are sweeps again on 7 November. Consumer-health swept on 29 October and no longer does: its problem-first shape now leads with academic and government sources by 1.4 points. Sales-gtm-crm is the new one, and only because it newly clears the evidence floor: on 29 October it had too few run dates to compare. A sweep is a reading of one release, and the margin in some shapes is thin.
AI visibility and GEO changed more. On 23 September it split three ways, with no class above 36.5%. On 29 October and again on 7 November vendor-owned sources lead all six shapes. Both a longer window and a changed set of classified domains could have moved it; the September release is not available to separate them, so this page does not say which.
What the AI visibility sweep looks like #
In ai-visibility-geo, vendor-owned sources hold 36.8% to 65.7% of classified citations by shape. Editorial publications hold 11.1% to 18.4%. Community and social platforms hold 7.3% to 25.5%, led by problem-first research, where a buyer describes a situation instead of naming a tool.
| Shape | Vendor-owned | Editorial publication | Most-cited vendor-owned domains (citations) |
|---|---|---|---|
| best_x | 65.7% | 11.1% | hubspot.com 51, therankmasters.com 44, semrush.com 39 |
| how_choose | 49.9% | 13.5% | hubspot.com 24, tryprofound.com 23, semrush.com 16 |
| is_x_worth | 46.1% | 16.0% | hubspot.com 54, digitalapplied.com 25, authoritytech.io 25 |
| problem_first | 36.8% | 18.4% | semrush.com 34, hubspot.com 19, derivatex.agency 14 |
| top_list | 62.3% | 13.3% | tryprofound.com 43, therankmasters.com 38, semrush.com 35 |
| x_vs_y | 50.5% | 15.8% | hubspot.com 79, semrush.com 46, monday.com 18 |
authoritytech.io carries the vendor-owned label in the release, with 52 citations across five of the six shapes, and is tied for second in is_x_worth. That reflects a vendor-owned page being cited. It says nothing about why.
Reading this #
This is a map of which source types the engines' answers drew on, per category and question, in this instrument's sample. It is not a finding about why, and it does not show that publishing in any class causes a citation or a recommendation.
Three consequences follow for someone choosing where to put effort.
- Check the category's class before copying a plan. A media plan built for consumer-finance, where editorial publications lead every shape, points at a pool that does not lead in cybersecurity or AI visibility, and the reverse holds too.
- A vendor-owned sweep means vendor domains are the largest classified group in the pool. In the five vendor-owned categories, a brand's own site competes with other vendors' sites for citations in every shape. The Index counts domains, not page types, so it does not show which kind of vendor page is cited.
- A sweep is not a ceiling for other classes. In AI visibility, editorial publications hold 11% to 18% of classified citations depending on the shape. That is a smaller share than vendor-owned, and not an absent one.
What this does not say #
It does not say why these categories behave this way, and the association is offered without a mechanism. It does not say a source outside the leading class cannot be cited: every sweeping category has non-leading classes with a nonzero share in every shape.
The largest limit is coverage. The "other observed source" label sits outside this study by construction, and the classified layer is a minority of every cell. In the AI visibility shapes the classified domains account for 25% to 36% of all domain-run citations; across the 138 category-shape cells compared, the classified share runs from 7% to 54%. The finding is about the classified layer. If the unclassified majority is distributed differently, the real picture differs. Labels are also assigned per domain, by review, site evidence or domain suffix, and a domain's label can change between releases, as the AI visibility movement shows.
It also does not extend to the one category still collecting, to questions outside the six shapes, or to what a person sees in a consumer app: the Index collects through API routes, and Google AI Mode collection changed between 10 and 30 August 2026, so its answers from 31 August are a separate collection regime. Run dates in the thinnest shapes are 7 to 8 for three of the eight sweeps. The release manifest marks coverage as degraded: Gemini returned citation-bearing answers on 149 of the 175 days.
Methodology and sources #
Machine Relations Index public view, contract machine_relations_index_public_view_v2.0, methodology v2.1, the November 7, 2026 Machine Relations Index release, generated 2026-11-07, read whole from machinerelations.ai/data/machine-relations-index.json and checked against the release manifest. Window 2026-05-10 to 2026-11-07, 175 observed days. The Index's evidence floor for a published segment is at least 10 observations across at least 7 run dates; this study additionally required at least 20 classified citations in each category-shape pair before comparing leaders, a stricter bar applied only here.
The nine source-role classes are defined by the release: editorial publication (editorial_media), vendor-owned (vendor_owned), market and company database (market_database), academic and government (academic_government), community and social (community_social), analyst and consulting (analyst_research), search or media platform (platform_search), wire and press-release distribution (wire_distribution) and other observed source (uncategorized_source), which is excluded from every denominator here. See Source Role and the sensitivity of rankings to source-role classification.
Per-domain source-role labels and per-category, per-shape citation counts were read from the release's own fields (the source_role of each domain and its per-category, per-shape runs_cited), not re-derived. The same procedure was run on the 29 October release (methodology v2.1, 166 days, 34,602 domains) for the stability comparison. A seventh shape, news topic, is published for seven categories and is outside the six buyer-question shapes compared here. In cybersecurity, fintech and martech-advertising, vendor-owned sources also lead it. For the live instrument and its current release, see the Machine Relations Index and the Answer-Source Fidelity measurement. For the same release's category concentration, see how many domains it takes to reach half of a category's citations.