Summary: The Machine Relations Index classifies every cited domain into one of nine source-role types — editorial publication, vendor-owned, market and company database, academic and government, community and social, analyst and consulting, search or media platform, wire and press-release distribution, or other observed source. Within the 13 subject categories that have enough evidence to compare all six buyer-question shapes (best tools, how to choose, is it worth it, problem-first research, top lists, head-to-head comparisons), the same classified source type leads every one of those six shapes in exactly four categories: consumer-finance and family-software, where editorial media wins all six; cybersecurity and fintech, where vendor-owned sources win all six. In the remaining nine, leadership rotates — a different class wins depending on which question a buyer asks.
An answer engine choosing a source for "best budget monitors" is not necessarily choosing from the same pool it draws on for "are budget monitors worth it" or "budget monitor vs premium monitor." The Machine Relations Index scores each domain per category and per one of six question shapes separately for exactly this reason: citation rate is not one number per topic, it is one number per topic-and-question-shape pair. This study holds the category fixed and asks whether the type of source that wins changes as the question shape changes.
What was measured #
Machine Relations Index public view, release mri_score_v2.0+2026-09-23+fdec6388001a, methodology version mri_score_v2.0, generated 2026-09-23. Window 2026-05-10 to 2026-09-23, 130 observed days. 23,280 cited source domains, 129,264 source events, 16,475 observed answer runs across six engines — ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.
For each of the 20 categories in the release's taxonomy, and each of its six published question shapes, every domain's citations were summed by that domain's own source-role classification, excluding the "other observed source" class — the majority, unclassified layer that sits outside the nine named types and is not a finding about any of them. A category-shape pair entered this study only if it cleared the evidence floor on the classified layer specifically: at least 20 classified-source citations across at least 7 distinct run dates. 13 of the 20 categories cleared that bar on all six shapes at once; the other 7 have at least one shape still too thin on classified evidence to compare, and are not in this study.
The result #
Within each qualifying category, the leading classified class was found separately for each of the six shapes. A category "sweeps" when the same class leads in all six.
| Category | Sweeping class | Share range across the six shapes |
|---|---|---|
| consumer-finance | Editorial publication | 39.8% to 73.9% |
| family-software | Editorial publication | 54.1% to 79.3% |
| cybersecurity | Vendor-owned | 30.3% to 51.7% |
| fintech | Vendor-owned | 32.5% to 52.2% |
| Category | Classes that lead at least one shape | Example split |
|---|---|---|
| ai-infrastructure | editorial_media, community_social, vendor_owned | community_social leads how_choose (22.9%) and problem_first (32.8%); vendor_owned leads top_list (35.0%) and x_vs_y (38.9%) |
| ai-security-privacy | vendor_owned, editorial_media, academic_government | academic_government leads only x_vs_y (37.8%); vendor_owned leads the other five |
| ai-visibility-geo | vendor_owned, editorial_media, community_social | three different classes lead across six shapes, none above 36.5% |
| consumer-health | editorial_media, academic_government | academic_government leads how_choose, is_x_worth, problem_first; editorial_media leads best_x, top_list, x_vs_y |
| consumer-products | editorial_media, community_social | community_social leads only problem_first (42.5%); editorial_media leads the other five |
| deep-tech-hardware | editorial_media, academic_government, community_social | three-way split, no class above 53.5% |
| education-training | editorial_media, community_social, academic_government | three-way split, no class above 41.1% |
| emergent-prosumer | editorial_media, community_social | community_social leads only problem_first (43.1%); editorial_media leads the other five |
| enterprise-software | market_database, vendor_owned | market_database leads best_x and how_choose; vendor_owned leads the other four, peaking at problem_first (56.1%) |
Enterprise-software is the closest a non-sweeping category comes to sweeping: vendor-owned sources lead four of its six shapes and reach 56.1% in problem_first, higher than the weakest shape in three of the four sweeping categories. What keeps it out of the sweep table is that market-and-company-database sources (G2-style listings, review aggregators) take the lead specifically in best_x and how_choose — the two shapes where a buyer is assembling a shortlist rather than deciding on one name. The same two classes swap places in enterprise-software depending on whether the question is "what are the options" or "why would I pick one."
Reading the sweeps against each other #
Consumer-finance and family-software sweep with the same class, editorial media, but at very different strength: family-software's floor (54.1%, top_list) sits above consumer-finance's ceiling in three of its six shapes. Cybersecurity and fintech sweep with vendor-owned sources at closer, more moderate margins — no shape in either category clears 52%.
Family-software's top_list share (54.1%) rests partly on one domain: apple.com, classified editorial_media in this release, contributes 40 of the shape's 72 editorial-media citations. Recomputed with apple.com removed entirely, editorial media still leads top_list — 32 of a reduced classified total of 93, or 34.4% — against community_social's 26. The sweep survives the removal of its largest single contributor; the margin does not.
Why this is worth separating from a per-category leaderboard #
A leaderboard that reports "editorial media leads consumer-finance" collapses six different citation pools into one number and loses the information in this study: whether a category behaves the same way across every kind of buyer question, or whether the source an engine reaches for changes with the question. Nine of the 13 measured categories change; four do not. A GEO strategy built category-by-category without checking this treats consumer-finance and enterprise-software as the same kind of problem, when one has a single gate to clear and the other has at least two, occupied by different kinds of sources.
What this does not say #
It does not say why these four categories behave this way, and correlation is not offered as a mechanism. It does not say a source outside the sweeping class cannot be cited in these categories — vendor-owned sources still take a nonzero share of citations in consumer-finance and family-software, they simply never lead a shape. It does not extend to the 7 categories excluded for thin classified evidence, or beyond the 20-category taxonomy this release scores. And the "other observed source" class, 60%-plus of total citations across the whole index, is outside this study by construction: this is a study of the classified minority, not the full citation pool.
Methodology and sources #
Machine Relations Index public view, contract machine_relations_index_public_view_v2.0, methodology version mri_score_v2.0, release mri_score_v2.0+2026-09-23+fdec6388001a, generated 2026-09-23, read whole from machinerelations.ai/data/machine-relations-index.json in an isolated compute environment on 2026-09-23. Window 2026-05-10 to 2026-09-23, 130 observed days. Evidence floor for a published segment is at least 10 observations across at least 7 distinct run dates; this study additionally required at least 20 classified-source citations per category-shape pair before comparing leaders, a stricter bar applied only for this comparison. Confidence grades are domain-level, not stratum-level, and are not aggregated here. Per-domain source-role classification and per-stratum citation counts are read directly from the release's own fields, not re-derived.
Nine source-role classes are defined in 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 this study's denominators throughout.
Illustrative citing domains named above and their own classification in the release: techradar.com, forbes.com, pcmag.com, nytimes.com, cybernews.com, medium.com, psychologytoday.com, finder.com, reddit.com, nerdwallet.com, versusreviews.com and apple.com, the last classified editorial media in this release and discussed above for its concentrated share of one shape.
For the live instrument and its current release, see the Machine Relations Index and the Answer-Source Fidelity measurement.