In the September 18 Machine Relations Index v2 release, editorial publications carry 5.9% of the citation weight in the AI visibility and GEO category. Across all 16 measured subject categories, editorial publications carry 11.3%. Only deep tech hardware (5.4%) and education and training (5.6%) sit lower.
The category that sells AI visibility is measured by answer engines at roughly half the editorial sourcing rate of the market it is being sold into.
That is the finding. The rest of this study shows what the engines cite instead, and why the answer is not a media story but a classification story: most of the sources engines cite in this category have not yet been assigned a role or a grade by the Index.
Editorial citation weight, by category #
Every domain an answer engine cites is assigned one source role. For each published category and question-shape segment, the Index counts how many observed answer runs cited each domain. Summing those counts by role gives the share of citation weight a role carries in that category.
| Category | Editorial share | Unclassified share | Citation weight |
|---|---|---|---|
| Deep tech and hardware | 5.4% | 75.6% | 4,112 |
| Education and training | 5.6% | 77.3% | 3,878 |
| AI visibility and GEO | 5.9% | 74.3% | 5,939 |
| AI security and privacy | 7.7% | 62.9% | 5,358 |
| AI infrastructure | 8.0% | 63.6% | 4,300 |
| Consumer health | 8.3% | 77.6% | 6,062 |
| Enterprise software | 9.6% | 50.4% | 8,206 |
| Fintech | 10.1% | 60.0% | 8,375 |
| Emergent prosumer | 10.2% | 81.0% | 3,229 |
| Consumer products | 12.1% | 78.3% | 5,757 |
| Martech and advertising | 14.3% | 39.9% | 4,602 |
| Cybersecurity | 14.5% | 51.7% | 9,573 |
| HR and talent | 15.8% | 45.0% | 4,666 |
| Family software | 16.4% | 76.0% | 3,424 |
| Consumer finance | 16.7% | 69.7% | 5,430 |
| Healthcare services | 19.3% | 42.5% | 4,630 |
Healthcare services, at 19.3%, cites editorial publications more than three times as often as deep tech hardware. That spread is the point: source selection is category-specific. An engine answering a healthcare question reaches for recognised publications. An engine answering an AI visibility question reaches somewhere else.
Where the citations actually go in AI visibility #
The clearest view is the buyer decision shape. In the ai-visibility-geo / how_choose segment — the question a buyer asks before a shortlist — the Index observed 113 answer runs across 7 distinct run dates, producing 448 domain-segment rows. These are the ten most cited domains. The segment is read in more detail in the companion study, the top editorial publication is cited in only 9.7% of how-to-choose runs.
| Rank | Domain | Source role | Runs cited | Citation rate | Confidence |
|---|---|---|---|---|---|
| 1 | YouTube | Search or media platform | 29 / 113 | 25.66% | A |
| 2 | Brandviz.ai | Unclassified | 15 / 113 | 13.27% | collecting |
| 3 | Gen-optima.com | Unclassified | 13 / 113 | 11.50% | collecting |
| 3 | Mybrandi.ai | Unclassified | 13 / 113 | 11.50% | collecting |
| 5 | Community and social platform | 12 / 113 | 10.62% | A | |
| 5 | Stackmatix.com | Unclassified | 12 / 113 | 10.62% | C |
| 7 | Airfleet.co | Unclassified | 11 / 113 | 9.73% | collecting |
| 7 | Averi.ai | Unclassified | 11 / 113 | 9.73% | C |
| 7 | Community and social platform | 11 / 113 | 9.73% | A | |
| 7 | Shadow.inc | Editorial publication | 11 / 113 | 9.73% | B |
The first editorial publication in the segment, Shadow.inc, is the tenth row of the table but holds a tied seventh rank: six domains are cited more often, and three others, Airfleet.co, Averi.ai and Reddit, are cited exactly as often at 9.73%. The next editorial publication is PR Newswire at 7.96%, then Gracker.ai at 6.19% and Demand Gen Report at 5.31%. Medium and HackerNoon each appear in 2.65% of observed runs. CMSWire, Search Engine Journal and CNBC each appear in a single run. The most cited domain the Index classifies as a vendor-owned source is HubSpot at 7.96%.
Twenty-two of the segment's 448 cited domains are editorial publications. Together they account for 6.1% of the segment's citation weight. Three hundred and seventy-three of them — 83.3% of the cited domains and 78.7% of the citation weight — are domains the Index has not been able to classify into any established source role.
The layer engines cite is mostly ungraded #
The confidence column is the second finding, and the more consequential one.
The Index grades each domain A, B, C or collecting by how much observation history stands behind it in the Index itself. A collecting grade means the Index has not yet observed the domain across enough runs and enough distinct dates to grade it. The grade says nothing about how old, how large or how authoritative the domain is anywhere else; it says only that the Index's own evidence about it is still thin.
Of the seventeen domains cited in eight or more of the segment's 113 observed runs, seven carry a collecting grade. Four of them sit in the top ten. Of the 373 unclassified domains in the segment, 356 are collecting. By the measurement's own standard, most of the sources engines assemble these answers from have not yet been observed often enough to be graded, and have not yet been assigned a source role.
This is what a category looks like before its reference layer has been measured. The demand is real — 113 observed answer runs in a single question shape, 5,939 units of citation weight across the category. The supply, as the Index currently sees it, is a field of domains without a role or a grade, a video platform, two social platforms, and a thin editorial tail.
What this means for the category #
Three readings follow from the data, and one does not.
For publications. The AI visibility category is an open supply-side position. The editorial share is 5.9% against a 11.3% market average, and the domains currently filling the gap are mostly ungraded and unclassified in the Index. A publication with durable, comparable, evidence-backed coverage of how buyers should evaluate AI visibility tooling is competing against sources with no Index grade yet, not against sources with an established citation record.
For vendors. Brand pages are not the mechanism here. In the buyer decision shape, engines mixed video, community, social, vendor-adjacent and editorial sources, and the single largest source was a video platform at 25.66%. A source strategy that assumes the trade press is the path into the answer is measuring a different category.
For the category itself. A market whose answers are assembled from sources the Index has not yet graded is a market where the reference layer is still being measured. Source roles in the Index are observed, not awarded; a domain moves from collecting to a grade by being cited consistently over time, not by claiming a position.
What does not follow: this study does not say editorial coverage is ineffective, and it does not rank any individual publication's quality. It does not say that unclassified or collecting domains are new, small or unreliable; the Index does not measure those things. It measures which domains six answer engines cited in observed runs. A domain's citation rate here is a marginal run rate — the share of observed runs in which the domain was cited at least once. Runs cite multiple domains, so these rates are not exclusive shares and must not be summed or narrated as one source displacing another.
Methodology #
Machine Relations Research analysed the public Machine Relations Index v2 JSON release generated on 2026-09-18. The release identifier is mri_score_v2.0+2026-09-18+8fa38e54dd0a, with public artifact SHA-256 8fa38e54dd0af0c5486b431b9ac3a8ca2c64159700c43094d2243cc005dfb988.
The release observes six answer engines — ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity — across a fixed basket of measured buyer and research prompts. Its window runs from 2026-05-10 through 2026-09-18, 125 days observed. It covers 22,179 cited source domains, 124,397 citation events, 15,782 observed answer runs and 912 eligible queries, organised into 157 category and question-shape strata under basket version mri_taxonomy_v2.0.
A stratum publishes a citation rate only after it clears the evidence floor of at least 10 observed runs across at least 7 distinct run dates. Eighty-five of the 157 strata are published in this release; the rest are marked collecting or empty and are excluded here.
The category comparison sums runs_cited across every published stratum in each category, grouped by the domain's source_role, and divides the editorial and unclassified totals by the category total. It covers the 16 subject categories with published strata. The legacy news-topic bucket is a pre-taxonomy holdover rather than a subject category and is excluded from the comparison and from the 11.3% average; including it moves the all-category editorial share to 12.2%.
The segment analysis selects, for each domain in the release, the mri_score_v2.strata entry matching category ai-visibility-geo and question shape how_choose, then sorts the resulting 448 rows by citation_rate descending. Source role, role label and overall confidence grade are read from the same public record. "Unclassified" in this study is the release's uncategorized_source role: a cited domain the classification has not assigned to an established role. A domain's confidence grade is its overall grade in the release, not a segment grade.
This study uses only public release fields. It does not use raw provider payloads, cited URLs or internal query identifiers.
FAQ #
What is the headline number? #
Editorial publications carry 5.9% of AI answer citation weight in the AI visibility and GEO category, against an 11.3% average across the 16 measured subject categories. It is the third-lowest editorial share of any category in the September 18 release.
Is 5.9% a market share? #
No. It is a share of citation weight: the sum of observed runs citing editorial-publication domains, divided by the sum of observed runs citing any domain, across published strata in the category. One run cites multiple domains, so these are not mutually exclusive shares.
Which source layer is dominant instead? #
Domains the Index cannot yet classify into an established source role. They hold 74.3% of citation weight in the category, and 78.7% in the how_choose buyer segment. Most carry a collecting confidence grade, meaning the Index has not yet accumulated enough observation history on them to grade them.
Does a low editorial share mean editorial coverage does not work? #
No. It describes observed source selection in one category and release window. It says the editorial layer is not currently the dominant source layer engines draw on for AI visibility buying questions — which is a statement about the present composition of the answer, not about the value of the coverage.
How often does this change? #
The Index publishes daily. Each release carries its own identifier and artifact hash, and every figure in this study is stated against mri_score_v2.0+2026-09-18+8fa38e54dd0a. Current values for any category or question shape are readable from the live Index.