475 of 840 Top AI Citation Slots Belong to No Source Class

Founder Decision Ledger #7.

I have spent six entries of this ledger asking whether a source class pays in your category. Academic research. Vertical trade press. International wires. The HR shortlist. Every one of those entries asked the same shape of question, and this week I measured how much of the board that question covers. It covers a minority of it.

The decision is this: stop sizing an AI visibility budget by source class, because in all 14 categories the Machine Relations Index measures across every buyer question, the largest single block of the top ten belongs to no class at all. Read the actual domains in your category's six leaderboards before you fund a program built around a class.

I run the Machine Relations Index, so once a week I pull a slice of it and let it argue with something I believe. This week it argued with the last six weeks of my own work.

The ledger decision

Field	Decision
What I believed	A category's answers are won by a source class, so a class-level program is the unit of spend
What the data says	475 of 840 top-ten slots across 14 categories carry no source role. The four classes a founder can actually buy into hold 268 between them
What I changed	The first read is the domain list, not the class table. Class programs get funded per category, against the slots they can reach
What I did not change	Class share still decides whether a class program is worth anything. It just cannot tell you your ceiling
Cost of being wrong	Funding a press, research or community program sized to a board it reaches a fifth of

What I measured

Release mri_score_v2.0+2026-09-26+2b779408cfda, window 2026-05-10 to 2026-09-26. 16,925 answer runs across six engines, 1,010 monitored prompts, 23,978 cited domains, 132,514 citation events.

The Index splits buyer demand into six question shapes and scores every category-shape pair separately, publishing a leaderboard only once that pair clears 10 observed runs across 7 distinct run dates. Fourteen categories currently publish all six. That is 84 leaderboards. I took the top ten of each one, which is 840 slots, and counted the source role the Index assigns to the domain holding each slot.

I used the top ten rather than the whole leaderboard on purpose. A buyer reads an answer, not a distribution. The tail below rank ten is real and it is mostly noise. If a class matters to a buying decision, it has to be in the visible part.

HR & Talent publishes only four of its six shapes and is excluded. Every number below is read from the segment's own public page.

The board

Category	Unclassified	Editorial	Vendor-owned	Community
AI Infrastructure	19	9	8	12
Cybersecurity	23	4	19	8
Consumer Finance	27	16	5	6
AI Security & Privacy	29	7	6	8
AI Visibility & GEO	31	4	8	10
Enterprise Software	31	4	11	4
Deep Tech & Hardware	34	4	1	10
Consumer Health	35	7	0	5
Consumer Products	38	10	0	5
Family Software	38	15	0	6
Fintech	39	3	12	3
Emergent Prosumer	40	10	0	7
Education & Training	45	2	1	8
Healthcare Services	46	8	2	0
All 840 slots	475	103	73	92

56.5% of the visible board carries the label the Index calls Other observed source, which means the classifier has not filed it under any of the nine named types. Editorial publications hold 12.3%. Vendor-owned sources hold 8.7%. Community and social platforms hold 11.0%. The remaining 11.5% is academic and government, market databases, search platforms, analyst research and wire distribution combined.

Unclassified is the plurality in 14 of 14 categories. There is no category where the named classes, added together, are a comfortable majority of the top ten, and in seven of them they are a minority.

What this number is not

It is not news that the catch-all is large in aggregate. My own company published that three weeks ago: across the whole index, 189 of 506 rated domains carry no source role and together hold more cited runs than the entire editorial class. That brief also said the honest thing about its own number, which is that most of those domains are long tail, cited a handful of times, below the evidence floor, and no taxonomy will ever reach them all.

That caveat is why I ran this. It is the difference between "the tail is unclassified" and "the top is unclassified." The tail being unlabelled costs a founder nothing. Every slot in the table above is above the evidence floor, inside a published leaderboard, in the ten positions a buyer actually sees.

It is also not a claim that these domains are unreachable. Unclassified is a state of our classifier, not a property of the domain. Some of that 56.5% is your competitors' own marketing pages, which are extremely reachable, just not by buying anything. Some of it is a trade blog with four readers and a sitemap. Telling those apart is manual work, and it is the work.

And it is not an argument against class programs. It is an argument against sizing one without looking.

The category that made me check the arithmetic twice

Education & Training holds 45 of its 60 top-ten slots in the catch-all, and gives editorial publications 2.

Here is the whole visible answer to the buyer's shopping question in that category, the segment the Index calls Best tools, measured on 106 runs across 7 dates out of 263 ranked domains:

Rank	Domain	Source role	Citation rate
1	Reddit	Community and social platform	33.96%
2	D2L	Other observed source	18.87%
3	YouTube	Search or media platform	16.04%
4	CYPHER Learning	Other observed source	15.09%
5	Valamis	Other observed source	15.09%
6	BestColleges.com	Other observed source	14.15%
7	Coggno	Other observed source	14.15%
8	TalentLMS	Other observed source	14.15%
9	Docebo	Other observed source	13.21%
10	PrepScholar	Other observed source	12.26%

Six of those ten are learning-platform companies writing on their own domains, filed under no class. The first domain the Index recognises as an editorial publication is at rank 12. Forbes is at 29 with 6.60%. Gartner is at 58 with 3.77%. G2 is at 23.

A founder in that category who reads a class table sees a market with no press and concludes there is nothing to buy. A founder who reads the domain list sees six competitors' websites in the top ten of the question their buyers ask first, and a completely different set of moves. Those are not the same briefing, and only one of them is actionable.

One more thing from that category, because it is the kind of detail a class table erases entirely. In Education & Training head-to-head comparisons, the most-cited source is University of the People at 28.00%, filed as an academic and government source, in a category where it competes commercially. The label is right by the taxonomy and useless as a buying signal.

Where the classes do pay, and pay well

The spread in that table is the actionable part, and it runs 19 to 46 out of 60.

Cybersecurity gives vendor-owned sources 19 of 60 top-ten slots, the highest own-domain payoff of any category measured, and it has the smallest catch-all. If you sell security software, your own documentation is a citation asset with evidence behind it.

Consumer Finance gives editorial publications 16 of 60, and Family Software gives them 15. In those two categories a press program is not a brand exercise, it is the single highest-leverage line in the budget, because a quarter of the visible answer is held by publications you can pitch. Capterra and the database class matter there too.

AI Infrastructure is the most balanced board in the index: 19 unclassified, 12 community, 9 editorial, 8 vendor-owned. Every class pays something and no class pays enough alone.

Compare that to Fintech, where editorial holds 3 of 60 and vendor-owned holds 12, and the correct budget is close to the inverse of Consumer Finance despite the two being adjacent markets. Sizing one from the other is the specific mistake this ledger entry exists to stop.

What I actually do differently now

Three steps, in this order.

Open your category's six leaderboards and read the domains. Not the class column. The domains. Name every one in the top ten and say out loud what it is: a competitor, a trade publication, a forum thread, a database listing, a university. That list is your board.

Then ask what each one costs to reach. A publication has a price and a pitch. A competitor's own page has no price and no pitch, and the only move against it is a better page that answers the same question. A forum has neither and cannot be bought at all, which is a fact about the forum and not a reason to pretend it is absent.

Then size the class programs against the slots they can reach in your category, not against the market in general. In Consumer Finance that number justifies a serious press budget. In Education & Training it justifies almost none, and the same money goes to displacing six vendor pages.

Christian works the publication half of this, and the bound he found is the companion to mine: of 1,238 publications AI engines cite, 47 hold a top-ten position on any buyer question and 35 of those hold exactly one. His number tells you a pitch list has to be built per question. Mine tells you how much of the question a pitch list can reach at all. Both have to be true before a budget is real.

I named this discipline Machine Relations because the question is not how a page ranks, it is which sources a machine reaches for when a buyer asks. AuthorityTech is where I put that into practice, and the reason I publish a number that shrinks the addressable market for my own category of service is that a practice that cannot tell you when its product is the wrong first move is not worth hiring.

What would change my mind

If the classifier's coverage of the top ten improved materially and the catch-all fell below a third of the visible board, the class question would be most of the board again and this entry would be wrong. That is a tractable engineering problem and I expect the number to move. Until it does, the domain list is the primary read and the class table is the second one.

Questions I get about this

Does this mean AI visibility cannot be bought? No. It means the purchasable fraction is a category-specific number between roughly a fifth and two thirds of the visible board, and you can read it before you commit.

Why measure the top ten instead of the whole leaderboard? Because a buyer reads an answer, and an answer is short. A domain at rank 60 of 263 with a 3.77% citation rate is not in the room where the decision happens.

Is "unclassified" the same as "low quality"? No. Healthline is the 15th most-cited domain in the index and carries no source role. The label reflects how far our classification has reached, not how much an engine trusts the source.

How often does this change? The Index republishes on a rolling window, so the slots move. The category spread has been stable across the releases I have read, and it is the spread, not the individual slot, that a budget should be sized against.

What is the single number to take away? In your category, how many of the 60 visible slots are held by something you can actually buy. If nobody on your team can answer that, no visibility budget in the company is sized correctly.