I Count the Research Record Before I Fund It

Before I fund or kill a research, standards, or technical-authority program as an AI visibility play, I check whether the scientific record shows up in that category's answers at all. The spread is large enough that the same program is the obvious spend in one market and a wasted citation budget in another.

The cross-category measurement is Four Categories One Source Class Wins; Nine Where None Does. That page is the deeper dataset. This essay is how I would use it as a founder: the count I run before the strategy conversation, the scope of what the count is allowed to decide, and the later release that would retire the rule.

The count, and the release it comes from

The Index sorts every cited domain into source classes and measures each subject category against six buyer question shapes: best tools, how buyers choose, is it worth it, problem-first research, top lists, and comparisons. Each pair is a segment with its own run denominator, published once it clears 10 observed runs across 7 distinct run dates. Those classes do not cover the whole board. Across the 840 top-ten slots in the 14 categories that publish all six shapes, 475 carry no class at all. I treat that unclassified share as its own decision, on its own page, because it changes what a class count can claim to cover.

The release I am reading was generated on 2026-09-26, the September 26, 2026 Machine Relations Index release, over the window 2026-05-10 to 2026-09-26: 23,978 cited source domains, 132,514 citation events, 16,925 answer runs against 1,010 monitored prompts, six engines, 95 of 151 measurable segments validated.

For five categories I took all six published leaderboards and counted how many of the 600 top-100 rows carry the academic and government label. That is a count of slots, not of distinct organizations. A source that wins three shapes is counted three times. Slots are what a buyer's answer is assembled from.

Deep Tech and Hardware and Enterprise Software are close neighbors on the measurement properties that could fake a difference. Their six segments measure 1,041 and 979 distinct domains. Run counts sit between 104 and 130 on one side and 99 and 114 on the other. Every segment in both has exactly seven run dates. The engines, the window, and the evidence floor are the same release. One of them draws on the scientific and standards record twenty-five times as often as the other.

Editorial publications sit between 26 and 53 rows in every one of the five categories. Journalism is present everywhere in this cut. What changes by market is whether anything from the research record is in the room.

Enterprise Software's two rows do not survive a look

Enterprise Software's two academic and government rows are cisco.com at rank 85 of 194 on "is it worth it" and suplari.com at rank 6 of 264 on top lists. Both are company domains. The label is a classifier output, and on these two rows it does not describe the organization behind the domain. The honest reading of that cell is at most 2 of 600, and on inspection closer to zero. The finding gets stronger when I check it.

The same kind of miss touches a few rows elsewhere: oreateai.com in deep tech and consumer health, bigcommerce.com in consumer health, lexisnexis.com in fintech, one slot each. That moves those counts by one or two and leaves the order of the five markets where it was. The academic rows in those four also include arxiv.org, nih.gov, sciencedirect.com, ieee.org, uspto.gov, researchgate.net, wikipedia.org, brookings.edu, and yale.edu. Those are what they say they are.

What the deep tech leaderboards actually contain

A preprint server is the second most cited source in the category when a buyer asks which tool is best. The patent office outranks most trade publications when a buyer asks how to choose. These are structural documents that answer a technical question directly. The same domain, the same category, the same release: arxiv.org is rank 2 on best tools and rank 65 on is-it-worth-it. Being the source engines reach for on one buyer question buys almost nothing on the next one.

Across all six deep tech leaderboards, 481 distinct domains appear, and 412 of them, 85.7%, appear in exactly one shape's top 100. Only reddit.com is in all six top tens. A program aimed at "deep tech authority" in the abstract is aimed at a board that does not repeat itself from question to question.

Two mistakes follow, and most companies make one of them. Enterprise software companies commission research programs into a market where the class barely exists. Hardware and deep tech companies treat technical documentation, standards work, and published papers as compliance overhead while those documents are already what the engines are quoting.

What sits beside the papers

The rest of the deep tech top tens are the trade's working infrastructure: analog.com, ti.com, avnet.com, arduino.cc, allaboutcircuits.com, anysilicon.com, semiengineering.com, eetimes.com, nvidia.com, plus reddit.com and youtube.com. In a market where the buying question is technical, the answer is assembled from technical primary sources: the datasheet, the standard, the paper, the application note, the forum thread where somebody built it.

That is the practice I mean by Machine Relations. The class names, including academic and government, are defined in the glossary. AuthorityTech is the agency I founded, and this count is the kind of read I want done before anyone proposes a research program as a visibility program.

What I would fund in each of the five

The table is a budget, once I stop averaging it.

Deep Tech and Hardware, 50 academic and government rows in 600 slots, is the market where I would fund papers, standards, and patent-facing documents as citation assets. arxiv.org at 17.36%, 21 of 121, rank 2 of 328 on best tools, and ieee.org at 19.05%, 20 of 105, rank 3 of 171 on comparisons, are already in the answer. uspto.gov at 6.15%, 8 of 130, on how buyers choose, says the patent office is in the consideration set. I would still fund press. Editorial holds 31 of those 600 slots, inside the 26-to-53 band that shows up in every category I counted. The papers do not replace the press. They are an extra 50 slots the other markets do not have. I would also refuse a single "thought leadership" URL aimed at all six shapes. Four hundred twelve of 481 deep tech domains in the top 100s, 85.7%, appear in exactly one shape. reddit.com is the only domain in all six top tens. A paper that wins best tools can sit at rank 65 on is-it-worth-it, which is where arxiv.org sits, 2 of 105 runs, 1.90%.

Consumer Health is the mixed case: 26 academic rows and 53 editorial rows, across 1,304 distinct domains. I would fund both, and I would expect the editorial line to be the larger of the two. nih.gov showing up in the deep tech top lists at 14.15%, 15 of 106, is a reminder that a government health domain can travel. I would not copy the deep tech paper budget into a consumer health plan without counting this category's own six boards. The health count is half the hardware count. That is still a program. It is not the same program.

Fintech at 8 academic rows and 26 editorial, across 838 domains, is where I get careful. Eight is not zero, and it is not fifty. I would not kill a research relationship that exists for regulators or for customers who never open an answer engine. I would not sell it internally as the AI visibility plan. The editorial 26 says a press line still has a slot count to work against. The academic 8 says I want to see the actual domains before I add a standards partnership to the citation budget.

AI Visibility and GEO, 6 academic rows and 35 editorial, across 1,870 domains, is my own market on this cut. The domain count is the largest of the five and the academic count is near the bottom. More sources being measured did not produce more research citations. I would not commission a whitepaper to move citation in this category. I would spend against the editorial 35 and against whatever the domain list shows once the class column is set aside. Six of 600 is the number I use when someone proposes an academic partnership as the lever.

Enterprise Software, 2 of 600, both of them company domains on inspection, across 979 domains and the same seven run dates as deep tech, is the clean no. cisco.com at rank 85 of 194 and suplari.com at rank 6 of 264 do not become a research program because a classifier filed them that way. Editorial still holds 37 slots, so press is not the thing I am cutting. The research-for-citation line is the thing I would make justify itself somewhere else: an analyst, a procurement committee, a regulator. The engines, the window, and the floor are identical to the category that produced 50 rows. The difference is the market.

The decision

I am not changing my own category on the strength of this cut. AI Visibility and GEO sits at 6 academic and government rows out of 600, so a research-authority play is not the lever in that market.

What I would do, before a strategy conversation about whitepapers, standards, patent publicity, or an academic partnership sold as AI visibility, is pull the category's six leaderboards and count the class. If the count is near zero, the program has to justify itself on some other outcome: analysts, regulators, procurement committees, customers who never touch an answer engine. Six fetches. Every Index segment page serves a markdown twin, so a category's six leaderboards are a few seconds of work.

The scope is where a company spends earned-authority effort for AI answer visibility. It says nothing about whether research is worth publishing for human readers. It is not a causal claim. These are observed citation rates over a fixed window against a fixed prompt set. A category's composition describes what engines did. It is not an instruction they follow. It also does not say how to build a target list. That decision turns on a domain's observation count rather than on its class.

The segment pages on this release report domain-level confidence as unavailable, so every figure above stays attached to its evidence counts: cited runs, observed runs, and run dates. Every segment cleared the floor of 10 observed runs across 7 distinct dates.

I would narrow this to a dated observation, and stop treating it as a property of the markets, if a later release shows the academic and government class rising into double figures in Enterprise Software or Fintech without a classifier change. A classifier change that refiles cisco.com and suplari.com would not count. The new rows would have to be research and standards domains, the way arxiv.org, ieee.org, and nih.gov are research and standards domains in the deep tech cut.

FAQ

Which categories did I count?

Deep Tech and Hardware, Consumer Health, Fintech, AI Visibility and GEO, and Enterprise Software. All six published question shapes each, on the release generated 2026-09-26, the September 26, 2026 Machine Relations Index release.

Why count slots instead of distinct organizations?

A buyer asks one question at a time. The top 100 of that question's leaderboard is the pool the answer is drawn from. A source that wins three of the six shapes occupies three positions in the pool a buyer can actually hit. Distinct organizations answer a different question, about how concentrated the class is. Both counts are useful. The budget question is the slot count.

Should a hardware company stop doing press because papers rank?

No. Editorial publications hold 31 of the 600 deep tech slots, in the same band as every other category in the table, 26 to 53. The research and standards record is an additional 50 slots. I would fund the papers and the press as different lines, against the shapes where each one appears. arxiv.org at rank 2 on best tools and rank 65 on is-it-worth-it is the reminder that one line does not cover the other shape.

How do I run this on a category that was not in the five?

Open the category on the Index, take each of its six question-shape pages, and count the source-role column. If the category is still collecting observations rather than published, nobody can tell you yet, including me. I would not fund the program on a general belief about credibility while the segment is unpublished.

What if the label is wrong on my row?

I open the domain. Enterprise Software's two "academic" rows were company domains, and the cell moved from 2 toward zero when I looked. A label that does not describe the organization is not a reason to fund a research partnership, and it is not a reason to ignore a real preprint server sitting at rank 2. The check is the domain, then the class.