Academic Sources Earn AI Citations in Only Some Categories

Founder Decision Ledger #6.

The decision is this: before you fund or kill a research, standards or technical-authority program for AI visibility, check whether the scientific record shows up in your category's answers at all. The spread between categories is large enough that the same program is obviously right in one market and obviously wasted in another.

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 I went looking for a general rule about academic and government sources. There is not one. What there is instead is a category effect so large that I had to check the arithmetic twice.

The ledger decision

Field	Decision
Decision	Treat "does peer-reviewed and standards material earn AI citations" as a per-category question with a measurable answer, never as a general belief about credibility.
Evidence	In the release mri_score_v2.0+2026-09-26+2b779408cfda, counting the top 100 of each of the six published buyer-question leaderboards per category, so 600 slots each: Deep Tech and Hardware holds 50 academic and government rows. Enterprise Software holds 2, and both of those are company domains. The two categories measure almost the same number of domains, 1,311 and 1,241, over the same seven run dates.
Scope	Applies to where a company spends earned-authority effort for AI answer visibility. It says nothing about whether research is worth publishing for human readers, customers or regulators.
Action	Before funding a whitepaper, standards, patent-publicity or academic-partnership program as an AI visibility play, pull your own category's six leaderboards and count the class. If it is near zero, the program has to justify itself on some other outcome.
Reversal condition	If a later release shows the academic and government class rising into double figures in Enterprise Software or Fintech without a classifier change, the effect is a moment in time rather than a property of those markets, and this decision narrows to a dated observation.

What I counted, exactly

The Index sorts every cited domain into one of nine source classes, and it 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 of those pairs is a segment with its own run denominator, published only once it clears the evidence floor of 10 observed runs across 7 distinct run dates.

The release I am reading was generated on 2026-09-26 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 organisations, and a source that wins three shapes is counted three times. That is deliberate: slots are what a buyer's answer is assembled out of.

The count

Category	Academic and government rows in 600 slots	Editorial publication rows in the same 600	Distinct domains measured across the six segments
Deep Tech and Hardware	50	31	1,311
Consumer Health	26	53	1,632
Fintech	8	26	1,107
AI Visibility and GEO	6	35	2,661
Enterprise Software	2	37	1,241

Read the first and last rows against each other. Deep Tech and Hardware and Enterprise Software are close neighbours by every measurement property that could explain a difference. Their six segments measure 1,311 and 1,241 distinct domains. Their 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 identical because it is one release.

And one of them draws on the scientific and standards record twenty-five times as often as the other.

Now read the middle column. Editorial publications are between 26 and 53 rows in every one of the five categories. The press is roughly as present everywhere. What changes by market is not whether journalism gets cited. It is whether anything from the research record is in the room.

Enterprise Software's two rows are not what the label says

I am not going to round this in my own favour. 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 organisation behind the domain.

So the honest reading of that cell is: at most 2 of 600, and on inspection closer to zero. The finding gets stronger when you check it, which is the only direction a finding is allowed to move.

I have filed the classifier behaviour where it belongs. It does not change the count in the other four categories, whose academic rows are 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 look like

Source	Question shape	Rank	Citation rate	Cited runs
arxiv.org	Best tools	#2 of 328	17.36%	21 of 121
ieee.org	Comparisons	#3 of 171	19.05%	20 of 105
nih.gov	Top lists	#4 of 279	14.15%	15 of 106
ieee.org	How buyers choose	#5 of 195	13.85%	18 of 130
wikipedia.org	Is it worth it	#8 of 181	11.43%	12 of 105
sciencedirect.com	Top lists	#9 of 279	11.32%	12 of 106
uspto.gov	How buyers choose	#23 of 195	6.15%	8 of 130
arxiv.org	Is it worth it	#65 of 181	1.90%	2 of 105

A preprint server is the second most cited source in the entire 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 not prestige effects. They are structural documents that answer a technical question directly, which is the same property I found when I pulled the academic class on its own and watched a free preprint server outrank Stanford.

Then read the last line against the first. The same domain, the same category, the same release: rank 2 on best tools and rank 65 on is-it-worth-it. Being the source engines reach for on one buyer question buys you close to 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.

So there are two separate mistakes available here, 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 their technical documentation, standards work and published papers as compliance overhead, while those documents are already the thing the engines are quoting.

What sits alongside the research record in deep tech

The rest of those leaderboards are not the business press either. The top tens are built out of the trade's own 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.

That is a coherent picture rather than a strange one. In a market where the buying question is technical, the answer gets assembled from technical primary sources: the datasheet, the standard, the paper, the application note, the forum thread where somebody actually built it. That is what Machine Relations is a discipline about, and it is the whole argument for treating machine-facing authority as its own practice rather than as a variant of publicity. At AuthorityTech, where I am CEO, that is the work we do for clients, and this is the kind of read we run before proposing a program to any of them.

What I am doing with this

I am not changing anything about my own categories on the strength of it. AI Visibility and GEO sits at 6 rows of 600, so the research-authority play is not the lever in my market, and I already knew where my market's leverage is.

What changes is a question I now ask first, before the strategy conversation: pull the category, count the class, then decide. It takes six fetches. Every Index segment page serves a markdown twin, so a whole category's six leaderboards are a few seconds of work and no credential.

What this does not say

It does not say research is wasted in enterprise software. Research reaches analysts, regulators, procurement committees and customers who never touch an answer engine, and none of that is measured here.

It does not say the count is causal. These are observed citation rates over a fixed window against a fixed prompt set, so a category's composition is a description of what engines did, not an instruction they follow.

It does not cover how to build a target list. That decision is a different one, and it turns on a domain's observation count rather than on its class.

And it is one release. The reversal condition above is written so that a later release can take this decision away from me.

FAQ

Which categories did you check? 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.

Why count slots instead of distinct organisations? Because a buyer asks one question at a time, and the top 100 of that question's leaderboard is the pool their answer is drawn from. A source that wins three of the six shapes genuinely occupies three positions.

Does this mean hardware companies should stop doing PR? No. Editorial publications hold 31 of the 600 deep tech slots, which is in the same band as every other category I checked. The point is that the research and standards record is an additional 50 slots that most hardware companies are not treating as a visibility asset at all.

How do I run this for my own category? Open the category on the Index, take each of its six question-shape pages, and count the source-role column. If your category is still collecting observations rather than published, the honest answer is that nobody can tell you yet, including me.

What about the confidence grade? The segment pages on this release report domain-level confidence as unavailable, so every figure above is stated with its evidence counts instead: the cited runs, the observed runs and the run dates are on every row, and every segment cleared the floor of 10 observed runs across 7 distinct run dates.