Deep Tech AI Answers Cite Science Over Vendor Websites
Across six deep-tech buyer-question leaderboards, 53 of 600 ranked rows are academic or government and 9 are vendor-owned. Enterprise software runs 87.
If you market semiconductors, robotics, sensors or instruments, the AI-visibility advice written for software does not describe your category. It describes the opposite of it.
The Machine Relations Index publishes a ranked leaderboard for each subject category paired with each buyer-question shape, and Deep Tech & Hardware clears the evidence floor on all six shapes. Reading the top 100 rows of each of those six leaderboards gives 600 ranked positions in the category. Of those 600, 53 belong to academic and government sources: universities, journals, standards bodies, the patent office. Nine belong to vendor-owned sources.
Run the identical read on Enterprise Software, and the two counts swap: 87 of its 600 rows are vendor-owned and 5 are academic or government.
Same release. Same six question shapes. Same depth of leaderboard. Two categories that reward almost inverse media plans. And one detail the software playbook never mentions: every one of the six deep-tech leaders is a source all six measured engines cite.
What was measured
The October 14, 2026 Machine Relations Index release, generated 2026-10-14, observation window 2026-05-10 through 2026-10-14, 25,323 monitored answer runs index-wide. Six engines report runs inside that window: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.
A segment is one subject category paired with one buyer-question shape, and it publishes a ranked leaderboard only after clearing the Index's evidence floor. All six Deep Tech & Hardware shapes clear it, at 180 to 200 observed runs each across 10 run dates; the six Enterprise Software shapes run 146 to 154 each across 9. Runs are not independent questions: each deep-tech shape repeats 7 distinct questions (5 paraphrase groups) across engines and dates, so a rank describes the sampled questions, not the whole market. Rates below are the segment's own denominator, read from the segment's public page rather than recomputed.
The Index sorts every cited domain into one of nine source-role classes. Here is how the 600 ranked rows in each category divide.
| Source role | Deep Tech & Hardware | Enterprise Software |
|---|---|---|
| Other observed source, unclassified | 461 | 405 |
| Academic and government | 53 | 5 |
| Editorial publication | 32 | 37 |
| Community and social platform | 29 | 14 |
| Market and company database | 6 | 24 |
| Search or media platform | 6 | 2 |
| Vendor-owned | 9 | 87 |
| Analyst and consulting research | 2 | 25 |
| Wire and press-release distribution | 2 | 1 |
Two things in that table are worth a marketing leader's afternoon.
Your own domain is competing in the wrong weight class
The nine vendor-owned rows in Deep Tech & Hardware sit on eight domains: Nvidia twice, and seven smaller vendors and distributors, two of them inside a top ten. Nvidia's strongest showing is #2 of 238 on head-to-head comparisons at 16.67%, and Nvidia is the single most recognisable entity in the category. Enterprise software's 87 vendor-owned rows are spread across 55 distinct domains, ordinary vendors whose own sites answer their buyers' questions routinely.
The analyst row count says something similar about a second habit. Enterprise software carries 25 analyst-and-consulting rows, including Deloitte at #5 of 243 on how buyers choose at 11.56%, and Gartner in all six of its shapes. Deep Tech & Hardware carries two analyst rows, neither above 2.5%. Analyst relations is a real function with real value; in this category's measured answers it is not the source engines reach for.
One qualification belongs right here rather than in a footnote, because it changes what you do about it. The Index classifies Analog Devices and Texas Instruments under its unclassified "other observed source" bucket, not as vendor-owned, and both rank well: Analog Devices leads problem-first research outright and Texas Instruments is second behind it. So manufacturer sites do win these answers. They win them when the site is the primary reference for the question, the datasheet and the application note, rather than when it is the best-marketed page about a product. That is a different asset from the one a content programme produces.
What fills the space instead is the part that changes a budget. arXiv is cited in all six deep-tech shapes and now leads best tools outright, #1 of 480 at 20.00%. ScienceDirect is cited in five, reaching #7 of 373 on top lists at 10.75%. NIH appears in three, reaching #4 of 373 on top lists at 12.90%. IEEE appears in four, reaching #4 of 238 on comparisons at 14.52%. ResearchGate appears in four, and the USPTO in three, best at #13 of 260 on how buyers choose: engines citing a patent filing as the answer to a buyer's question about hardware. PatSnap, a patent-analytics database the Index leaves unclassified, is in all six shapes and inside the top eight of three.
The six leaders, and the engine column that decides your plan
| Buyer question | Leading domain | Source role | Citation rate | Engines citing it, index-wide |
|---|---|---|---|---|
| Best tools | arXiv | Academic and government | 20.00% | 6 of 6 |
| How buyers choose | AnySilicon | Other observed source | 35.68% | 6 of 6 |
| Is it worth it | VLSI Shuttle | Other observed source | 31.11% | 6 of 6 |
| Problem-first research | Analog Devices | Other observed source | 19.02% | 6 of 6 |
| Top lists | Unmanned Systems Technology | Other observed source | 27.42% | 6 of 6 |
| Comparisons | YouTube | Search or media platform | 17.20% | 6 of 6 |
Every leader is a source ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity all cite somewhere in the Index. The domain that is conspicuously not in that column is Reddit. It is cited in all six deep-tech shapes and sits inside the top four of four of them (#3 on best tools, #3 on how buyers choose, #4 on problem-first, #3 on comparisons), which is unusual for the most technical category the Index measures, and the reflex is to read it as an instruction to go build presence in engineering subreddits.
Read the engine breadth before you act on it. Reddit's domain profile records 4 observed engines across the window: Gemini, Google AI Mode, Google AI Overviews and Perplexity. arXiv's records 6, ScienceDirect's 5, NIH's 6, Semiconductor Engineering's 6, YouTube's 6, Analog Devices' 6. Reddit's confidence grade is A and so are arXiv's, NIH's and YouTube's, so this is not a data-thinness artefact.
The practical translation: a community-first plan in deep tech is a plan aimed at the Google answer surfaces and Perplexity. A plan built on peer-reviewed and standards-body corroboration, primary technical references and explanatory video reaches those and ChatGPT and Claude as well. If your buyers are the ones using the assistants you cannot see into, the second plan is the one that gets there.
The engineering trade press, with the caveat attached
The editorial publications ranking inside these six segments are a usable pitch list, and nearly every one of them is a specialist:
| Publication | Segment position | Segment rate | Engines citing it, index-wide | Index-wide confidence |
|---|---|---|---|---|
| AIMultiple | Best tools, #2 of 480 | 16.50% | 4 | B |
| EE Journal | How buyers choose, #4 of 260 | 14.57% | 4 | C |
| Electronic Design | How buyers choose, #5 of 260 | 14.07% | 4 | C |
| EE Times | How buyers choose, #6 of 260 | 13.07% | 4 | C |
| Power Electronics News | Comparisons, #12 of 238 | 9.14% | 5 | Collecting |
| Semiconductor Digest | How buyers choose, #23 of 260 | 5.53% | 3 | Collecting |
| EEVblog | Problem-first, #21 of 266 | 3.80% | 2 | Collecting |
| LIDAR Magazine | Top lists, #45 of 373 | 3.76% | 3 | Collecting |
The two right-hand columns are the honest part, and they are why this list is a shortlist rather than a ranking. Confidence grades in the Index are domain-level, not segment-level: Power Electronics News holds 29 cited runs across the whole 25,323-run index, which is why its index-wide grade reads Collecting even while it sits twelfth in a published deep-tech segment. EE Times, at 43 cited runs, grades C. Read those rows as evidence that these titles enter deep-tech answers, and not as a durable rate you can forecast against. arXiv, NIH, YouTube and Reddit are the grade-A domains here.
That distinction matters for where a pitch goes. A placement in EE Times is a placement in a publication that engines demonstrably pull into this category's answers, on four of the six engines. It is not, on this evidence, equivalent to a citation from a journal or a standards body that six engines read. AuthorityTech keeps the full per-shape list of which publications AI engines cite for deep tech and hardware buyers, rebuilt on every Index release.
What this changes for a deep-tech media plan
Three shifts, in the order they pay.
Move corroboration ahead of publication. In enterprise software, a well-structured page on your own domain is a live candidate for citation; 87 rows say so. In deep tech, 9 rows say it is not the lever. The claim about your product has to exist somewhere you do not own (a paper, a standard, a patent, an independent teardown, a specialist title) before an engine treats it as checkable.
Treat technical publishing as a distribution channel, not an academic indulgence. An application note, a measured benchmark, a conference paper, a patent filing, a recorded teardown: in this category those are the artefacts already sitting in the cited set, across all six engines. Engineering organisations often produce them and then fail to make them findable, canonical and citable. That is a fixable gap, and it is cheaper than the content programme that the software playbook would have you run.
Pitch the specialist press, and measure it per engine. The trade titles above are the editorial route into these answers. Because each is cited by two to five engines rather than six, a single blended visibility number will hide whether a placement reached the engine your buyers actually use.
What this does not say
It does not say deep-tech brands cannot be cited on their own domains. Analog Devices leads problem-first research at 19.02%, Texas Instruments ranks in five shapes and reaches #2 on problem-first at 13.59%, and Nvidia ranks in two. Note that the Index files Analog Devices and Texas Instruments under the unclassified "other observed source" bucket rather than vendor-owned, so the classified count of 9 understates how often a manufacturer's own site ranks. The sharper version of the finding is that it happens for component manufacturers whose datasheets and application notes are the reference material for the question, not because the site is well optimised, but because it is the primary source.
It does not explain why the two categories differ, and correlation is not offered as a mechanism. It does not extend past the top 100 rows of each leaderboard, past this release, or to the unclassified majority: 461 of the deep-tech rows sit in the Index's "other observed source" bucket, which is read here for its winners and is not compared as a class. Engine breadth is a domain's count across the whole Index, not within this category. A class-level study of how source-role leadership shifts by category, including the deep-tech split, is published separately on the neutral layer: source class capture by category. The neighbouring comparison for software categories is vendor sites winning cybersecurity but not AI infrastructure.
And it is a market finding, not a statement about any particular company's citation status. Your own category-and-shape leaderboards are public on the Index; read the six that match your buyers before acting on the six above.
How to check your own category
- Open your category on the Machine Relations Index and confirm all six shapes clear the evidence floor. A shape marked collecting is the instrument declining to publish a number, which is a different state from a low one.
- Count each leaderboard's rows by source role rather than reading the leader. The leader tells you who won once; the class distribution tells you what kind of source your category rewards.
- Check engine breadth on every domain you are planning around. A domain cited by four engines and a domain cited by six are different assets.
- Separate domain-level confidence from segment position before you forecast anything off a rate.
That work, building the independent corroboration, the extractable structure and the entity presence that put a brand inside a category's cited set, is what Machine Relations names as a discipline. AuthorityTech is the practice that applies it, and earned media in the titles a category's engines actually read is the part of it we run for deep-tech and hardware brands.
Figures above are read from the October 14, 2026 Machine Relations Index release, window 2026-05-10 to 2026-10-14, at machinerelations.ai/index on 2026-10-14. First published 2026-09-27 on the September 26, 2026 Machine Relations Index release; updated 2026-10-14 to the current release.