AI Recommends Life Sciences 2026: 40 Capped Category Panels and the Citations Behind Them

TL;DR

A commercial lead screenshots an answer to “best ultrasound machines.” Fourteen names. None of them is theirs. The instinct is to treat that list as a market census — or as a ranking the company failed. It is neither.

VayoMed Research computed 40 public “AI Recommends” category reports (541 stored recommended-brand slots; 415 unique names after light normalization) and 188 public brand reports, accessed 15 August 2026. The category snapshots run from 17 February to 6 June 2026; most were generated on 6 June. The recommended-brand panels are not raw model lists: a second Gemini extraction step combines four answers, sorts names by platform count and mentions, and stores at most 15. This is a tracked, capped dataset, not a live census of every life-sciences answer on the web [1][2].

What the measurements show:

• The stored panel is not the clearance register. Diagnostic-ultrasound product codes IYN, IYO and ITX contain 2,475 FDA 510(k) records and 620 unique applicants. The “best ultrasound machines” report stores 14 strings — and some of those strings are product names (Lumify, Soloscan, DRSONO), not manufacturers [1][3][4].
• Concentration has two shapes inside the capped panels. 358 of 415 unique names (86%) appear in exactly one tracked category panel. Ten normalized names occupy five or more stored slots: Philips 15, Siemens Healthineers 10, GE HealthCare 10, Medtronic 8, Mindray 6, Canon Medical Systems 6, Roche 5, Abbott 5, Epic Systems 5, NextGen Healthcare 5. Those are panel-occupancy counts, not market share [1].
• A stored 15-name panel is not 15-model agreement. 211 of 541 rows (39%) are attributed to one platform; 143 are attributed to three or more. Because the pipeline favors multi-platform names before imposing the cap, these shares describe the curated panel, not every name in the raw answers [1].
• Category-answer citations seldom point to a stored recommended brand’s domain. Of 1,689 citation URLs inside the four model responses, 209 (12.4%) resolve to a recommended brand’s own site. Eleven land in an “authority” bucket. Another 372 rows (22.0%) are vertexaisearch.cloud.google.com — Gemini’s grounding wrapper, not a publisher [1][5].
• The 12.4%/98.5% contrast is descriptive, not causal. A separate August 2026 press-release study of brand-name searches found 98.5% own-site citations. The studies use different query sets and sampling designs, so they show a sharp association with query context; they do not isolate query class as the cause [6].
• This dataset cannot test search volume against raw list length. Twenty-seven of 40 stored panels hit the pipeline’s 15-name ceiling. The aiSearchVolume field is a separate vendor metric, and counts at the ceiling are right-censored [1].
• The 188-brand panel is a pyramid, not a census. Mention sum 246,397. The top 10 brands hold 55.3% of mentions; the top 20 hold 73.4%. Twenty-four reports have zero mentions; 47 have fewer than ten. Visibility-score median 44 [1].

Google’s public description of AI Overviews and AI Mode is that they surface links, may use query fan-out, and use the same helpful-content and technical requirements as Search. There is no separate application form. We do not claim a ranking algorithm we cannot see [7].

The operational conclusion is narrow. A company missing from a “best X” screenshot has not been erased from the market. It is absent from one dated model answer or from this pipeline’s capped derived panel. The work that improves the public evidence available to buyers and retrieval systems — owned pages, source-grounded articles, a monthly public record, monitored queries — is the growth engine. It is not a promise to appear in ChatGPT.

1. What this dataset is — and is not

Takeaway: 40 dated category reports and 188 dated brand reports, already public. Screenshots are observations on generatedAt. They are not a ranking, a live crawl, or an industry census.

Every number in this article that is not attributed to FDA, Google, or a prior VayoMed Deep Research piece comes from one compute, accessed 15 August 2026, over two public collections [1][2]:

• 40 category reports under VayoMed’s “AI Recommends” set. Each stores a query (best ultrasound machines, best GLP-1 weight loss drugs, best FDA 510k consulting firms, and 37 others), a generatedAt stamp, a derived recommendedBrands panel, stored llmResponses from ChatGPT, Claude, Gemini and Perplexity, citation URLs inside those responses, and a separate sourceDomains field. The panel is produced by a second Gemini extraction pass, aggregated across platforms, sorted by platform count then mention count, and capped at 15.
• 188 brand reports on named companies. Each stores a visibility score (0–100 in the files), a mention total, and platform cards. These are the public panel at /ai-visibility, not a private audit dump.

Category generatedAt stamps cluster tightly. Thirty-one reports were generated on 6 June 2026, eight on 17 February 2026, and one (global ultrasound repair) on 2 June 2026. Answers are non-deterministic; a screenshot taken in August will not match a JSON file stamped in February, and we do not pretend otherwise. Brand reports include later stamps (several of the zero-mention files are dated 25 June 2026). Treat every list as observed on the stamp, not as current rank.

The stored model set inside llmResponses is four platforms, 40 responses each, 1,689 citation rows:

Citation volume is a platform habit. Claude emits nearly three times as many citation rows as ChatGPT on the same 40 queries. Mixing those rows into one pool is valid for asking “where did the stored answers point?” It is not valid for asking “which publisher is most important to Google.”

A second trap sits on the reports themselves. Some metrics.byPlatform cards list only Google AI Overview and ChatGPT. The stored answers underneath still include Claude, Gemini and Perplexity. We use the stored responses and their citation URLs. We do not treat the summary card as the full model set [1].

Google AI Overview is a real surface, and Google documents how AI Overviews and AI Mode relate to Search: they surface supporting links, may issue query fan-out across subtopics, and require the page to be indexed and snippet-eligible under ordinary Search technical rules. Specific extra optimization is not required; the same people-first content bar applies [7][8]. This dataset does not store AI Overview citation URLs the way it stores ChatGPT and Claude citations. Where we talk about “what the answers cite,” we mean the 1,689 URLs inside the four stored LLM responses.

The 40 queries are not a random sample of all life-sciences language. They are the category prompts VayoMed already tracks in public: imaging hardware, digital health software, diagnostics platforms, clinical-operations tools, regulatory services, CROs, a pharmaceuticals query, a biologics query. That selection is itself a finding. A “best X” engine will look more like a comparison-blog engine in categories where comparison blogs exist, and more like a drug-monograph engine where Drugs.com and WHO pages exist. The inventory:

Twenty-seven of 40 panels contain exactly 15 names because 15 is the storage ceiling. The observed floor is 6 (CT, MRI, GLP-1); median stored length is 15; range 6–15; 541 rows in total [1]. Counts of 15 are right-censored: the raw union may have been longer. Counts below 15 also remain dependent on the secondary extractor’s recall.

The public copies of these reports are the evidence surface, not the product. A reader can open the ultrasound category report and the underlying brand reports without buying anything [2][9]. What they cannot do is treat a June snapshot as a ranking, recover raw model-list length from the capped panel, or add the sourceDomains metric to citation counts — that last error is Section 6.

2. The shortlist is not the register

Takeaway: 620 unique 510(k) applicants in three diagnostic-ultrasound codes versus 14 strings in one derived panel. The panel is far smaller than the FDA applicant set, and it is not sampled from that set.

FDA 510(k) is a premarket notification. A substantial-equivalence order clears a device for commercial distribution; it does not approve it [10]. Diagnostic ultrasound systems and transducers typically travel under three classification product codes that appear together on the same submissions [4][11][12]:

All three are Class II radiology devices. We took the FDA’s downloadable 510(k) files and counted every row whose product code is IYN, IYO or ITX: 2,475 records, 620 unique applicant strings [3]. That applicant count is not 620 independent companies — subsidiaries, name changes and “Inc.” variants inflate it — but it is the public register of who has filed. It is the closest thing this industry has to a denominator.

The “best ultrasound machines” report, generated 6 June 2026, names 14 strings [9]:

Read that table as a pipeline artifact derived from four generations, not as a manufacturer census.

Company and SKU sit on the same panel. Lumify is a Philips handheld, stored against philips.com, attributed only to Perplexity. Soloscan and DRSONO are product-brand strings, not the applicant names a 510(k) row would carry. GE and GE HealthCare are the same company, split because our normalizer collapses “GE Healthcare” / “GE HealthCare” but leaves a bare “GE” untouched. Butterfly and Butterfly Network are the same company, counted twice. Canon appears as “Canon,” not “Canon Medical Systems.” These are tokens recovered by a secondary extractor, not a cleaned OEM register.

Consensus is thin. Only Mindray is named by all four stored models. Clarius is named by three. Five of the 14 strings are single-platform. A buyer who screenshots Claude and a buyer who screenshots ChatGPT do not see the same fourteen names.

The stored citations do not point to the 510(k) register. The stored ChatGPT answer cites a Medical Design & Development product announcement and two Reddit threads. The stored Claude answer cites handheld roundups on todopocus.com, marcroftmedical.com, ultrasoundportables.com, elzhen.com, pref-med.com, and DRSONO’s own “Top 5 Clinical Ultrasound Machines” blog. The stored Gemini answer’s citation URLs are almost entirely vertexaisearch.cloud.google.com/grounding-api-redirect/... wrappers whose display titles point at refurbished-equipment and dealer sites (sentinelimaging.net, heartmedical.com, probomedical.com). Those attached citations expose a comparison/dealer neighborhood, not the applicant column in PMN96CUR.ZIP; they do not prove which unexposed sources influenced generation [3][9].

We matched the 14 strings loosely against 510(k) applicant names. Several incumbents match (Mindray, Philips Ultrasound, Siemens Medical Solutions, Samsung Medison, GE, Clarius, Butterfly Network, FUJIFILM Sonosite). Lumify, Soloscan and DRSONO do not match as applicant strings — because they are not applicant strings. We do not claim the models “ignore 606 ultrasound makers.” We claim something narrower and more useful: the stored panel is far smaller than the FDA applicant set, and it is not sampled from that set.

A wider net on surgical robots (robot / computer-assisted surgical classification names) produced 3,170 510(k) rows and 1,004 applicants against 15 AI names. That net is too inclusive — it pulls devices that are not what a hospital buyer means by “surgical robot” — so we do not headline it. Directionally it rhymes with ultrasound: the answer is a short SKU-and-platform list, not a clearance census.

The adjacent Deep Research question is what happens after clearance on the company’s own site. In the July 2026 commercial-readiness audit, 56 of 60 first-time non-U.S. 510(k) applicants could be connected to an official domain, but the newly cleared product term was found on the bounded page set for only 27 of 60, and a K-number for only three [13]. A company can be on the register, have a website, and still give a model nothing citable to retrieve. That is the other half of why the shortlist does not look like the census.

The portable-ultrasound sister query, generated the same day, names 15 strings and repeats the same blur: Mindray, Philips, Clarius, Butterfly Network, Butterfly, GE HealthCare, GE, plus Exo, Sonoscanner, Clarius Mobile Health, Philips Healthcare, Fujifilm Sonosite, DRSONO, EchoNous, Suresult. Three labels for Clarius. Two for Philips. Two for Butterfly. Two for GE. The model is listing what it retrieved from handheld roundups, not collapsing a corporate tree [1].

3. Concentration has two shapes

Takeaway: Within the capped derived panels, 86% of unique names appear in one tracked category and ten names occupy five or more slots. This measures panel recurrence, not company specialization or market share.

Across 40 queries there are 541 recommended-brand slots. After collapsing a short list of obvious pairs — Philips / Philips Healthcare, GE Healthcare / GE HealthCare, Siemens / Siemens Healthineers, Abbott / Abbott Laboratories, Roche / Roche Diagnostics, Canon / Canon Medical Systems, Epic / Epic Systems — that is 415 unique names. 358 of them (86%) appear in exactly one tracked category. 57 appear in two or more. 10 occupy five or more slots [1].

The 358 one-category names form a long tail inside this panel: handheld SKUs, billing-software vendors, 510(k) consultancies and digital-pathology platforms. Some may be specialists; the capped data cannot establish that. A name may also appear once because the extractor missed a variant or because another occurrence fell below the top-15 cutoff. The safe conclusion is recurrence, not corporate focus.

The repeating names are the spine.

The slot count and the query count diverge because the same report sometimes lists both “Philips” and “Philips Healthcare,” both “Epic” and “Epic Systems,” both “Abbott” and “Abbott Laboratories.” Light normalization collapses the string and still double-counts the slot when a single query emits both variants. Philips’s 15 slots therefore span 12 distinct queries: CT, digital pathology, ultrasound repair, medical-equipment companies, MRI, PACS, patient monitoring, portable ultrasound, top device companies, ultrasound machines, ventilators, X-ray. That is still a cross-category imaging-and-monitoring franchise. It is not 15 independent markets.

Abbott’s five slots span three queries. Epic’s five span three software queries. NextGen’s five are five distinct digital-health queries with no within-query double listing. The chart therefore uses a stored-slot axis, not a category axis. A few slots are name-variant duplicates inside one panel; the distinct-query column exposes that effect.

Just below the five-plus line, the two-to-four-slot names tell a cleaner product-family story. Tebra and AdvancedMD each appear in four digital-health queries (EHR, healthcare software, medical billing, practice management). Johnson & Johnson appears in equipment, surgical robots and “top device companies.” Thermo Fisher and Illumina appear in lab-equipment and molecular-diagnostics queries. Clarius, Butterfly Network and DRSONO each appear in both ultrasound queries — the SKU layer repeating next to the OEM layer. Emergo by UL appears in both consulting queries. That is adjacency, not world domination.

Three method limits belong next to this chart, not in a footnote. First, normalization is incomplete: Butterfly versus Butterfly Network, GE versus GE HealthCare, Stryker versus Stryker Corporation, athenahealth versus Athenahealth still sit as separate names. Second, we did not merge SKUs into parents (Lumify → Philips, Hugo RAS → Medtronic, Mako → Stryker). Third, every panel was sorted and capped before this cross-category count. We leave the extracted strings visible because they are what the published reports store, not because they reproduce every raw answer exactly.

The commercially literate reading is not “AI prefers the West.” The data show an association: recurring names sit in answers whose citations include comparison publishers, dealers and software roundups. The corpus does not prove that those pages caused panel inclusion, and the top-15 sort itself favors names recovered from more platforms.

4. Consensus names versus single-platform names

Takeaway: In the stored top-15 panels, 211 of 541 rows (39%) are attributed to one model and 143 to three or more. Median single-platform share per panel is 40%. These values are descriptive after a sort that favors multi-platform names.

Each recommended-brand row stores mentionedBy — which of the four stored models named that string. Pooling 541 rows [1]:

The median category has 40% of its list coming from a single model. The mean is in the same neighborhood. Extremes are more informative than the average.

High single-platform share inside the capped panel. best FDA 510k consulting firms: 12 of 15 rows are single-platform; only one reaches three-plus. best pharmacovigilance software: 12 of 15. global ultrasound repair: 11 of 15, with zero three-plus rows. best PACS systems and best practice management software: 10 of 15. Endoscopy systems: 9 of 15 and zero three-plus. These panels show low cross-model overlap among the rows retained; they do not show that a model “padded” to 15.

Low single-platform share. GLP-1 drugs, insulin pumps, CT scanners, MRI scanners, molecular diagnostics, lab-equipment suppliers, CROs and regulatory consultancies have zero single-platform rows in the stored panels. Those panels are shorter (6–13) and show more agreement. The dataset can describe that pattern; it cannot determine whether a “public canon,” extractor behavior or answer format caused it.

Ultrasound sits in the middle: 5 of 14 single-platform, 2 of 14 with three-plus (Mindray and Clarius). A founder who is missing from “the ChatGPT list” may be present on Gemini, or present as a SKU string, or present on none. One screenshot is a sample of one model on one day.

This is also why “how do I get into ChatGPT” is the wrong specification. ChatGPT’s 40 stored answers carry 251 citations; Claude’s 40 carry 741. The union of four models is a different object from any one model’s list. Paying to chase a single screenshot optimizes a draw from a non-deterministic union.

Google’s own documentation already tells site owners not to expect a separate AI-Overview application: eligibility is indexation plus snippet eligibility plus helpful content [7][8]. The four LLM lists in this file are even less of an application process. They are generations over whatever the retriever returned.

5. What the stored answers actually cite

Takeaway: 12.4% of category-answer citations resolve to a stored recommended brand’s own site; 22% are Gemini’s Vertex wrapper; authority domains are 11 rows. A separate branded-query study found 98.5% own-site. The contrast is large but not a controlled causal test.

The 1,689 citation rows are URLs the stored model answers actually attached. We bucketed them with a domain heuristic: Gemini’s Vertex wrapper is classified first; then a host matching a recommended brand’s stored domain counts as own-site; then short lists cover authority, journal, wire, review-site, Wikipedia, YouTube, Reddit, social and comparison-blog hosts; everything else is other_web [1].

Three readings, in order of how much they should change a commercial plan.

First, own-site share is 12.4% here and 98.5% in a separate branded-query study. In the August 2026 press-release benchmark, a 1,440-keyword sweep across six generative surfaces for one life-sciences brand put 451 of 458 citation rows (98.5%) on that brand’s own website. The four wire citations in that sweep occurred on queries for the company’s name; unbranded category queries in that study produced zero release citations [6]. The comparison is informative, but the datasets differ in brand, query set, surfaces and sampling. It supports a hypothesis about query context; it does not prove that changing only the query class causes the share to move from 12.4% to 98.5%.

That contrast is the reason this article exists next to the PR paper rather than instead of it. Wire distribution creates a stable public copy of approved language. In these observations, wire URLs are rare in “best X” citations, owned pages dominate the separate branded-query sample, and third-party comparison hosts are common around category answers. That is a planning signal, not a guarantee that publishing any one page changes a model’s output.

Second, the single most-cited host is not a publisher.

372 of 1,689 (22.0%) are Gemini grounding redirects on vertexaisearch.cloud.google.com. Google documents Grounding with Google Search as a retrieval step that returns supporting links; the stored Gemini responses in this file expose those links as grounding-API redirect URLs, with a display title that sometimes names the underlying publisher [5]. Counting the wrapper as “Google is the publisher” is a misread. Counting it as 372 extra FDA pages is a worse misread. It is Gemini’s citation transport. Strip it, and the remaining host list is comparison sites, market-research shops, a used-imaging dealer, Wikipedia, YouTube, and a training publisher.

The named hosts under the wrapper are the actual retrieval neighborhood for category answers:

• medsoftwares.com (21) and getreskilled.com (15) are software-education / vendor-list publishers.
• verifiedmarketresearch.com (20), marketsandmarkets.com (18) and researchandmarkets.com (9, just off the chart) are paid market-research pages that rank for “best [category] market.”
• blockimaging.com (19) is a used-and-refurbished imaging dealer that publishes buying guides. Ultrasound and CT answers borrow dealer content.
• ultrasoundportables.com (9) and sonoransurplus.com (9) are the same pattern in handheld and surplus channels.
• quanticate.com (14) and elexes.com (9) are recommended brands whose own sites appear as citations — the 12.4% bucket in miniature, concentrated in consulting and CRO queries where the “best X” page is often written by a firm on the panel.
• drugs.com (12) appears because the GLP-1 query is a drug-monograph query, not a device-listicle query.

The labelled comparison_blog bucket is only 12 rows because the host list in the classifier is short. Many listicles landed in other_web. After separating 372 wrapper rows, 976 residual rows remain: dealer blogs, SaaS roundups, vendor-written “top 10” posts (including DRSONO citing drsono.com), and unlabeled hosts. Hand-labeling them would change the pie slices. The supported headline is descriptive: third-party pages are common in this category-answer corpus.

Third, authority is scarce, and the 11-row bucket is mostly PMC. The classifier treats any host containing nih.gov, fda.gov, who.int and similar as authority before it tests for journals. Of the 11 authority rows, nine are pmc.ncbi.nlm.nih.gov, one is accessdata.fda.gov (FDA’s list of recognized 510(k) third-party review organizations, cited on the consulting query), and one is WHO’s GLP-1 Q&A [14][15]. After that rule, three rows remain in peer_review — all nature.com. FDA.gov does not appear on the host leaderboard. A 510(k) clearance page is almost invisible as a citation on “best ultrasound machines,” which is consistent with Section 2: the models are not reading the register.

The 12.4% own-site figure is a domain-suffix heuristic, not a legal-entity match. It may under-count regional or distributor domains and over-count unrelated hosts on broad domains. We corrected a prior ordering bug by classifying the Vertex wrapper before brand-domain suffix matching; that moved seven wrapper URLs out of the earlier own-site/residual allocation. The corrected count is 209 of 1,689.

News and wire together are 21 rows (1.2%). That is real and small. It matches the PR paper’s observation that syndication-tail domains did not appear as AI citations, while owned pages did — on branded queries [6]. Category queries do not suddenly start citing PR Newswire copies either.

Google’s site-owner guidance remains the right ceiling for what anyone can honestly claim: appear in Search under ordinary rules; AI features may then surface supporting links; measure in Search Console; do not invent a separate ranking product [7][8].

6. sourceDomains is not a citation

Takeaway: YouTube’s 15,444 sourceDomains metric and 1,689 citation URLs are different fields. Adding them invents a finding. The exact meaning and collection method of the vendor-provided sourceDomains metric are not established by these JSON files, so this report does not relabel it as open-web popularity.

Each category report stores sourceDomains[] with a mentions count. Summed across 40 files, that overlay is dominated by large consumer and employment hosts [1]:

The field’s next hosts are Reddit (about 8,800), Indeed, Wikipedia, LinkedIn, ZipRecruiter, Forbes, GoodRx, BLS and PMC. Without vendor documentation in this research bundle, those values cannot be interpreted as search volume, citation count or a complete map of where a topic is discussed. They are only a separate stored metric and are not URLs attached to the answers.

YouTube as a citation host in the LLM responses is 18 rows (1.1%). YouTube as a sourceDomains overlay is 15,444 mentions. Mixing those sentences produces “AI cites YouTube 15k times,” which this file does not show. Reddit is 9 citation rows versus thousands of overlay mentions. Indeed and ZipRecruiter barely exist as citations and dominate the overlay because “best healthcare software” and “best practice management software” leak into job-board language.

We keep the two fields in separate charts because the failure mode is attractive. The vendor metric is large; the citation layer is the smaller, directly inspectable set of URLs attached to stored answers. That set includes comparison hosts, 372 wrappers, and a 12.4% own-site bucket.

If a vendor shows you a dashboard tile labelled “AI mentions” in the tens of thousands, ask which field it is. If the top domain is YouTube and the query is “best ultrasound machines,” you are probably looking at an overlay. If the top domain is a dealer blog or a Vertex redirect, you are looking at citations.

7. The 188-brand mention pyramid

Takeaway: In VayoMed’s public panel of 188 brand reports, the top 10 names hold 55.3% of 246,397 tracked mentions. Twenty-four reports have zero. Being in the panel is not the same as being visible, and being visible on a brand report is not the same as making a category shortlist.

The 188 files are public. They are also a panel: companies VayoMed already tracks, weighted toward firms that are already in the corpus. Thermo Fisher at 42,177 mentions does not mean Thermo Fisher is 17% of the life-sciences industry. It means Thermo Fisher is 17% of this panel’s mention sum (42,177 / 246,397). Panel composition is a method finding, not a nuisance [1][2].

Top 20 share is 73.4%. Visibility-score median is 44 on the files’ 0–100 scale. Scores are observed values in a tracked snapshot, not ranks we sell.

The next ten include QIAGEN, Boston Scientific, Agilent, Stryker, GE HealthCare, Bio-Rad, Becton Dickinson, ResMed, Grifols, Fujifilm. The overlap with the category incumbent spine is partial: Philips, Medtronic, Abbott, Roche, GE HealthCare appear in both. Thermo Fisher and Quest dominate the brand-report pyramid because the panel includes life-sciences tools and diagnostics names that the 40 “best X” queries only graze. Eli Lilly appears here because GLP-1 is in the category set and because Lilly is a tracked brand; Lilly is not a 15-category imaging incumbent. Different question, different concentration.

The 24 zeros are public. On the 25 June 2026 stamp they include SIUI, Landwind, Sonostar, Aohua Endoscopy, Wandong, Jafron Biomedical, Shandong Weigao, Sansure Biotech, Getein Biotech, and others — many of them China-based device and IVD manufacturers that are in the panel because they are FDA-relevant commercial entities, not because English-language AI answers already discuss them [16]. A zero in this window is not a moral judgment. It is a measurement: on the tracked queries behind that report, the stored surfaces did not mention the name.

The more precise commercial point is the gap between a brand report and a category shortlist. Chison Medical’s public brand report records 137 mentions and a score of 45 — not a zero — and Chison still does not appear among the 14 ultrasound name strings [17][9]. SonoScape’s public report records 4 mentions and a score of 15, and SonoScape is also absent from that shortlist [18]. SIUI’s public report records 0. Three imaging OEMs, three different mention intensities, one shared outcome on “best ultrasound machines”: not named. Clearance and panel membership do not mint a category-list slot.

We do not publish unpublished named-account scores in this article. Every company named in this section already has a public report. The pyramid is what the public panel looks like. A company that is not in the 188 is not “below zero”; it is off this particular sheet.

8. Why this dataset cannot test volume against list length

Takeaway: Twenty-seven of 40 stored panels hit a 15-name ceiling imposed by the report-generation pipeline. Those observations are right-censored, so they cannot support a claim that search volume does or does not predict raw answer-list length.

The tracked aiSearchVolume field is a vendor metric stored with each snapshot. The research bundle does not establish its collection method well enough to treat it as traffic. More importantly, the recommendedBrands pipeline takes a secondary Gemini extraction from each model answer, aggregates names, sorts by number of platforms and mention count, and returns brands.slice(0, 15). Therefore a stored count of 15 means “15 or more recovered before the cap,” not “the models produced exactly 15.” The table is useful as a data-quality audit, not a correlation test [1]:

The GLP-1 panel stores six high-agreement rows and the consulting panel stores 15 rows, twelve single-platform. That contrast is real inside the published panels. What is not known is the uncapped union for consulting, the secondary extractor’s recall for GLP-1, or whether the two prompts elicited comparable answer formats. The citation neighborhoods also differ: GLP-1 includes Drugs.com, WHO, newsrooms and Nature, while consulting answers include firm pages and “best of” posts [14]. Those observations can guide qualitative review; they cannot identify volume as the driver.

For 510(k) consulting, the published panel is a full-looking but censored 15. The pipeline may have recovered more names and discarded them after sorting. Its 12 single-platform rows still show low overlap among retained names, but “padding” is not observable from this artifact alone.

Healthcare software stores 15 names with only three three-plus-platform rows. Practice-management software stores 15 with zero three-plus rows. This supports a claim about retained-panel agreement, not about how demand changes list length.

CT and MRI store six names each with no single-platform rows. Ultrasound stores 14 with five. The plausible explanations include answer format, extractor behavior, name normalization and the available source set. This dataset does not distinguish among them.

For a company missing from a panel, the first diagnostic is therefore the pipeline: inspect the raw answers, preserve all extracted names, record per-model order, and rerun without the 15 cap. Only then should the team test whether the vendor volume metric, answer format, category type or citation neighborhood is associated with list length. Until that rerun exists, this article treats volume and stored panel size as separate descriptors.

9. What a company that is not on the list actually builds

Takeaway: Category answers cite third-party pages. Branded answers cite the owned site. The work is to make both layers true — website, source-grounded pages, a monthly public record, monitored queries — not to buy a ChatGPT ranking. VayoMed runs that as one $24,000/year Done-for-You engine. Cadence is guaranteed. Appearance is not.

Start from the screenshot. The missing company is usually doing one of four things, and the four are distinguishable in this dataset plus two papers we have already published.

It is on the register and off the shortlist. Ultrasound’s 620 applicants versus 14 strings. That is the normal state for most cleared devices [3].

It has a website that does not yet carry the commercial evidence a retriever can use. The after-510(k) audit found official domains for 56 of 60 first-time non-U.S. applicants, contact routes on 46 of 49 reachable sites, and the newly cleared product term on only 27 of 60 bounded audits. K-numbers appeared for three [13]. A model asked for “best ultrasound machines” will not retrieve a homepage that never names the system, the indication, or the clearance. Google’s AI-features guidance is blunt on the same point: important content needs to be available as text; there is no extra AI-Overview form [7][8].

It is in the brand panel and still off the category list. Chison (137 mentions), SonoScape (4), SIUI (0): three public imaging reports, zero occupancy on the 14-name ultrasound shortlist [17][18][16][9]. Brand-query visibility and category-query occupancy are different measurements. Optimizing only the former — “mentions of our name” — can leave the “best X” layer untouched.

It ships announcements onto a wire and expects those copies to become category citations. The PR study measured the opposite on branded queries: 98.5% of citations went to the owned site; zero of 307 syndication-tail domains appeared in sampled citation evidence. Wire copies earned citations when someone already searched the company by name [6]. This dataset’s category citations put news/wire at 21 of 1,689. Distribution is a public-record tool. It is not the category-listicle layer.

What, then, does a missing company actually publish if it wants to change the citation neighborhood this file measured? The neighborhood is specific: dealer buying guides, handheld roundups, software “top 10” posts, market-research landing pages, Wikipedia, an occasional PMC review, and the brand’s own domain when that domain already has a page the retriever can quote. The company does not control Block Imaging’s editorial calendar. It does control whether a citable, source-grounded page exists on its own domain in the same shape those hosts already quote — system family, indications, form factor, service footprint, clearance language that is accurate (cleared, not approved) [10] — and whether a third-party public record repeats the same entity facts so a branded query has somewhere to land.

That is a system, not a ranking hack. The loop is the one VayoMed already runs as six services, not as an AI-visibility SKU [19][20]:

The six services are the website build, website operations, LinkedIn page and content, domain portfolio, the content engine and GEO monitoring, and monthly PR. LinkedIn and domains do not appear as citation hosts in this file; they are still part of the entity-consistent public record a buyer (and a model) can reconcile. We do not claim they “get you into ChatGPT.”

The coordination-tax paper rebuilt the in-house alternative from BLS medians at $220,888 a year for a realistic 1.5 FTE configuration, and stated the honest limit: most companies never spend that; they assign fragments of the work to people who already have jobs [21]. $24,000 does not buy 1.5 FTE of attention. It buys a productised system. Deep regulatory, clinical and MLR work stays with the client. Content follows the client’s review path; it does not replace it [20].

Two promises this dataset forbids, stated as such:

• We will not guarantee a slot on a 15-name list, an AI recommendation, a ranking, traffic, or leads. The lists move. 39% of slots are one-model names. Snapshots are dated.
• We will not redefine VayoMed as an AI-visibility vendor because this week’s paper is about AI Recommends reports. The reports are a proof surface. The offer is the engine.

A company that wants only a dashboard of the 40 queries can read the public reports for free [2]. A company that wants the public-evidence loop operated — site, articles, PR cadence, monitoring — is the $24,000 conversation. The screenshot of someone else’s shortlist is the symptom. The missing citable pages are the work.

10. Methodology and limitations

Category corpus. 40 JSON category reports in VayoMed’s public AI Recommends set, accessed 15 August 2026. generatedAt ranges from 17 February 2026 (8 reports) to 6 June 2026 (31 reports), plus 2 June 2026 (1 report). Fields used: query, metrics (including aiSearchVolume and byPlatform cards), recommendedBrands[] (name, domain, mentionedBy, mentionCount), llmResponses[].citations[], sourceDomains[]. Public copies: vayomed.com/ai-visibility [1][2]. recommendedBrands is built by a secondary Gemini extraction from each of four answers, then aggregated, sorted by platform count and mention count, and capped with slice(0, 15). It is a derived panel, not a verbatim or complete list of names in the raw answers.

Brand corpus. 188 public brand reports, excluding the best/ collection. Mention sum 246,397; visibility scores as stored (median 44). Several zero-mention reports carry a 25 June 2026 stamp. Panel, not census.

Normalization and censoring. Regex collapses a short list of corporate pairs (Philips, GE HealthCare, Siemens Healthineers, Abbott, Roche, J&J, Boston Scientific, Medtronic, Intuitive, Stryker, Olympus, Canon Medical Systems, Mindray, FUJIFILM, Epic Systems). Incomplete: Butterfly / Butterfly Network, GE / GE HealthCare, Stryker / Stryker Corporation, athenahealth / Athenahealth remain split. SKUs are not merged into parents. Five-plus counts are normalized stored-slot occupancy; Philips 15 slots span 12 distinct queries because some reports list both “Philips” and “Philips Healthcare.” Twenty-seven panels hit the 15-row ceiling, so prevalence and single-platform shares are conditional on the retained top-15 rows and biased toward multi-platform names.

Citations. 1,689 URL rows from ChatGPT (251), Claude (741), Gemini (373), Perplexity (324). Host bucketing is a suffix heuristic plus keyword lists. Gemini’s vertexaisearch.cloud.google.com wrapper is classified before brand-domain matching (372 rows); recommended_brand_site is 209 (12.4%); other_web is 976. authority (11) fires on nih.gov before the journal rule, so nine of 11 are PMC, one is FDA accessdata, one is WHO. comparison_blog is a short allow-list. sourceDomains values are never added to citation counts, and this report does not infer their undocumented collection semantics.

FDA contrast. Downloadable 510(k) files, product codes IYN, IYO, ITX: 2,475 rows, 620 unique applicant strings, accessed 15 August 2026 [3][4]. Applicant strings are not de-duplicated corporate families. Surgical-robot classification-name net (3,170 rows / 1,004 applicants) is recorded as too wide to headline.

Platform cards. Some reports’ metrics.byPlatform list only Google AI Overview and ChatGPT. Stored llmResponses still include four models. AI Overview citation URLs are not stored in parallel with the LLM citation arrays; we do not invent them.

Limitations. Snapshots are dated; generations are non-deterministic; recommended-brand panels are second-model extractions capped at 15; scores are not rankings; 188 brands are a biased panel; 40 queries are VayoMed’s tracked set, not all life-sciences language; we did not hand-label 976 residual hosts; we do not claim concentration is unfair, causal or purchased. Google does not publish the ranking function behind AI Overviews or AI Mode [7].

Frequently asked questions

Why does ChatGPT keep recommending the same medical device brands?

In 40 tracked “best X” queries, 358 of 415 unique names appear in only one capped category panel, while ten normalized names occupy five or more stored slots — Philips, Siemens Healthineers, GE HealthCare, Medtronic, Mindray, Canon Medical Systems, Roche, Abbott, Epic Systems, NextGen Healthcare [1]. The answers cite comparison blogs, dealer guides, market-research pages and a Gemini wrapper, not the FDA applicant file. This shows recurrence in a derived panel, not why any model selected a name.

How do I get my medical device cited in ChatGPT or Google AI Overviews?

Google’s published rule for AI Overviews and AI Mode is the Search rule: indexed, snippet-eligible, helpful people-first content. There is no separate application [7][8]. In this dataset, category answers cited a stored recommended brand’s own site 12.4% of the time; a separate branded-query PR study found 98.5% [1][6]. The practical work is citable owned pages plus a third-party public record, measured on dated snapshots — not a purchased slot. VayoMed will not sell a guarantee of appearing in ChatGPT.

Are AI “best ultrasound” lists the same as FDA-cleared ultrasound manufacturers?

No. IYN / IYO / ITX contain 2,475 510(k) records and 620 unique applicants; the 6 June 2026 ultrasound report stores 14 strings, mixing OEMs (Philips, GE HealthCare, Mindray) with SKUs (Lumify, Soloscan, DRSONO) and double-counting Butterfly / Butterfly Network and GE / GE HealthCare [3][9]. A 510(k) order clears a device; it does not approve it, and it does not place the applicant in a derived recommendation panel [10].

Why did my competitor appear on Gemini but not on ChatGPT?

Within the capped derived panels, 211 of 541 rows (39%) are single-platform and the median panel’s single-platform share is 40% [1]. Ultrasound’s 14 strings include five one-model rows. Consulting and pharmacovigilance panels are 80% one-model. One screenshot is one model on one date; the stored panel is a secondary extraction across four answers and is a different object.

Do press releases get you into AI answers for “best medical devices”?

On branded queries, the August 2026 PR study found 98.5% of citations on the brand’s own site and zero of 307 syndication-tail domains in sampled citation evidence [6]. On category queries in this file, news/wire is 21 of 1,689 citations (1.2%) [1]. A monthly release is a durable public record of approved language. It is not the comparison-blog layer that “best X” answers borrow. VayoMed’s PR guarantee is production and distribution cadence, not coverage or AI placement [19].

What does VayoMed actually sell if it cannot guarantee an AI recommendation?

A Done-for-You Global Growth Engine: website build and operations, LinkedIn, domains, 110–180 source-grounded articles per year, one press release per month with global distribution, and AI-visibility monitoring of tracked queries. One package, $24,000 per year, 12-month term, 40+ markets [19][20]. The public reports at /ai-visibility are the proof surface for concentration like this paper’s. They are not a standalone ranking product. Rankings, AI recommendations, traffic and leads are not guaranteed.

Conclusion: the list is a retrieval, the register is a census, the asset is still the website

Four findings survive the method limits.

First, the shortlist is the wrong denominator. Fourteen ultrasound name strings against 620 510(k) applicants is not a market ranking. It is a short, SKU-mixed, third-party-cited generation. Most cleared manufacturers will not be on it. That is the expected state.

Second, concentration is a spine plus a long tail. 86% of names appear once. Ten names occupy five-plus slots. Chasing “why are we not Philips” is a different job from “why are we not on the one category we actually sell.”

Third, the citation layer is the part a company can improve without controlling the model. This corpus is 12.4% own-site with a 22% Gemini wrapper; a separate branded-query study is 98.5% own-site. Dealer and roundup hosts appear in the residual. A company controls whether accurate, useful owned pages exist; it does not control whether a model retrieves or cites them.

Fourth, the current pipeline cannot answer the list-length question. Twenty-seven panels hit a hard 15-name ceiling after secondary extraction and sorting. Rerun uncapped, preserving raw per-model order, before making any claim about volume and list length. The retained rows can still describe agreement within each published panel.

The screenshot will keep arriving. The honest response is to measure which query class it was, which model, which hosts were cited, and whether the company’s own public evidence could have been retrieved. Then build the loop — site, pages, public record, monitoring — as one system. That is the $24,000/year engine. Appearance on a 15-name list is a possible downstream observation. It is not the deliverable.

---

Sources

1. VayoMed analysis of 40 public “AI Recommends” category reports and 188 public brand AI-visibility reports (category generatedAt 17 February 2026 – 6 June 2026, most 6 June 2026; brand reports include later stamps). Compute accessed 15 August 2026. Method in Methodology. Public copies: AI Visibility reports.

2. VayoMed, AI Visibility reports for life sciences, accessed August 2026.

3. U.S. Food and Drug Administration, Downloadable 510(k) Files — VayoMed analysis of product codes IYN, IYO and ITX (2,475 records; 620 unique applicant strings), accessed 15 August 2026.

4. U.S. Food and Drug Administration, Product Classification: IYN — System, Imaging, Pulsed Doppler, Ultrasonic (21 CFR 892.1550, Class II), accessed August 2026.

5. Google Cloud, Grounding with Google Search (Vertex AI / Gemini grounding redirects observed in stored Gemini citation URLs as vertexaisearch.cloud.google.com), accessed August 2026.

6. VayoMed Research, Medical Device Press Release Distribution Benchmarks 2026: What Six Campaigns Actually Produced, 10 August 2026 (branded-query citation sweep: 98.5% own-site).

7. Google Search Central, AI features and your website (AI Overviews, AI Mode, query fan-out, same technical and helpful-content requirements as Search), accessed August 2026.

8. Google Search Central, Creating helpful, reliable, people-first content, accessed August 2026.

9. VayoMed, Best Ultrasound Machines — AI Recommends category report (generatedAt 6 June 2026; 14 recommended name strings), accessed August 2026.

10. U.S. Food and Drug Administration, Premarket Notification 510(k) (a substantial-equivalence order “clears” the device), accessed August 2026.

11. Electronic Code of Federal Regulations, 21 CFR 892.1560 — Ultrasonic pulsed echo imaging system (product code IYO), accessed August 2026.

12. Electronic Code of Federal Regulations, 21 CFR 892.1570 — Diagnostic ultrasonic transducer (product code ITX), accessed August 2026.

13. VayoMed Research, After 510(k), Can Buyers Actually Find You? A MedTech Commercial-Readiness Audit, 21 July 2026.

14. World Health Organization, Obesity: GLP-1 therapies (cited in stored ChatGPT GLP-1 citations), accessed August 2026.

15. U.S. Food and Drug Administration, Current List of FDA-Recognized 510(k) Third Party Review Organizations (the FDA host appearing in stored Perplexity citations on the 510(k)-consulting query), accessed August 2026.

16. VayoMed, SIUI AI visibility report (0 mentions; score 0; public panel), accessed August 2026.

17. VayoMed, Chison Medical AI visibility report (137 mentions; score 45; public panel), accessed August 2026.

18. VayoMed, SonoScape AI visibility report (4 mentions; score 15; public panel), accessed August 2026.

19. VayoMed, Done-for-You Global Growth Engine for Life Sciences ($24,000/year; six services; 40+ markets; PR guarantee is cadence and distribution), accessed August 2026.

20. VayoMed, Content Engine for Life Sciences (110–180 source-grounded articles per year; client review path), accessed August 2026.

21. VayoMed Research, What It Costs to Run Your Own Growth Engine, 27 July 2026.

Original analysis: 40 category reports and 188 brand reports retained as public VayoMed AI-visibility snapshots; citation-host bucketing, brand-name normalization, 510(k) IYN/IYO/ITX contrast and mention-pyramid shares computed 15 August 2026. Chart data files published alongside this report preserve every displayed series.