Two hundred million clinical studies is the sort of number that can end a sales conversation before the useful questions begin. It sounds comprehensive. It suggests that whatever your clinical question, the answer must be somewhere inside an enormous library.
It also deserves a definition.
In an earlier company LinkedIn post, CaryHealth promoted Clair with access to more than 200 million clinical studies and 147,000 FDA package inserts. This is CaryHealth's clinical information assistant, a separate product from the Clair Health wearable.
The number is doing substantial persuasive work. A clinician can reasonably read it as a count of distinct pieces of clinical research. If that is the intended meaning, a source inventory should make it straightforward to demonstrate. If the counting unit is something else, that unit belongs in the headline.
First, decide what a study is
A study, a publication and a database record are different objects. One trial can generate a protocol, a registration, conference abstracts, a principal paper, subgroup analyses and follow-up reports. Each may be useful. Counting them separately does not create additional independent trials.
A retrieval system can further divide one paper into passages. That can be a sensible engineering choice: answering a narrow question may require a paragraph rather than an entire document. But a paragraph stored for retrieval is not a new clinical study either.
None of these distinctions tells us which counting method CaryHealth used. That is precisely the information the headline leaves the reader needing. There is no reason for clinicians to reverse-engineer a vendor's database vocabulary from a promotional number.
The PubMed description provides a useful reference point: it describes more than 40 million citations and abstracts, using a defined bibliographic unit. PubMed is not a ceiling on the world's research, and its total cannot establish that another collection is impossible. The comparison simply shows why a much larger claim labelled clinical studies needs its own equally intelligible accounting.
The appropriate response is neither automatic applause nor a declaration of mathematical impossibility. It is a request for the data dictionary.
A library needs shelves, not just a turnstile counter
A useful inventory would distinguish what enters the system, how it is counted and what the user can inspect. The following is a proposed disclosure structure, not a description of Clair's implementation.
| Collection component | Counting unit worth publishing | Clinical question it helps answer |
|---|---|---|
| Research literature | Unique documents, plus linked study identifiers where available | How much distinct research is represented? |
| Guidelines | Current guideline documents and versions | Which recommendations can the system retrieve? |
| Drug information | Products, labels and historical versions separately | Is the answer using the relevant product information? |
| Retrieval passages | Chunks, explicitly labelled as chunks | How has material been prepared for search? |
| Duplicates and updates | Records removed, linked or superseded | Does the headline repeatedly count the same underlying material? |
The table matters because different counting choices can produce spectacularly different totals from the same intellectual content. A vendor should be able to explain those choices without revealing its model weights or commercial contracts.
Freshness also belongs in this account. A collection can grow by retaining superseded documents. That may be valuable for historical questions, but it creates a separate obligation to recognise which version should govern a current answer. More records can mean more work for the retrieval system, not automatically more clinical confidence.
Access is the start of the evidence chain
CaryHealth's Clair product page presents the service as a clinical information tool. That puts the important test at the point where a clinician receives an answer.
Consider an illustrative drug information request involving a patient with several relevant complications. A large collection may contain the correct guidance, an older recommendation, a study in a different population and a superficially similar case report. The product's value depends on retrieving the appropriate material, recognising its limitations and supporting the actual recommendation it produces.
An answer with twelve references can still fail that task. The references might address only the uncomplicated version of the question. They might support the background explanation while leaving the proposed action unsupported. Or they might represent several publications from one underlying study, giving the appearance of independent corroboration.
The useful evaluation therefore follows the answer backwards: recommendation, supporting passage, source, applicable population and date. Counting the entire library does not complete any of those steps.
This is also why a smaller, carefully maintained collection can be competitive for a defined task. That is an argument about relevance and maintenance, not a claim that smaller databases are inherently better. The size question becomes informative only after the intended clinical job is specified.
There is an answer-quality study to discuss
A University of Illinois Chicago study published in May 2026 tested 30 drug information questions across Clair, OpenEvidence, GlassHealth, DougallGPT and ChatGPT. Three pharmacy faculty members assessed responses using a 25-point framework. OpenEvidence scored highest at 17.82; every chatbot scored below the faculty responses. Supporting evidence was the weakest scoring component across the AI systems.
That is a more clinically interesting conversation than a contest over collection size. The study examined answers to identifiable questions with an explicit assessment method. Its abstract does not supply a separate numerical result for Clair, so it cannot support an invented Clair score or a permanent ranking of today's products.
Its relevance to the headline is more basic. Evidence needs to survive the journey from a collection into an answer. A corpus can be extensive while the final response remains inadequately supported. Buyers should ask to see that journey tested on the questions they actually need answered.
The questions a buyer should put on the table
The first question is whether the promotional count refers to distinct studies, documents, database entries or retrieval passages. The second is which sources contribute to it. The third is how duplicate publications and superseded versions are handled.
Then the discussion should move to performance: can evaluators inspect the retrieved passages, test difficult questions and identify precisely why an answer succeeded or failed? Are comparisons performed against the same questions, versions and assessment criteria? Can the vendor describe where expanding the collection improved results, rather than simply enlarged the number?
For iatroX, the same standard starts with material readers can inspect through how the product works and its methodology. A competitor's ambiguous metric is no reason to introduce a grander number of our own.
CaryHealth can make the claim more useful by defining the inventory and connecting it to evaluated answers. Until then, 200 million remains a striking headline with a missing unit of analysis. Medicine already has enough impressive numbers that require somebody else to explain what they mean.
Frequently asked questions
Is CaryHealth's Clair the same as Clair Health's wearable?
No. CaryHealth's Clair is a clinical information assistant. Clair Health markets a wearable related to hormone and cycle information. Their evidence, products and promotional claims should be assessed separately; a criticism of one is not evidence about the other.
Does the PubMed total prove that CaryHealth's figure is wrong?
No. Different collections can cover different materials, and PubMed does not represent every research record worldwide. Its clearly described counting unit illustrates the disclosure required to interpret CaryHealth's figure. A direct numerical comparison without matching the units would repeat the mistake being criticised.
What would make a large corpus claim clinically persuasive?
A defined inventory, transparent treatment of duplicates and updates, and evidence that retrieval improves answers on relevant tasks. The strongest demonstration connects the sources used to the recommendations produced and makes errors inspectable. A larger total alone does not establish that connection.
Examine the evidence behind a clinical answer with Ask-iatroX ->