Doximity Ask

A language model does the answering, and the acquirer's own account of what it bought points somewhere else: one of the largest structured datasets in medicine, assembled over seven years, covering close to every guideline, drug and landmark trial. The price paid was for the data and the physicians who curated it as much as for the model.

That is the moat is the dataset case, and it holds this at B. The distinction matters to a buyer because it explains why the answers are good: not because the model is unusually capable in general, but because what it retrieves from has been curated by clinicians for years. A competitor with an equally good model and a worse corpus would produce worse answers.

Minimal autonomy by design and stated plainly, which is the right posture for a reference tool. The product answers a question; the clinician reads the answer, follows the citation if they choose, and decides.

The predecessor product's own guidance was explicit that the assistant is a supplemental resource and that outputs should be independently verified before use in patient care. That is a clearer statement of limits than most vendors in this index offer.

The practical tension is worth naming rather than treating as a defect. The product is sold on giving answers in seconds instead of minutes, and a clinician who independently verifies every answer against the cited literature has given back the time saving. The instruction is correct and the workflow it describes is not the one the product is marketed for.

Better than most in this category because a named benchmark with a number is published rather than a general claim of accuracy: a reported 96 percent on the United States Medical Licensing Examination, with an earlier version reported above 97 percent of human test takers and exceeding a major technology company's flagship medical model on the same test.

The corpus is also described in useful terms, as guidelines, drugs and landmark trials across specialties, with answers referenced to peer reviewed literature so a reader can check the basis.

What is not published is the model itself, the retrieval method, or how currency is maintained as guidelines change, which is the operative question for a reference product whose value is being up to date.

The exposure in a reference tool is not the record but the query, and that framing matters more here than the missing enumeration. Clinicians asking about a real patient type in the details that make the question answerable, covering age, presentation, medications and results, so those queries are patient information in everything but label, and whether they are retained, reviewed or used to improve the system is exactly what a buyer should establish.

Nothing published addresses it: no retention schedule, encryption detail or model training position specific to this product was located, and no model, hosting arrangement or sub processor list was named, with the acquirer describing compliant infrastructure in general terms. One complication is specific to this company and is worth resolving before relying on anything.

A sibling product from the same parent publishes specific and creditable commitments, including that prompt data is never shared and never used to train the company's models, and this index has recorded that the products are named inconsistently across the company's own material, with a support article carrying one name at an address carrying another.

So a buyer cannot establish from public material whether those prompt commitments cover this product or only its sibling, which is precisely the kind of question a naming inconsistency turns from trivial into material. Ask which product the published commitments govern, in writing.

A licensing examination score is a benchmark, not clinical evidence, and the distinction should be stated plainly because this category leans on it heavily.

An examination measures the ability to answer questions that have known correct answers, in a format designed to test humans, on material selected for testability. Bedside questions are not like that: they are underspecified, they concern a particular patient with competing problems, and often no single correct answer exists. A high examination score establishes that the system has the knowledge. It does not establish that a clinician using it makes better decisions, faster, or with fewer errors.

Nothing published measures that. Adoption figures are substantial, more than a million registered professionals under the predecessor product across 180 countries, and a strongly positive satisfaction score, but this index does not treat use or satisfaction as evidence of benefit.

Graded on an honest basis. The acquirer describes operating health privacy compliant infrastructure; no retention schedule, encryption detail or model training position specific to this product was located.

The exposure in a reference tool is not the record but the query. Clinicians asking about a real patient type in the details that make the question answerable: age, presentation, medications, results. Those queries are patient information in everything but label, and whether they are retained, reviewed or used to improve the system is exactly what a buyer should ask and what is not published.

Graded on an honest basis. Compliant infrastructure is referenced in company material; no agreement posture was located.

The relationship here is unusual and worth understanding. A physician using a free tool on their own device, through their own professional account, is not procuring it through their employer, so no institutional agreement governs the use and the hospital's information technology function may not know it is happening. That is a common pattern for clinician facing tools and it means the usual contractual protections are simply absent rather than weak.

Recorded honestly and provisionally: the dedicated trust and security search this index requires was not run in this pass, and no attestation specific to this product was encountered.

The parent is a listed company, so its annual filing carries mandatory cybersecurity risk management and governance disclosure. That filing is the authoritative source here and it was not checked. Read it before quoting this grade, as with the other listed company records in this index.

No device authorisation and none needed, and the reason is the most interesting regulatory fact on this record.

United States law excludes from the definition of a medical device software that supports clinical decisions where the clinician can independently review the basis for the recommendation and is not intended to rely primarily on it. A reference tool that answers with citations to peer reviewed literature, and instructs the user to verify independently, is built to sit inside that exclusion.

So the citation design is not only a quality feature; it is what keeps the product outside device regulation. That is worth understanding rather than criticising, and it has a consequence a buyer should hold onto: the boundary is crossed not by the tool becoming more accurate but by it becoming more directive, and nothing external audits which side of the line an answer falls on.

Nothing published on monitoring, error rates in practice, hallucination handling or performance variation, and the reach makes that omission larger than it would be elsewhere.

A free tool distributed inside a platform reaching more than 80 percent of United States physicians is potentially the most widely consulted clinical reference in the country. If its retrieval favours certain guideline bodies, certain trial populations or certain framings of a condition, that influence is applied at national scale and is invisible to the clinician reading a fluent answer.

Two specific gaps. No account of how currency is maintained when guidelines change, which for a reference product is a safety property rather than a feature. And no statement on how the system behaves at the edges of its corpus, where the honest answer is that the evidence does not settle the question.

A named benchmark with a number is published rather than a general claim of accuracy, including a reported score on a licensing examination and an earlier version reported above a stated share of human test takers and exceeding a major technology company's flagship medical model on the same test. Comparing against a named model on a named test is more useful than a bare figure, because the comparator is public and the claim is answerable by the party it names.

The corpus is described in useful terms as guidelines, drugs and landmark trials across specialties, with answers referenced to peer reviewed literature so a reader can check the basis. Held at C on two grounds. A licensing examination measures recall and reasoning over textbook medicine under exam conditions, which is a different thing from point of care reference accuracy: the questions are self contained, the answer is known to exist among the options, and nothing is stale.

A product a clinician consults about a real patient faces none of those conditions. And nothing is published about how currency is maintained as guidelines change, which is the operative question for a reference product whose whole value is being up to date, since a confidently delivered superseded recommendation is the characteristic failure of this category. No warranty, indemnity or remediation commitment was located. Ask how the corpus is refreshed and superseded guidance retired, and for an evaluation on real clinical questions rather than examination items.

Not an integration story. The product reaches clinicians through the parent's own application and network rather than through the record system, which is a deliberate distribution choice: it meets the clinician where they already are instead of asking the institution to install anything.

That has a real consequence for the answer quality. A reference tool inside the record can see the patient; this one knows only what the clinician types. It cannot ground an answer in the actual medication list or the actual results, which is precisely what the electronic record embedded tools in this index can do. No interface or record integration was located.

A hosted consumer style application delivered to individual clinicians, with no customer controlled deployment option, on premises variant or residency choice available or applicable.

Under the predecessor product the user base spanned 180 countries; the acquiring platform is oriented to United States physicians, and whether the free product is available on the same international footprint was not established. No hosting or retention detail was located.

The price is published and unambiguous: the product is free to clinicians, where the predecessor charged 300 dollars a year for its premium tier. A zero with no asterisk is more transparency about price than most of this index offers.

What is not disclosed at the point of use is who pays instead. The parent's business is built on reaching physicians on behalf of others, principally pharmaceutical and healthcare marketing, so a clinical reference tool given away inside that platform is funded by the same commercial relationships that other vendors in this category advertise themselves as being free of. One indexed competitor states explicitly that it carries no pharmaceutical advertising, positioning that as a deliberate contrast with ad supported reference products.

Nothing suggests the answers are influenced. The point for a buyer is narrower and worth holding: free is a price, not an absence of one, and the consideration here is attention and reach within a commercially monetised network. Graded B rather than A on that basis.

Broad by construction. The corpus is described as covering close to every guideline, drug and landmark trial across the major specialties, and the predecessor product reported use across 33 specialties in more than 180 countries.

Distribution is the more striking figure: a network stated to include more than 80 percent of United States physicians, which is a reach almost nothing else in this index can claim. The setting is the point of care in the general sense, meaning wherever a clinician has a question and a phone, rather than a specific unit or workflow. Held at B because the product does not extend to nursing, allied health or patient facing use, and because coverage outside the United States after the acquisition was not established.