It is a Tuesday afternoon and the chair of medicine at a 400-bed hospital is trying to name a new director of the hospitalist division.
She has three internal candidates. All three are competent by any formal measure: board certified, good quality scores, no complaints of note. She pulls up each one's Doximity profile because it is the fastest thing on her desk. One has 8,900 followers and posts regularly about practice management. Another has 1,100 followers and posts almost nothing. The third does not have a claimed profile at all.
The follower count tells her nothing about the question she actually needs answered, which is: when a hospitalist on that floor is stuck on a genuinely hard case at 2 a.m., whose name do they say out loud? Who do the nurses ask for by name when a family is angry and nobody can explain what is happening? Who do the other two candidates, privately, admit they would call first?
She does not have a number for that. Nobody does. So she does what every department chair in the country does in this situation: she starts making phone calls, catches five people between meetings, gets five partial and slightly contradictory answers, and makes the call on incomplete information dressed up as due diligence.
Medicine has spent fifteen years building an industry-scale measurement of who physicians are visible to. It has never built a measurement of who physicians actually trust, because the two are not the same thing, and only one of them can be sold to a pharmaceutical marketer.
The number everyone can see is the wrong number
Start with what is actually being measured when a platform shows a follower count.
A follow costs nothing. It requires no evaluation, no risk, and no memory of an actual interaction. Following a colleague on Doximity is closer to a passive acknowledgment that a profile exists than to any statement about competence. It correlates with how much content someone posts, how long they have had an account, and whether their specialty tends to be active on the platform. It does not correlate with whether their colleagues would send them a hard case.
LinkedIn's endorsement feature makes the mechanism even more explicit. Endorsements are frictionless, reciprocal, and almost entirely ungoverned, which is exactly why the feature is widely regarded inside and outside medicine as gamed and untrustworthy. Nobody treats a LinkedIn endorsement as a competence signal, because everybody understands, correctly, how cheap it was to give.
A trust nomination is the opposite kind of act. Naming a specific colleague as the person you would call about a specific kind of problem is a costly, meaningful, reputation-bearing statement, both about the person you name and about your own judgment in naming them. It requires the nominator to accept some accountability for the claim. That is precisely why almost nothing in medicine collects it: collecting it safely requires an environment where identity is verified, where the data cannot leak to the person being evaluated in a damaging way, and where the platform has no incentive to sell the answer to someone with money riding on the outcome.
Every commercial network physicians currently use was built to maximize the first kind of signal, because the first kind of signal is monetizable and the second kind is not.
What the real trust graph looks like when someone actually measures it
The useful fact in this problem is that the trust graph is not hypothetical. It has been measured, more than once, and the measurements agree with each other.
The best-validated proxy comes from a 2011 Health Services Research study by Barnett, Landon, O'Malley, Keating and Christakis, built on 2006 Boston-area Medicare data. The researchers tested whether physician patient-sharing, meaning two physicians treating overlapping patients, predicts whether those physicians would self-report a "recognized professional relationship" when surveyed directly. Among 616 physicians surveyed, with a 63 percent response rate, patient-sharing predicted a recognized relationship with an AUC of 0.73 (95 percent CI 0.70 to 0.75). At nine or more shared patients, relationship recognition rose to 82 percent.
Read that result honestly in both directions. An AUC of 0.73 is a real signal, well above chance, and it means an administrative proxy built entirely from billing data can approximate something as intimate as professional trust. It is also, bluntly, not that close to perfect: claims data gets the relationship wrong roughly a quarter of the time, and it is retrospective, lagging, and dependent on data access that most organizations do not have.
The referral data tells the same story from a different angle. A 2021 Health Services Research study by Pany and McWilliams, examining 40,495 referrals, found that primary care physicians refer to their own former residency or fellowship co-trainees at 26.2 percent of referrals, compared with 21.4 percent to non-co-trained specialists. That gap, roughly five points on a huge base of real-world referral volume, is a structured trust bias operating right now, informally, at scale, worth billions of dollars in routed care. It is evidence that physicians already act on a trust graph every day. They just act on the small, personal, memory-bound version of it, built entirely from who they happened to train alongside.
Direct sociometric surveys, the kind where researchers simply ask physicians "who do you turn to for advice," produce even richer data, including a companion line of work by Doumit and colleagues in 2011. These studies prove the method works. They are also, without exception, one-off academic projects: a single city, a single funding cycle, a response rate in the 37 to 63 percent range, and then nothing. No refresh. No continuity. The trust graph gets measured once, published, and immediately begins to decay, because physicians change jobs, retire, and shift specialties within a few years of the survey closing.
Why the gap persists: nobody who could build it wants to
This is the part of the problem that looks like an oversight and is actually a business model.
Doximity's revenue is built on physician attention sold to pharmaceutical marketers, historically making up the large majority of the company's revenue. A follower graph, a content-engagement graph, an audience graph: these are the exact assets that ad-based revenue depends on, because reach is what you sell to a marketer. A trust graph is close to worthless to that buyer and actively threatens the premise the whole product is built on, because it would demonstrate, publicly and repeatedly, that the physicians with the biggest audience are frequently not the physicians their peers actually rely on. No company builds a product that undermines its own pitch deck.
LinkedIn's endorsement system is the closest thing to an attempt, and it demonstrates the failure mode rather than the solution. Endorsements have no verification step and no cost to give, so they accumulate without meaning and nobody uses them for anything consequential.
Academic researchers have proven the method and have no reason or funding to run it continuously. A university grant pays for one measurement of one city's physicians in one year. It does not pay for a standing service that keeps the graph current as people change practices, retire, or move specialties.
Medical societies and hospital quality departments run the closest thing to a live version of this, in the form of one-off opinion-leader surveys funded by a quality-improvement grant to identify who should lead a guideline-implementation effort. These surveys work, they are expensive, they are run once, and the resulting list is stale within two years, because nobody is paying to keep it current.
So the pattern across every plausible builder is identical: the method is proven, and the incentive to run it as a living, continuously updated asset does not exist anywhere in the current market.
The hidden network problem: what makes a trust edge different from a follow
It is worth being precise about why this cannot simply be added as a feature to an existing platform.
A follow is a one-directional, frictionless, essentially free action. It requires no governance because it carries no information about the person being followed beyond the fact that someone clicked a button.
A trust nomination is reputation-bearing information about a third party. Naming Dr. Ramirez as who you would call about a difficult vascular case is a statement that, if mishandled, could be seen by Dr. Ramirez's employer, competitors, or a plaintiff's attorney trying to establish a standard of care. It needs an accountable environment: verified identities on both sides, a clear rule that nominations are never sold or shown to pharmaceutical or device companies, and a design that captures only positive nomination, never a negative or comparative ranking that could function as a public demotion of a named colleague.
That governance requirement is exactly why no low-friction consumer network has built this. Building it well requires the platform to be trusted enough, by verified professionals, that they will put a real name on a real, meaningful claim about a colleague. That trust has to be earned before the product exists, which is a chicken-and-egg problem every academic pilot has solved once, briefly, and no commercial platform has solved on an ongoing basis.
Why now, specifically
The value of this gap is not static, and the direction it is moving in should worry anyone still treating follower counts as a proxy for expertise.
AI answer engines are rapidly commoditizing the part of medical knowledge that used to require finding a smart colleague: the factual lookup, the differential, the drug interaction check. Tools built for exactly this purpose are already routing straightforward clinical questions away from human consultation. That is, on its own terms, a good thing for efficiency.
It also means the remaining scarce resource is precisely the thing nobody has captured: whose judgment do peers actually trust when the textbook answer runs out. As the easy questions get absorbed by tools, the hard questions, the ones that require someone who has actually managed the specific situation, become relatively more valuable, and the absence of any way to find that person becomes relatively more costly.
What would actually work
Continuous collection, not a one-time survey. The Barnett and Doumit studies prove the method works. What has never existed is a standing collector that refreshes the graph as physicians change jobs, specialties, and networks, rather than freezing it at the moment a grant-funded survey closes.
Positive nomination only, never negative ranking. The product must capture "who would you ask," full stop. A feature that lets physicians rank or downgrade named colleagues creates immediate defamation exposure and destroys the trust required to participate honestly. This is a design constraint, not an afterthought.
Verified identity on both sides of every edge. A trust nomination is only meaningful, and only safe to collect, inside an environment where both the nominator and the nominee are confirmed real, licensed, identifiable professionals. An anonymous or unverified network cannot host this safely.
A hard governance wall against monetizing the nominations themselves. The moment trust data is sold to a pharmaceutical company, a device maker, or an expert-witness broker, physicians stop nominating honestly, and the entire asset degrades back into a gameable popularity contest. This is the single differentiator from every follower-based platform, and it has to be structural, not a policy line in terms of service.
Benchmarking against an independent proxy. The claims-based patient-sharing signal, at an AUC of 0.73, is not a replacement for direct nomination, but it is a useful, continuously available check on whether the nomination-based graph is behaving sensibly, and a way to validate new entrants before enough direct nominations exist for them.
Density before breadth. A trust graph is only useful once a meaningful share of a given specialty's actual trusted contacts are present to be nominated. A thin, broad rollout across every specialty produces a graph too sparse to answer any real query; a deep rollout inside a smaller number of specialties and regions produces something a physician can actually use on day one.
A byproduct model, not a survey model. The most durable version of this asset is generated as the exhaust of real accountable activity, an answered consult, a countersigned attestation, a referral that was actually routed, rather than asking people to recall and name trusted colleagues out of context. Byproduct data is harder to game and does not decay the way a point-in-time survey does.
What you can do now
If you are a clinician
Write down your own list. Who are the five people you would actually call about the five hardest categories of case in your specialty? Most physicians have never made this list explicit, even to themselves, and making it explicit is the first step toward it being useful to anyone besides you.
Stop treating follower count as a proxy for anything clinical. When you are choosing who to ask for a second opinion, a large public following is, at best, irrelevant information and, at worst, a signal that correlates with self-promotion rather than the kind of quiet reliability colleagues actually depend on.
Offer to be findable for what you are actually good at. The informal trust network runs on people being willing to say, explicitly, "call me about this specific thing." Almost nobody does this proactively, and it is the cheapest possible contribution to fixing the discovery problem.
If you lead a department or a quality program
Separate influence from expertise explicitly when you choose opinion leaders. A one-off sociometric survey, run properly, will identify who your staff actually turn to, and it will not be the same list as who has the most social media presence or the most publications. Fund it as a recurring exercise rather than a one-time grant deliverable, because the list decays within about two years.
Ask the "who would you ask" question directly, in writing, at intervals. It is a simple survey instrument with a long academic track record. The barrier has never been methodology; it has been that nobody keeps running it after the first grant cycle ends.
Treat this as a leadership-selection tool, not just a quality-improvement tool. The chair choosing a division director, the CMO choosing a service-line lead, and the QI director choosing a guideline champion are all trying to answer the same underlying question with no shared infrastructure to answer it.
If you build systems or research this space
Replicate the Barnett methodology with current data. The strongest available proxy validation is now more than a decade old and comes from a single city. A multi-city, current replication using commercial claims, not just Medicare, would meaningfully update the field's confidence in the AUC 0.73 figure.
Design the governance layer before the product layer. The technical challenge of collecting nominations is smaller than the trust challenge of collecting them safely. Any serious attempt has to start with the rule that nominations are never sold, never shown as negative rankings, and never exposed to the nominee's employer without consent.
Frequently asked questions
How do physicians find out who other doctors actually trust? Right now, mostly by word of mouth and personal networks built during training. The one validated large-scale proxy is patient-sharing data, which predicted a self-reported "recognized professional relationship" with an AUC of 0.73 in a 2011 Health Services Research study of 616 Boston-area physicians (Barnett et al.). No continuously updated, queryable version of this exists.
Does a large Doximity following mean a doctor is respected by peers? Not reliably. Follower counts measure visibility and platform activity, not clinical trust, and Doximity's revenue model depends on selling physician attention to pharmaceutical marketers, which creates no incentive to build or surface a separate trust metric. No published study directly correlates Doximity follower counts with peer-nominated trust.
Can patient-sharing data actually predict physician referral trust? Yes, with real but limited accuracy. The Barnett et al. 2011 study found an AUC of 0.73 (95% CI 0.70 to 0.75), rising to 82 percent relationship recognition once two physicians shared nine or more Medicare patients. It is a genuine signal, not a definitive one, and it is retrospective and dependent on claims-data access most organizations do not have.
What is a physician trust graph? A dataset of who clinicians actually turn to for advice on specific clinical problems, as distinct from who they are visible to or connected with online. It has been measured in academic pilots such as Barnett et al. (2011) and Doumit et al. (2011), but no platform maintains it continuously; every existing measurement is a one-time research snapshot.
Why do primary care physicians refer more often to doctors they trained with? A 2021 Health Services Research study by Pany and McWilliams, examining 40,495 referrals, found PCPs referred to their own former residency or fellowship co-trainees 26.2 percent of the time, versus 21.4 percent to non-co-trained specialists. That gap is evidence of a real, informal trust bias already shaping billions of dollars of referral routing, built entirely from personal relationship rather than any measured competence signal.
What is the difference between influence and expertise in medicine? Influence, as commercial platforms measure it, tracks audience size, posting frequency, and public visibility. Expertise, in the form colleagues actually rely on, is closer to what sociometric surveys capture: a costly, specific nomination of who someone would call about a hard case. The two are frequently uncorrelated, and the most-followed physicians on a platform are often the most prolific self-promoters rather than the physicians their peers quietly rely on.
The bottom line
The chair of medicine in that Tuesday afternoon office is not short on data about her three candidates. She has quality scores, complaint histories, and, if she wants it, a follower count accurate to the last digit.
What she does not have is the one number that would actually answer her question, because that number has only ever been measured in isolated academic pilots that expire the moment the grant does. The method is not the obstacle. Barnett proved patient-sharing data gets it right, at an AUC of 0.73, three-quarters of the time. Doumit proved direct sociometric surveys work even better when people are willing to name names. The Pany and McWilliams referral data proves physicians are already acting on an informal trust graph every single day, worth billions of dollars in routed care.
What has never existed is a standing collector: an entity physicians trust enough to name real colleagues to, on a continuous basis, without fear the answer becomes ad inventory for a pharmaceutical company. Every commercial incentive in the industry points toward measuring visibility instead, because visibility is what can be sold.
So the chair makes five phone calls and takes a partial answer, the same way every department chair, every QI director, and every patient searching for a second opinion has always done it. The trust exists. It just lives entirely in memory, dispersed across thousands of people who have never been asked to write it down anywhere durable.
And the follower count, the wrong number, stays the only number anyone can see.
Part of a series on the missing professional infrastructure of healthcare. Previously: The Surge Registry: Credential Verification That Cannot Happen at Disaster Speed
Evidence note: the core validation figure, an AUC of 0.73 for patient-sharing as a proxy for self-reported professional trust, comes from Barnett ML, Landon BE, O'Malley AJ, Keating NL, Christakis NA, "Mapping physician networks with self-reported and administrative data," Health Services Research, 2011, based on a single-city (Boston), single-year (2006) sample of 616 physicians with a 63 percent survey response rate; it should be read as strong but dated and geographically limited evidence. The referral co-training figure (26.2% vs 21.4%) is from Pany MJ, McWilliams JM, Health Services Research, 2021, examining 40,495 referrals; the original source URL for this figure was not independently re-verified in this pass and is cited as reported in the underlying research dossier. The claim that Doximity's revenue is predominantly built on physician attention sold to pharmaceutical marketers is drawn from prior competitive-landscape research cited in the dossier rather than an independently re-pulled financial filing in this pass, and should be treated as directionally reliable but not freshly audited. Doumit et al.'s 2011 sociometric work is referenced as a companion academic validation of the direct-survey method; this article did not independently re-verify its specific figures. Nothing in this article should be read as a claim that any named commercial platform currently offers a validated trust-graph product.
