A physician might need to check a guideline, compare evidence, or make sense of a tricky clinical question. Instead of hunting through journals and reference tools, they can ask an AI-powered service and get a synthesized response. That workflow is drawing attention in US healthcare - and one company, OpenEvidence, says doctors are using it at a striking scale.
The Miami-based startup, founded in 2022, has reportedly reached a $15 billion valuation and built an advertising-funded business around its medical search product. Its rise suggests something practical: clinicians are willing to try AI when it fits the work they already do. It also raises uncomfortable questions. How reliable are the answers? What does it mean to insert advertising into a research workflow? And when a doctor relies on a system that summarizes medical evidence, who's accountable?
You can't treat OpenEvidence's growth as proof that AI is ready to make medical decisions. The real issue is separating four things: adoption, clinical usefulness, evidence quality, and business incentives. They overlap, but they aren't the same question.
What OpenEvidence does - and why it is gaining traction
OpenEvidence is an AI-powered search tool for physicians. Doctors can ask questions and use it to search medical journals and clinical guidelines, aiming to surface useful evidence without manually wading through large collections of sources.
That's a specific job, and clinicians tend to care about specificity. They don't need another general chatbot just to have one. They need a faster way to locate and interpret information while still juggling a heavy workload. A product that reduces friction in that process has a better shot at adoption than a tool that only works if doctors change how they practise medicine.
According to figures reported in the source material, roughly one million US physicians processed more than 42 million prompts in a recent month. The reporting also describes adoption approaching two-thirds of US doctors. Those numbers point to real use - but you have to read them with care. A prompt isn't automatically a unique doctor. It isn't automatically an independent clinical decision. And it isn't automatically a verified improvement in patient outcomes. Usage can signal workflow fit, not clinical performance.
Why a narrow workflow can spread quickly
Medical software has often been slow to integrate generative AI. The source account notes that Wolters Kluwer introduced an AI version of its established UpToDate product in October 2025 - nearly two years after OpenEvidence entered the market. That delay helps explain the opening: a newer company can move faster when an incumbent is still figuring out how AI belongs inside a long-standing product.
There's also a basic product advantage in shortening the distance between a question and a useful starting point. If a doctor has to search multiple sources, compare results, and then find the relevant guideline, an interface that compresses that process can save time. But that doesn't mean it should replace judgment. If the tool is going to be credible, the most defensible role is as a research assistant - helping clinicians find and assess material - rather than acting like an autonomous decision-maker.
The source reports that senior physicians interviewed praised the tool for easing workloads. That's meaningful as a snapshot of perceived value, but it's not the same thing as a controlled check of accuracy, safety, or patient outcomes. Both can be true: a product can genuinely make a task easier and still need careful validation before anyone treats its output as dependable clinical guidance.
Adoption is evidence of demand, not proof of safety
A product spreading through word of mouth among professionals is a signal worth noticing. It suggests the tool solves a problem people recognize and can explain to colleagues. But healthcare adoption comes with a higher bar than popularity in a consumer app. Guidance changes. Information can be incomplete. Patient details matter. A smooth answer can be more dangerous than an obvious mistake if it leads users to stop checking the underlying evidence.
For buyers and health systems, that means the practical questions are sharper than "Are doctors using it?" They include:
- Can users see the evidence behind an answer and verify that it actually supports the summary?
- Does the product make uncertainty and conflicting guidance visible?
- Are information retrieval and clinical recommendation clearly separated?
- Can the tool be evaluated on tasks that match the specialty and setting where it will be used?
- Is there a way to report errors and update the product over time?
Those questions turn a headline adoption figure into a procurement and safety discussion. They also help avoid a common trap: treating usage metrics as a stand-in for evidence of clinical performance.
The advertising model changes the trust equation
OpenEvidence's business model is part of the story, whether we like it or not. The platform is ad-powered, with commercial messages placed alongside a product used to find medical information. The reporting says the company booked $100 million in "ad buys" for the year at the end of 2025. The source defines that as covering current and future advertising commitments. A person close to the company said it was on track to generate between $400 million and $500 million through advertising by the end of 2026.
Those figures aren't interchangeable. "Booked" advertising commitments aren't the same as revenue already recognized, and a forecast isn't the same as a result. Those distinctions matter when you're looking at a private company whose valuation and growth projections get a lot of attention.
The business logic is understandable. A free or low-friction tool can spread quickly among clinicians, and pharmaceutical and healthcare advertisers pay to reach a professionally valuable audience. That can support growth without requiring every doctor to pay a subscription. The trade-off is that the product's revenue depends partly on advertisers - while its users need to trust that evidence discovery isn't being nudged by commercial incentives.
The key question is separation, not the mere presence of ads
Advertising isn't automatically incompatible with medical information. The real issue is whether paid placement can influence what a doctor sees, how recommendations are framed, or which evidence ends up most prominent. A clearly labeled advertisement that's technically and editorially separated from search results carries a different risk than a sponsored message that looks like an evidence-based answer.
A responsible system should make that distinction clear at the moment of use. Doctors need to know whether something is an advertisement, a source document, or an AI-generated summary. The source material doesn't provide enough detail to judge OpenEvidence's specific ad controls. So it would be premature to conclude either that the system is safe or that advertising has compromised its results. The sensible move is to ask how separation works and how it's audited.
A practical review should look at:
- Placement: Are ads visually and functionally separate from clinical search results?
- Labelling: Can a user spot sponsored content without relying on subtle design cues?
- Targeting: What information is used to choose an advertisement, and what limits apply in sensitive health contexts?
- Influence: Can advertisers affect rankings, generated answers, or which sources the system prioritizes?
- Governance: Is there a documented process for investigating complaints and correcting problematic placements?
These aren't abstract policy debates. In a clinical workflow, interface design can shape what users notice and how quickly they move from a search result to a decision. Trust depends not just on whether a company says it keeps ads separate, but on whether the product makes that separation visible and verifiable.
Revenue can fund access, but incentives need controls
An ad-supported model can lower the price barrier for doctors and widen access. That's a real benefit if the product is useful and if its commercial design doesn't erode information quality. A subscription model has its own downsides: it shifts cost to users or institutions and may limit adoption to organizations that can pay. Every business model involves incentives; the key is making them explicit and constraining their influence.
Healthcare organizations considering an ad-supported AI tool should treat the commercial model as part of the product review - not something you shrug off as "secondary" after a pilot goes well. If procurement evaluates only features and response speed, it misses a central part of what clinicians are being asked to trust.
Clinical accuracy is a workflow problem, not just a model score
The source account says OpenEvidence's growing influence has led researchers to scrutinize the accuracy of its results. It doesn't point to a specific study or provide an error rate. So there's no basis here for calling the tool accurate or inaccurate overall. What matters is that scrutiny becomes more consequential as reliance grows.
Medical questions are hard in ways that simple benchmarks can miss. The right answer may depend on age, medical history, other treatments, or the most recent guideline update. A response can sound confident while leaving out an important qualification. It can also cite a legitimate source and still summarize it incorrectly, or combine pieces of evidence that don't apply to the same patient population.
So accuracy isn't only about whether the model can produce a fluent sentence. It's about the full workflow. The system has to retrieve relevant material, represent it fairly, show where the information came from, communicate uncertainty, and help the clinician notice when extra verification is needed. A polished interface doesn't remove those requirements. If anything, it can raise the stakes by making a fluent answer easier to accept without checking.
Evaluate the task doctors actually perform
An evaluation should reflect how the tool will be used, not just how it performs in general. Testing a broad set of questions can tell you something about overall behavior, but it may not show how the product handles a specific specialty, ambiguous prompts, or situations where sources disagree.
For an organization assessing a clinical AI search product, a sensible process is:
- Define the intended use. Is it for finding sources, summarizing guidelines, generating a differential, or supporting a treatment choice? Each use case has a different risk level.
- Select representative questions. Use real, de-identified, or carefully constructed cases from the intended setting, including ambiguous and edge cases.
- Review sources and summaries together. A correct citation doesn't automatically mean the summary is faithful. Clinicians should be able to inspect both.
- Record failure modes. Track unsupported claims, missing caveats, outdated material, misleading confidence, and failures to surface relevant evidence.
- Test change over time. Models, indexes, and source collections can change. A one-time evaluation can't establish ongoing reliability.
- Set escalation rules. Define when users must consult primary sources, established references, or a specialist rather than relying on a generated response.
That's more operational work than asking a few clinicians whether the interface "feels good." But it's also more useful. A tool can score well in tests and still fail in real workflows where time is tight. A tool with known limitations can be used responsibly if those limitations are understood and the task boundaries are clear.
Keep clinical responsibility explicit
AI search can help a doctor gather information, but it doesn't become the clinician responsible for a patient. That distinction has to show up in training, interface design, and organizational policy. If a system is positioned as a shortcut to evidence, users need a clear way to inspect that evidence. If it starts pushing stronger recommendations, the standard for validation and oversight should rise with it.
The deployment failure to avoid isn't just "the model made an error." It's a chain of small assumptions: the interface looks authoritative, the answer sounds plausible, the source isn't checked, and the output enters a clinical decision without anyone noticing where uncertainty remained. Good deployment design interrupts that chain.
Competition will test whether adoption becomes durable
OpenEvidence has attracted nearly $1 billion in investment over the reported 18-month period, with backers including Andreessen Horowitz, Blackstone, Thrive Capital, and Sequoia. Its September funding round reportedly raised $250 million at a $15 billion valuation. Those figures suggest investor confidence, but they don't prove product quality, long-term profitability, or a defensible advantage.
The company isn't competing in a vacuum. Established medical information providers already have relationships with clinicians and extensive reference products. The source points to Wolters Kluwer's UpToDate as one example of an incumbent that added an AI version of its subscription service. The reporting also says large AI companies, including OpenAI and Anthropic, are moving into healthcare.
For OpenEvidence, the strategic challenge isn't just maintaining a lead in a new interface. It has to keep earning clinician trust while competing with companies that may have strong models, better distribution, existing customer relationships, or established medical content. Those assets don't automatically translate into a better clinical product. But they do raise the bar for a startup whose first advantage comes from being early and easy to adopt.
What could make the advantage durable
In healthcare, a product's long-term position depends on more than model quality. A durable tool needs a reliable evidence base, a usable workflow, transparent sourcing, governance users can understand, and a business model that doesn't create unacceptable conflicts. It also has to fit into the systems and processes medical organizations already run.
A fast-growing tool can win attention before it has fully answered those questions. That's normal in a new market. It also creates work for both the company and its customers. Physicians need to understand what the tool is designed to do. Health systems need evidence about performance and risk. The vendor needs to show that the product can improve - or at least remain auditable - as it scales.
So the valuation is best read as a market bet. Investors are betting that AI-assisted medical information becomes a large category. Whether OpenEvidence captures that category depends on what happens after the first adoption surge. In this market, the product that gets repeat use isn't always the same one that earns durable trust. It needs both.
A decision framework for clinicians and health systems
If you're a clinician, treat an AI search answer as a starting point, not the last word. If the decision matters, check the underlying source - especially when the answer involves a patient-specific treatment choice or offers a confident summary without enough context.
If you're evaluating the product for a practice or health system, ask vendors for evidence and controls instead of relying on growth claims. A short review can cover four areas:
- Purpose: What use cases are supported, and what's explicitly out of scope?
- Evidence: Can users inspect sources, and how is the product evaluated for the intended setting?
- Incentives: How are advertisements labeled and separated from clinical content?
- Operations: Who monitors issues, handles incidents, and communicates product changes?
A pilot should define success before rollout. That might include time spent locating evidence, clinician ratings of source relevance, observed error types, and adherence to verification procedures. It shouldn't rely on speed or user satisfaction alone. Faster answers help only if they don't undermine how the decision process works.
Conclusion: measure the system, not the valuation
OpenEvidence's rise shows that clinicians will adopt AI when it tackles a concrete task and fits into how they work. Its reported reach, investment, and advertising goals make it a notable case study in healthcare technology. But those numbers don't answer the central questions about clinical reliability, advertising safeguards, or patient outcomes.
The right standard isn't to dismiss the tool because it's commercial, or to trust it because it's popular. Evaluate what it does, inspect the evidence behind its answers, understand how ads are handled, and define the boundaries of use. If you're responsible for deploying AI in a clinical setting, run a task-specific pilot with explicit quality and safety measures - not just speed and adoption.
For the original reporting and figures discussed here, see the source article.