The Label Is Not the Answer.

Pharma cannot buy its way into an AI answer and cannot promote inside one. That leaves a single compliant lever: the evidence layer the models read, and the discipline to monitor what they say about your products as seriously as you monitor adverse events.

A briefing you did not approve

Every approved claim a pharmaceutical company makes passes through medical, legal and regulatory review before it reaches a patient or a physician. The most widely read summary of your product this year will pass through none of it.

On 7 January 2026, OpenAI disclosed that more than 230 million people ask ChatGPT health and wellness questions every week, and launched ChatGPT Health, a dedicated space that can connect medical records and wellness apps. By 23 July it was available to all adult users in the United States. On the professional side, the AMA’s 2026 physician survey found 81% of US physicians using AI, up from 38% in 2023, with medical research summarisation among the leading uses. OpenEvidence alone reports more than 757,000 clinicians using its tool and over 20 million clinical conversations a month.

So the question is no longer whether AI talks about your molecule. It is what it says, to whom, and on which sources.

Two audiences, two machines

A general assistant answering a patient draws on two machines. The first is parametric memory: what the model absorbed about the drug during training, which includes label language but also forums, press coverage and controversies that the label long ago resolved. The second is live retrieval: the pages it fetches and cites at answer time. A clinical AI answering a physician leans harder on the second machine, grounded in journals and guidelines, and its framing is only as current as the evidence it can cite.

Fig. 1. The same molecule is described to patients by general assistants and to physicians by clinical AI. In both cases, the answer is assembled from sources, not from your approved materials.

Neither answer passes through an MLR process. Neither is required to provide fair balance. And both will be read as neutral, precisely because no brand appears to be speaking.

What the answers actually look like

The best independent measurement so far predates the current generation of models, and it is worth reading for its mechanism rather than its model. In a study published in BMJ Quality & Safety in October 2024, researchers asked an AI-powered search engine ten patient questions about each of the 50 most prescribed drugs in the United States. Around a quarter of the answers did not match reference drug data, and 39% contradicted scientific consensus. Of a subset of answers assessed by experts, 42% were judged capable of moderate or mild harm, and 22% of severe harm or death. Completeness was weakest on the question patients most need answered: what considerations apply when taking this drug.

Fig. 2. The failure is not hallucination in the abstract. It is missing warnings, outdated consensus and dense language, delivered with confidence.

Models have improved since. The architecture has not changed: an answer is only as accurate as the sources that win retrieval, and only as complete as the sources that were written to be complete.

The most widely read summary of your drug this year will never go through medical, legal and regulatory review.

Why pharma cannot buy its way out

In most categories the reflex is to buy the placement. In pharma, that door is closed twice.

First, the platforms have closed it. ChatGPT ads, now self-serve in more than 40 countries, exclude sensitive categories including health. Even where sponsored units exist, OpenAI states that advertisers have no ability to shape, rank or alter the response itself.

Second, the regulator is tightening the other side. Following a presidential memorandum on 9 September 2025, the FDA sent more than 100 enforcement letters to pharmaceutical and compounding companies and announced it would extend oversight to social media, influencer content and AI-generated health content, using AI tools of its own to surveil promotion. Any attempt to “optimise” an AI answer with promotional language is an attempt to create exactly the kind of content that is now being watched.

Clinical AI is a partial exception: OpenEvidence is funded by pharmaceutical advertising. But the ad sits beside the answer, and the answer is grounded in peer-reviewed evidence. A sponsored message next to a response that frames your therapy as second-line does not reverse the framing.

The compliant lever: the evidence layer

What remains is the part pharma already owns and rarely treats as a channel: the approved, factual, citable record of the product. Label text, SmPC and prescribing information, patient information leaflets, trial publications, safety communications, medical information responses. These are not marketing assets. They are exactly what a grounded model is looking for, and a KDD ’24 study across roughly 10,000 queries found that content carrying quotations, statistics and citations gained around 40% in AI-answer visibility.

The work is not to say more. It is to make the approved truth the easiest thing for the model to find, parse and cite, and to know within days when an answer drifts from it. We organise that into three registers.

Fig. 3. Three registers, three owners. The fix is always applied to the source the model cites, never to the prompt.

Label fidelity. Does the answer match the approved label, in each market? Wrong dosing, withdrawn indications and implied off-label use are the most common drifts, and the most exposed.

Safety completeness. Are contraindications, interactions and current warnings present? This is where AI monitoring starts to look like pharmacovigilance, and where it should borrow its discipline: logged findings, escalation paths, documented correction of the cited source.

Comparative framing. How does the model position your therapy against the class, against a rival, against a biosimilar or generic? This is the only register where brand has a legitimate voice, and only through published, comparable evidence.

What this changes inside the company

AI visibility in pharma does not belong to digital marketing alone. It sits between Medical Affairs, Pharmacovigilance, Regulatory and Brand, and today it usually belongs to none of them. The companies that move first will not be the ones that publish the most. They will be the ones that can show a regulator, at any moment, what the leading AI systems say about their products, which sources they cite, and what was done when an answer was wrong.

The playbook

1.  Audit before you optimise. Run the patient and HCP question sets for each priority molecule and market across the main assistants and clinical AI. Score every answer on the three registers. Keep the evidence.

2.  Fix the source, not the sentence. For each drift, trace the cited page and correct, update or publish the approved record it should have found. No promotional rewriting, no prompt games.

3.  Give it an owner and a cadence. Treat material misstatements like safety signals: logged, escalated, closed. Report AI answer fidelity to Medical and Brand leadership on the same calendar as the rest of the product dashboard.

The label will always be the legal truth about a medicine. It is no longer the version most people read. The work now is making sure the version they do read still says the same thing.

References

· OpenAI, “Introducing ChatGPT Health,” 7 January 2026 (updated 23 July 2026). openai.com/index/introducing-chatgpt-health/

· American Medical Association, 2026 Physician Survey on Augmented Intelligence (81% of physicians using AI, up from 38% in 2023). ama-assn.org

· MedCity News, “Thunderstruck by OpenEvidence’s $12B valuation? Don’t be,” February 2026 (757,000+ clinicians, 20M+ clinical conversations per month, pharmaceutical advertising model).

· Andrikyan W. et al., “Artificial intelligence-powered chatbots in search engines: a cross-sectional study on the quality and risks of drug information for patients,” BMJ Quality & Safety, October 2024.

· OpenAI, “Testing ads in ChatGPT” and “A milestone in expanding access to AI,” 2026 (sensitive categories including health excluded; advertisers cannot shape responses).

· Latham & Watkins, “FDA Begins Crackdown on Direct-to-Consumer Pharmaceutical Advertising,” 24 September 2025 (memorandum of 9 September 2025; 100+ enforcement letters; oversight extended to social, influencer and AI-generated content).