The study
Mount Sinai researchers, publishing in The Lancet Digital Health on 11 November, gave 12 AI models from Google, OpenAI, Anthropic and others physician-written clinical cases. When a clearly labelled drug ad sat alongside the question, the models picked the advertised drug far more often.
47.6%
of answers picked the advertised drug when a labelled ad was present. 35% without it.
Source: The Lancet Digital Health, Mount Sinai (11 Nov 2026), 12 AI models
The nuance
No unsafe picks. Just a nudge between good options.
The shift happened where several treatments were clinically reasonable. The ads did not clearly push unsupported or suboptimal choices, and the study did not measure real prescribing.
Why it matters in the UK and EU
Advertising prescription medicines to the public isn’t allowed here, so buying a spot next to the answer is off the table anyway. But HCPs and patients increasingly arrive with AI-shaped questions, and since OpenEvidence withdrew from the UK and EU in April 2026, some clinicians have moved towards general-purpose tools.
We can’t buy the spot.So earn the answer.
What I’d be doing now
Credible. Findable. Balanced.
01
Make it credible
Keep medical and scientific content accurate, balanced and well structured, so people and machines can both trust it.
02
Own it together
Give medical and digital joint ownership of the source content, rather than treating it as a web task.
03
Check the answer
Monitor what the main AI tools say about your therapy area today, and package the routine so affiliates can run it locally.
Turns out the best AI strategy looks a lot like good scientific integrity.