Five situations where the AI answer decided the outcome.
Each one starts from the prompts real HCPs, patients, pharmacists and payers typed - then shows the sources driving the answer and the compliant work that changed it. Brands are anonymised; the mechanics are not.
A GLP-1 obesity brand losing the launch conversation to compounded copies
Two weeks before a national obesity launch, the brand team had no view of how AI engines answered weight-management questions. Consumer prompts were dominated by compounding pharmacies and telehealth resellers; HCP prompts leaned on a competitor's head-to-head trial.
- “What is the most effective weight loss injection available in Germany?”
- “Semaglutide vs tirzepatide for a patient with BMI 34 and prediabetes”
- “Is compounded semaglutide the same as the branded version?”
Across ChatGPT, Gemini and Perplexity the brand appeared in 41% of HCP answers but only 12% of patient answers. The three most-cited sources were a telehealth blog, a 2021 meta-analysis predating the pivotal trial, and a national obesity society page that had not been updated since approval.
- -Medical affairs submitted an evidence update to the obesity society guideline page, with the pivotal trial and current label
- -The brand's HCP site was restructured into question-shaped, citable pages on dosing escalation, GI tolerability and CV outcomes
- -A compounding-risk explainer was published under medical information, not promotion, so engines had an authoritative counter-source
An originator biologic being written out of switching answers
When hospital pharmacists asked AI engines about switching stable patients from the originator to an adalimumab biosimilar, the answers read as a straightforward cost decision. Nothing in the answer mentioned device differences, citrate-free formulation or nurse-support programmes.
- “Should stable rheumatoid arthritis patients be switched to an adalimumab biosimilar?”
- “Interchangeability and nocebo effect when switching biologics”
- “Cost per patient per year adalimumab originator vs biosimilar in the Nordics”
Answer quality scoring showed the originator mentioned in 88% of answers but framed negatively or neutrally in 71% of them. Driver analysis traced the framing to two HTA summaries and one hospital procurement PDF that the engines treated as the canonical economic view.
- -Real-world persistence and nocebo data packaged as an open-access publication summary with structured metadata
- -Device and administration differences documented on a market-access resource page with plain, factual language
- -Weekly monitoring set on payer prompts so a shift in HTA-driven framing is caught within days, not quarters
An engine stating the wrong dose reduction for a targeted oncology therapy
A medical information team noticed a pattern of enquiries citing a dose-reduction schedule that did not exist in the SmPC. The source turned out to be an AI answer reproducing an early-phase protocol as if it were the approved regimen.
- “Dose reduction schedule for grade 3 hepatotoxicity on [molecule]”
- “Can [molecule] be combined with a strong CYP3A4 inhibitor?”
- “First-line treatment options for EGFR exon 20 insertion NSCLC”
Accuracy alerts flagged three engines returning a phase-I dose schedule and one omitting a boxed contraindication entirely. Every occurrence was reproducible and time-stamped, which is what medical affairs needed to act.
- -Evidence pack with reproducible answer transcripts filed to the affiliate medical governance committee
- -SmPC-aligned dosing and interaction content published in machine-readable form on the medical portal
- -Daily accuracy monitoring kept on all safety-critical prompts for this molecule
A rare disease invisible in the questions that precede diagnosis
The commercial problem was not brand share, it was that patients waited an average of six years for diagnosis. Increasingly, the first place a symptomatic adult or a puzzled GP describes those symptoms is an AI chat.
- “Adult with unexplained proximal muscle weakness and elevated CK, what could it be?”
- “When should I test for acid maltase deficiency?”
- “Which specialist do I see for progressive muscle weakness and breathlessness at night?”
The condition appeared in only 9% of symptom-description answers, and referral-pathway answers routinely stopped at neurology without naming a diagnostic test. Two national patient-organisation pages were the only sources engines trusted, and both were thin.
- -Disease-awareness content written to symptom-shaped questions rather than brand-shaped ones, non-promotional and unbranded
- -Diagnostic pathway and testing criteria co-published with two patient organisations to strengthen the sources engines already use
- -Prompt panel expanded to caregiver phrasing and five local languages
Adult vaccine answers drifting toward hesitancy sources
In three markets, answers to routine adult RSV and shingles vaccination questions began hedging, quoting forum threads and a retracted preprint alongside national immunisation guidance.
- “Is the RSV vaccine safe for a 68-year-old with COPD?”
- “Shingles vaccine side effects, should I worry about Guillain-Barré?”
- “Do I need the RSV vaccine every year?”
Sentiment tracking showed a measurable slide over eight weeks, correlating with a single high-traffic forum thread that had become a top-cited source in two engines. Guidance bodies were cited, but ranked below it.
- -Safety and revaccination-interval content published as concise, citable Q&A aligned to national guidance
- -Pharmacist-facing counselling resources structured for retrieval, since pharmacist prompts drove a third of the volume
- -Sentiment alerting configured to notify medical affairs on a two-week negative trend