Insights · For the moments that define the job
How do you monitor what AI says about a drug or a disease?
Which questions should you track?
Start from what people actually ask, not from your key messages. A useful question bank has four layers:
- The disease. "What is [condition]?" "How is it treated?" "What are the newest treatments?"
- The product. "What is [product]?" "What is it approved for?" "Who should not take it?" "What are the side effects?" "How much does it cost?"
- The moments. "What did [trial name] show?" "Did the FDA approve [product]?" "Is [product] better than [alternative]?"
- The concerns. The questions your medical information team and patient support lines already hear most often.
Keep the wording fixed from one check to the next, so a change in the answer means the answer changed, not the question. Refresh the bank deliberately, a few times a year, from what people are really asking.
Should you ask as a patient or as yourself?
Both, and the difference is measurable. In Signal Fidelity Group's Disclosure Effect study, adding patient or caregiver context to the same health question changed the sources cited past the study's threshold in 29.4% of eligible comparisons. One system did so in 48.3% of comparisons; another in 5.1%. The study is not peer reviewed and did not test medical accuracy, but the lesson for monitoring is direct: a program that only asks as an anonymous user sees one slice of what patients, caregivers and physicians see. Ask each important question in each audience's voice. What the study found.
How often should you check?
- A standing cadence, monthly at minimum, for the core question bank.
- Around events, on a tighter cycle. A sensible pattern after a readout or a regulatory decision is to check at 24 hours, seven days and 30 days.
- After any correction, to see whether the fix reached the answer.
What should you record each time?
For each question, engine and audience voice: the date, the full answer, every source cited, and a short judgment against a fixed list. Keep the raw answers; a summary alone cannot show what changed.
Be clear about what you are observing. Answers fetched through an engine's developer interface with web search are not identical to what a signed-in person sees in a consumer app, which can add memory and personalization. Say which you used.
What counts as a problem?
A fixed list keeps the judgment consistent:
| Finding | Example |
|---|---|
| Wrong | A dose, an indication or a trial result stated incorrectly |
| Dated | Describes an approved product as investigational, or omits a newer label change |
| Broader than the label | The trial population, or an off-label use, presented as the approved use |
| Missing safety context | Benefits listed without the important safety information |
| Unbalanced comparison | "Better than" a competitor, with no head-to-head evidence |
| Thin | A generic answer with weak sources, where a clear, citable source would help |
| Absent | Your product, trial or disease area does not appear where it should |
Not every finding needs the same response. A wrong dose is urgent; a thin answer is an opportunity.
What do you do when an AI answer is wrong?
You cannot edit an answer engine. You can change what it reads.
- Find the source. Most answers cite their sources. Identify the page the error came from.
- Correct the source. If it is yours, fix it and date the change. If it is a third party's (a news story, a patient organization's page, an encyclopedia entry, a directory), the remedy is ordinary communications work: a correction request with the evidence, or updated materials for the partner.
- Publish the clear answer. Where no good source exists, publish a short, dated, sourced answer that cites the primary evidence: the label, the publication, the registry entry. Fresh pages matter: Seer Interactive, looking at citations to four brands from March to June 2026, found that 75% of the cited pages had been updated within the past year.
- Check again. Re-ask the same question on the same schedule and record whether the answer changed.
Remember that being found and being cited are different. In an Ahrefs analysis, only 37.9% of URLs cited in Google's AI Overviews also appeared in the first ten result blocks, so a page does not need to rank near the top to be the source an answer uses.
What about adverse events and medical questions?
Monitoring sometimes surfaces things that are not communications problems. If a check, or something a person shared alongside it, describes a possible adverse event, follow your company's pharmacovigilance procedure, exactly as you would for social listening. If an answer raises a medical question, route it to medical information. A monitoring program should have those routes agreed before its first check.
How does this fit with media monitoring?
AI answers are now part of the same picture as news coverage and social conversation, and they draw on both. A story that gets something wrong today can become an AI answer that gets it wrong for months. It makes sense to read coverage and AI answers side by side, and to trace the errors in one back to the other.
How does CommsRx approach it?
CommsRx treats answer engines as intermediaries to be measured and briefed, not gamed. AI Pathways tests stakeholder questions to see how AI answers frame a market and which sources they cite; the Morning Briefing (Brand Coverage and Strategic Watch) follows the coverage around them; and Ally's AI answer-surface playbook takes a team from a measured check to a map of sources, an owned answer pack, outreach to the explainers that matter, and a re-check. Every answer drafted in CommsRx is built from confirmed messages and goes through the team's own review.
Related questions
Sources
- Signal Fidelity Group, The Disclosure Effect, 2026. https://signalfidelitygroup.com/research/the-disclosure-effect (checked October 19, 2026). Signal Fidelity Group was founded by Abhi Basu.
- Ahrefs, Louise Linehan, March 2, 2026, "Update: 38% of AI Overview Citations Pull From The Top 10." https://ahrefs.com/blog/ai-overview-citations-top-10/ (checked October 19, 2026)
- Seer Interactive, Sonny Vasquez, July 24, 2026, content recency study. https://www.seerinteractive.com/insights/study-content-recencys-impact-on-ai-visibility-in-2026 (checked October 19, 2026)
About this page
General guidance, not legal, regulatory or pharmacovigilance advice. Drafted with AI assistance from public research and CommsRx's own playbooks, then edited and checked by Abhi Basu, founder of CommsRx and Signal Fidelity Group. Last reviewed October 19, 2026. Corrections: help@signalfidelitygroup.com.