Your Brand Has an AI Reputation. Who’s Managing It?

In a KFF poll fielded February 24 to March 2, 2026, 32% of US adults said they had used an AI chatbot for health information in the past year. [1] The same year, OpenAI reported that more than 40 million people send health-related prompts to ChatGPT every day. [2] Every one of those conversations can mention a drug, a manufacturer, or a competitor. No label, no medical information letter, and no brand website controls what the model says.

This article asks a narrow question: inside a pharmaceutical company, which function owns what AI systems say about its products? We reviewed the public evidence on AI drug-information accuracy, the sources AI systems cite, and the FDA documents that touch the issue. We then built an ownership model and an escalation matrix from that evidence.

We did not run new AI queries for this article. Every statistic comes from a cited source. Every calculation is labeled as our analysis. The frameworks are our own proposals, not industry standards. DrugChatter publishes this article and sells AI-monitoring software, so we describe that tool in one section and keep it out of the evidence sections.

The Short Answer: Who Owns a Brand’s AI Reputation?

What Is an AI Reputation for a Drug Brand?

An AI reputation is the set of statements AI systems make about a drug, its manufacturer, and its competitors when people ask real questions. It includes what the answer says, which sources it cites, and whether it mentions the brand at all.

Who Manages It Today?

In our search of public sources, we found no survey that measures which pharma function owns AI reputation. We found vendor reports on AI visibility, peer-reviewed accuracy studies, and FDA documents. None of them assign ownership. Our working conclusion is that responsibility is usually split across Medical Information, Medical Affairs, Pharmacovigilance, Regulatory Affairs, Legal, Brand, and Corporate Communications, with no single accountable owner. That conclusion is an inference from the absence of data, and we label it as one.

The Five Findings That Matter

  1. A KFF poll found 32% of US adults used AI chatbots for health information in the past year. [1]
  2. In a BMJ Quality & Safety study of Bing Copilot, 126 of 484 assessed answers (26.0%) had accuracy below 100% against reference drug data, and 16 of 484 (3.3%) were fully inconsistent with it. [3]
  3. In a Long Island University study, pharmacists judged 10 of 39 ChatGPT answers to drug questions satisfactory. [4]
  4. In 5W’s Q2 2026 index, Eli Lilly and Novo Nordisk together held an estimated 24.0% citation share across 25 pharma companies (our sum of two reported figures). [5]
  5. In the Worldcom Healthcare Monitor, nih.gov accounted for 51% of all AI citations, while the best corporate domain, AbbVie, held 1.8%. [6]

32% of adults say they used AI chatbots for health information in the past year. (KFF Tracking Poll on Health Information and Trust, March 25, 2026; 1,343 US adults, margin of error plus or minus 3 points) [1]

What Is a Brand’s AI Reputation, and How Does It Differ From SEO?

Definition: AI Reputation in Pharma

We define AI reputation as the combination of four measurable properties: accuracy against the current label, source selection, completeness of safety information, and presence in the answer. This is our definition. It is not an established industry term.

AI Reputation vs. Search Ranking vs. Brand Reputation

ConceptWhat it measuresWho usually owns itControlled by the company?
Search rankingPosition of a page in a results listDigital marketing, SEOPartly
Brand reputationPerception among patients, HCPs, payersBrand, Corporate CommunicationsPartly
AI reputationWhat a model says, cites, and omitsNot clearly assignedNo, only the source material

Source: our analysis. The ownership column reflects our inference, not survey data.

Why Pharma Differs From Consumer Brands

A wrong answer about a sneaker is a service problem. A wrong answer about a drug interaction can affect a patient’s treatment. The Long Island University team gave one example: ChatGPT said no interactions had been reported between Paxlovid and verapamil, while the lead author said the two drugs have the potential to interact and may lower blood pressure excessively. [4]

How Many People Ask AI About Drugs and Health?

What Does the KFF Poll Say About AI Health Use?

KFF found that 32% of adults used AI chatbots for health information in the past year. [1] The poll surveyed 1,343 adults in English and Spanish. [1] Among users who asked about mental health, 58% did not follow up with a doctor or other health professional. Among users who asked about physical health, 42% did not. [1]

How Many Health Questions Does ChatGPT Receive?

OpenAI’s January 2026 report, as covered by Medical Economics, says more than 40 million people submit health-related prompts each day. [2] It also says roughly one in four of its more than 800 million regular users submits at least one health-related prompt per week. [2] More than 5% of all ChatGPT messages globally concern health care. [2]

Our calculation: one in four of 800 million is about 200 million weekly users. That figure is an estimate derived from OpenAI’s rounded statement. We did not retrieve the OpenAI report itself, so we cite the press coverage of it.

What the Usage Numbers Do Not Say

Neither figure tells us how many of these questions name a specific drug. Neither tells us how often the answers are wrong. Usage data show exposure. Accuracy studies show risk. The two must be read together.

How Accurate Are AI Answers About Drugs? What the Studies Show

Bing Copilot and 50 Drugs: The Erlangen Study

Andrikyan and colleagues at Friedrich-Alexander-Universität Erlangen-Nürnberg queried Bing Copilot in April 2023. [3] They asked 10 patient questions about each of 50 drugs. That produced 500 answers containing 1,727 statements. [3]

The mean Flesch Reading Ease score was 37.2, which the authors describe as difficult to read. [3] Mean accuracy was 88.7% and mean completeness was 76.7%. [3] In 126 of 484 assessed answers (26.0%), accuracy was below 100%. [3] In 16 of 484 (3.3%), the answer was fully inconsistent with the reference. [3]

The harm assessment covered a subset. Experts reviewed 20 answers. They judged 66% potentially harmful, 42% potentially causing moderate or mild harm, and 22% potentially causing death or severe harm. [3] The 95% confidence interval for the 22% figure runs from 10% to 40%. [3] With 20 answers, we treat these percentages as a signal, not a rate.

ChatGPT and 39 Real Drug Information Questions

A Long Island University team took real questions from its drug information service, collected over 16 months in 2022 and 2023. [4] Pharmacists answered 45 of them. The team put 39 to the free version of ChatGPT. [4]

Pharmacists judged 10 of the 39 answers satisfactory. [4] The other 29 were not. [4] ChatGPT supplied references in only eight responses, and the coverage reports that the references included nonexistent ones. [4]

Our calculation: 10 of 39 is 25.6% satisfactory, and 29 of 39 is 74.4% not satisfactory.

Do Clinically Focused Chatbots Perform Better?

Ipema and colleagues tested five chatbots on 30 drug information questions that pharmacists had already answered. [7] The chatbots were OpenEvidence, Clair, Glass, DougallGPT, and ChatGPT. OpenEvidence scored highest, at a mean of 17.82 out of 25 on the CLEAR instrument. ChatGPT scored 15.72. All chatbot responses scored statistically lower than the pharmacists’ responses. [7]

A separate study by Surbaugh and Moeller, published online March 23, 2026, tested four tools on dosing of long-acting injectable antipsychotics. [7] OpenEvidence was 95% accurate. ChatGPT was 40% accurate. [7]

The spread matters for brand teams. One study tool scored 95% and another 40% on the same task. A brand cannot assume that one model’s answer predicts another’s.

What About Health Misinformation Beyond Drug Labels?

Tiller and colleagues audited five chatbots on 50 prompts about cancer, vaccines, stem cells, nutrition, and athletic performance. [8] The chatbots were Gemini 2.0, DeepSeek V3, Meta AI (Llama 3.3), ChatGPT 3.5, and Grok 2, all in free versions. [8] They collected the 250 responses on February 1, 2025. [8]

Of the responses, 49.6% were problematic: 30% somewhat problematic and 19.6% highly problematic. [8] By chatbot, Grok had 58% problematic responses, ChatGPT 52%, Meta AI 50%, DeepSeek 48%, and Gemini 40%. [8] The differences between chatbots were not statistically significant overall (p=0.566). [8]

The citation results are specific. The chatbots returned 1,013 of 1,250 required references. [8] Median reference completeness was 40%, and no chatbot produced a fully complete and accurate reference list for any prompt. [8] Our calculation: 1,013 of 1,250 is 81.0%, so 237 required references (19.0%) were missing.

This study does not test drug labels. We cite it for citation behavior, which applies to any health topic.

Does Better Model Knowledge Mean Better Patient Outcomes?

Bean and colleagues ran a randomized, preregistered study in the UK with 1,298 adults. [9] The models were GPT-4o, Llama 3, and Command R+. [9] Used alone, each model identified at least one relevant condition in more than 90% of cases. [9] Participants who used the models identified relevant conditions in fewer than 35% of cases, which was significantly worse than the control group. [9]

The study did not test drug information. It shows that a correct model output and a correct patient understanding are different things. For a brand team, that means monitoring the model answer is necessary but does not measure patient outcomes.

Summary Table of the Accuracy Evidence

StudyProduct testedSampleKey resultSource
Andrikyan et al., BMJ Qual Saf 2025Bing Copilot (April 2023)500 answers, 50 drugs26.0% of 484 assessed answers below 100% accuracy; 3.3% fully inconsistent[3]
Long Island University / ASHP 2023ChatGPT (free)39 questions10 satisfactory[4]
Ipema et al., JACCP 20265 chatbots30 questionsOpenEvidence 17.82/25; all below pharmacists[7]
Surbaugh and Moeller, AJHP 20264 toolsLAI antipsychotic dosingOpenEvidence 95%; ChatGPT 40%[7]
Tiller et al., BMJ Open 20265 chatbots (Feb 2025)250 responses49.6% problematic[8]
Bean et al., Nature Medicine 2026GPT-4o, Llama 3, Command R+1,298 UK adultsUsers identified conditions in under 35% of cases[9]

Why the Published Evidence Lags the Models

Our calculation, using dates in the cited papers: Andrikyan and colleagues queried Bing Copilot in April 2023. [3] The paper was first published on October 1, 2024, about 17 months later. [3] Tiller and colleagues collected data on February 1, 2025. [8] The study appeared in BMJ Open volume 16, issue 4, and was covered in the press in April 2026, roughly 14 months later. [8]

Both studies tested products that had been replaced or updated by the time readers saw the results. Peer-reviewed literature gives a baseline. It cannot tell a brand team what a model says this week.

Which Sources Do AI Systems Cite About Pharma Companies?

What Does the 5W Pharma AI Visibility Index Show?

5W’s index covers 25 pharmaceutical companies in Q2 2026. [5] It ran more than 60 consumer and patient prompts across six sub-categories. Each prompt ran five times per engine in clean sessions. The engines were ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. [5] Our calculation: 60 prompts times 5 engines times 5 runs gives at least 1,500 responses.

5W describes its citation-share percentages as estimates and “directional measures of relative visibility.” [5] The company framed prompts at the company and category level, not as treatment-selection queries. [5]

RankCompanyEstimated citation share, Q2 2026
1Eli Lilly12.5%
2Novo Nordisk11.5%
3Pfizer8.5%
4Johnson & Johnson7.0%
5Merck6.0%
6AbbVie5.0%
7AstraZeneca4.0%
8Moderna3.5%

Source: 5W Pharma/Rx AI Visibility Index 2026, top eight of the published top 15. [5]

Concentration: Our Calculation From the 5W Figures

Using 5W’s published percentages, the top two companies sum to 24.0 points. The top five sum to 45.5 points. The top 15 sum to 72.5 points. [5] The page does not say whether shares across all 25 companies total 100%, so we report the sums only as sums of the published figures.

Do Pharma Company Websites Get Cited?

The Worldcom Healthcare Monitor, published July 8, 2026, analyzed ten large pharma companies. [6] It reports that nih.gov accounts for 51% of all citations. [6] Among corporate domains, AbbVie leads at 1.8%, followed by Pfizer at 1.5% and Novartis at 1.3%. [6]

The report names Pfizer as the leader on awareness and share of voice, with a GEO awareness score of 5.3%. [6] Bristol Myers Squibb scores 8.6 out of 10 on sentiment, the highest of the ten, while ranking near the bottom on awareness. [6] The page does not state the number of prompts or platforms. We therefore treat these figures as a company-reported snapshot.

What Citation Share Does Not Measure

Citation share counts how often a source appears. It does not measure whether the cited source supports the claim in the answer. The Tiller audit found that references were frequently incomplete. [8] A high citation share and an accurate answer are separate outcomes.

Our interpretation: if nih.gov supplies half of the citations, then the quality of public NIH-hosted drug records matters more to AI reputation than most corporate websites do. That is an inference from one report, not an established finding.

What Does FDA Regulate When AI Answers a Question About Your Drug?

What Has FDA Said About Drug Advertising Since September 2025?

On September 9, 2025, FDA announced it was “sending thousands of letters warning pharmaceutical companies to remove misleading ads” and “issuing approximately 100 cease-and-desist letters to companies with deceptive ads.” [10] FDA also said it was starting rulemaking to close the “adequate provision” loophole created in 1997. [10]

Covington’s review reports that CDER sent more than 50 Warning Letters and more than 50 Untitled Letters in the third quarter of 2025. [11] It also notes one CBER Untitled Letter to AstraZeneca about FluMist, dated September 9, 2025. [11]

These actions target company promotional communications. An AI answer is not a company communication. Our interpretation is that the enforcement climate raises the cost of any company content that feeds AI systems and could be read as promotional. That interpretation is not an FDA statement.

Can a Company Correct Misinformation About Its Own Drug Online?

FDA’s June 2014 draft guidance on correcting independent third-party misinformation says firms “should respond, if they choose to respond,” to misinformation about their own approved products. [12] The document applies to misinformation created or disseminated by independent third parties, wherever it appears. [12] It is a draft and states that it contains non-binding recommendations. [12]

What Is Not Settled

The 2014 draft predates generative AI. We found no FDA guidance that addresses chatbot answers about a company’s product directly. Whether a model’s answer counts as third-party misinformation under that draft is a legal question. Regulatory and Legal teams should answer it for their own products. We do not offer a legal conclusion.

Does an Inaccurate AI Answer Trigger Adverse Event Reporting?

Not by itself, in our reading. Under 21 CFR 314.80(b), an applicant must promptly review all adverse drug experience information “obtained or otherwise received,” including information from commercial marketing experience and the scientific literature. [13] For serious and unexpected events, the alert report is due no later than 15 calendar days from initial receipt. [13]

A wrong answer in a chatbot is not an adverse event. A patient report that links a chatbot’s advice to a harm is information the company has received. Pharmacovigilance decides whether it meets reporting criteria. The practical point is that a monitoring program needs a defined path to PV before it finds the first case.

Do FDA and EMA AI Principles Cover Chatbot Answers About Drugs?

FDA’s January 2025 draft guidance sets a risk-based credibility framework for AI used to produce data that supports regulatory decisions about drug safety, effectiveness, or quality. [14] On January 14, 2026, FDA and EMA published 10 guiding principles for AI across the drug product life cycle. [15] Both documents address AI that sponsors use to generate evidence. Neither addresses what a consumer chatbot says about a marketed drug. That is our reading of their stated scope. [14][15]

Who Owns AI Reputation Inside a Pharma Company Today?

Which Functions Have a Claim on the Problem?

Eight functions each own a piece. The table is our mapping of function to interest. It is not survey data.

FunctionWhy it has a claimTypical blind spot
Medical InformationAnswers unsolicited drug questions todayRarely monitors third-party AI answers
Medical AffairsOwns scientific exchange and evidence interpretationTreats AI as a channel for others
PharmacovigilanceReviews information received about safetyTuned to case reports, not model outputs
Regulatory AffairsOwns the label, the source of truthLabel changes are not pushed to AI sources
Legal / MLRJudges promotional exposureReviews company content, not model content
Brand Marketing / DigitalOwns web content and visibility metricsMeasures visibility, not accuracy
Corporate CommunicationsOwns reputation and media responseMonitors news, not model answers
Competitive IntelligenceTracks competitor positioningSamples a few prompts, rarely at scale

Four Ownership Failure Modes (Original Taxonomy)

We propose four labels for how ownership breaks down. The labels are ours. They are not an industry classification.

Orphaned

No function monitors the question. A wrong dosing statement can persist because nobody checks. This is the default when a risk crosses function lines.

Split

Each function monitors one aspect. Brand measures visibility. Medical Information checks a few answers. Nobody combines the results, so a brand can rank first in visibility while its safety information is missing from the answer.

Duplicated

Two functions buy separate monitoring tools or run separate prompt sets. The results conflict because the prompts, models, and dates differ. Leaders then cannot tell which result to trust.

Deferred

Each function assumes another one is responsible. Brand assumes Medical Affairs handles accuracy. Medical Affairs assumes Brand handles AI. The absence of a named owner is itself the failure.

A Five-Layer Model of AI Reputation (Original Framework)

Why Five Layers Instead of One Score?

A single AI visibility score hides which part failed. We separate five layers so that each has a verification source and a natural owner. This framework is our proposal.

LayerQuestion it answersVerification sourceNatural owner
1. SourceAre the authoritative documents current and findable?Current FDA label, EMA product informationRegulatory Affairs
2. RetrievalWhich sources do AI systems cite?Citation logs per engineDigital, Corporate Communications
3. AnswerDoes the answer match the label?Label sections, trial publicationsMedical Information, Medical Affairs
4. HarmCould the answer affect patient safety?Boxed warning, contraindications, interactionsPharmacovigilance
5. ClaimsDoes the answer imply unapproved use or comparison?Approved indicationsRegulatory, Legal

How Do the Five Layers Connect to the Evidence?

The Worldcom data speaks to layer 2: nih.gov supplies 51% of citations and corporate domains supply small shares. [6] The Erlangen and Long Island University studies speak to layers 3 and 4. [3][4] The Tiller audit speaks to citation integrity inside layer 2 and 3. [8] Layer 5 has the least public evidence, which is why we flag it as the open legal question.

A Proposed Ownership Model: One Accountable Owner, Five Layer Owners

Who Should Be Accountable for the Program?

Our recommendation: name one accountable executive for the monitoring program, with one budget, one prompt library, and one reporting cadence. The five layers keep their functional owners. The accountable executive does not decide medical or legal questions. That person makes sure someone does. We offer this as opinion based on the failure modes above, not as a finding.

Who Does What? A RACI-Style Proposal

R means responsible, A means accountable, C means consulted.

ActivityMed InfoMed AffairsPVRegulatoryLegalBrand / DigitalCorp Comms
Prompt library designRCCCCRC
Label verification setCCCA / R
Answer accuracy reviewA / RRCC
Safety-gap escalationCCA / RCC
Claims and off-label reviewCCRA / RC
Source and citation actionsCCA / RR
External responseCACR

Source: our proposal. It is not drawn from a survey of company practice.

Escalation Matrix by Severity (Original Classification)

SeverityDefinitionExample patternFirst responder
S1Answer could affect patient safetyWrong interaction statement, missing contraindicationPharmacovigilance
S2Answer contradicts the label without direct safety impactWrong indication or populationMedical Information, Regulatory
S3Answer is outdated or incompleteSuperseded label languageRegulatory, Medical Information
S4Brand is absent or misframed with no label conflictCompetitor named, brand omittedBrand, Competitive Intelligence

Source: our classification, for use as a starting point. Each company should calibrate it with its own safety and legal teams.

Scenario Analysis: Three AI Errors and Who Should Handle Them

Scenario A: The Answer States a Wrong Drug Interaction

A monitoring run finds an AI answer saying no interaction exists between your drug and a common drug. The current label lists a clinically significant interaction. This resembles the Paxlovid and verapamil example in the Long Island University study. [4]

Who Should Act in Scenario A

Capture the prompt, model, date, and full answer. Classify it as S1. Notify Pharmacovigilance, which decides whether any reporting duty applies under its procedures. [13] Medical Information prepares a label-based response for inquiries. Regulatory confirms the label language is current.

Scenario B: The Answer Describes an Unapproved Use

An AI answer says your drug is appropriate for an indication that FDA has not approved. The answer cites a mix of sources, none from your company.

Who Should Act in Scenario B

Classify it as S2, and treat it as a layer 5 claims risk. Regulatory and Legal decide whether any response is appropriate and under which framework. The 2014 draft guidance on third-party misinformation is the document to review first, with its draft status in mind. [12] No one should post a correction until Legal approves the wording.

Scenario C: A Competitor Appears and Your Brand Does Not

You ask a category question and the AI answer names two competitors, not your brand. This matches the pattern in 5W’s data, where citation share concentrates in a few companies. [5]

Who Should Act in Scenario C

Classify it as S4. Brand and Digital review which sources the engine cited. They check whether public records and publications describing the drug are current and findable. The action is about source quality, not about asking the model to change its answer.

How to Monitor AI Drug Information: A Reproducible Method

This section describes a method. It is not a description of tests we ran for this article.

What Should a Monitoring Protocol Record?

  • Platform and model version for each run
  • Exact prompt text and any system settings
  • Date and time of the run
  • Whether web search was on, and whether citations were requested
  • Geographic context and language
  • The full response and every cited URL

How Should the Verification Set Be Built?

Build a verification set from the current label: indications, dosing, contraindications, warnings, interactions, and adverse reactions. Record the label version and date. When the label changes, rerun the prompts. The Surbaugh result, 95% for one tool and 40% for another, shows why each platform needs its own results. [7]

How to Handle Duplicate and Variable Responses

Models return different text for the same prompt. 5W ran each prompt five times per engine in clean sessions. [5] We suggest the same practice: repeat each prompt, store every response, and report the share of runs with a discrepancy, not a single pass or fail.

How to Classify Discrepancies

Use two reviewers per discrepancy. Assign severity S1 to S4 from the matrix above. Record disagreements and resolve them through a named medical reviewer. Do not let an AI system grade another AI system’s medical accuracy without human review.

Limitations of Any Monitoring Program

A sample of prompts does not represent all questions. Results change as models change. A prompt written by an analyst may differ from what patients type. Report the sample size and dates with every figure, and do not generalize beyond them.

Where DrugChatter Fits

DrugChatter is one approach to this protocol. DrugPatentWatch’s description says it runs structured query batteries across ChatGPT, Gemini, Claude, Copilot, and Perplexity, then logs and scores outputs for accuracy, regulatory alignment, sentiment, and competitive framing. [16] That page is a vendor description and reports no methodology for its audit statistics, so we do not use those statistics here.

A tool like this speeds up capture and comparison. It does not replace the label, the clinical evidence, or a medical reviewer. The same applies to any vendor.

What This Means for Each Function

What This Means for Medical Information

Your team already answers unsolicited questions with label-based responses. AI answers are the same questions arriving through a channel you do not control. Add a quarterly comparison of model answers against your standard response documents.

What This Means for Pharmacovigilance

Write down, before any finding arrives, how a monitoring flag reaches PV and who decides on reporting. Section 314.80(b) refers to information “obtained or otherwise received,” so the intake path needs to exist. [13]

What This Means for Regulatory Affairs and Legal

Decide in advance how your company reads the 2014 third-party misinformation draft for AI outputs. [12] Define who approves any public correction. Track the “adequate provision” rulemaking, because FDA said it was starting it in September 2025. [10]

What This Means for Brand and Digital Teams

Pair visibility metrics with accuracy checks. A brand with high citation share and wrong safety statements has a risk, not an asset. The Worldcom data suggests corporate domains hold small shares of citations. [6] Spend effort on the authoritative records AI systems actually cite.

Timeline: The Documents Behind This Analysis, 2014 to 2026

The timeline lists dated events from the cited sources. It shows how the regulatory documents and the evidence arrived in different years.

DateEventTypeSource
June 2014FDA draft guidance on correcting third-party misinformationRegulatory (draft)[12]
April 2023Bing Copilot queried in the Erlangen studyData collection[3]
December 2023Long Island University ChatGPT results reportedStudy[4]
October 1, 2024Erlangen study first publishedStudy[3]
January 7, 2025FDA draft guidance on AI for regulatory decision-makingRegulatory (draft)[14]
February 1, 2025Data collected for the BMJ Open chatbot auditData collection[8]
September 9, 2025FDA announces advertising enforcement and rulemaking planRegulatory[10]
January 7, 2026OpenAI health-usage report covered by the pressUsage data[2]
January 11, 2026Google removes AI Overviews for named liver-test queriesPlatform action[17]
January 14, 2026FDA and EMA publish 10 AI guiding principlesRegulatory[15]
February 10, 2026Nature Medicine randomized study covered by the pressStudy[9]
March 25, 2026KFF poll publishedUsage data[1]
Q2 20265W tracks 25 companies across five AI enginesVisibility data[5]
July 8, 2026Worldcom Healthcare Monitor article publishedVisibility data[6]

Our calculation: from the June 2014 draft guidance to the March 2026 KFF poll is about 11 years and 9 months. The only FDA document in the table that addresses third-party misinformation is the 2014 draft. The two AI documents address AI used to generate regulatory evidence. [12][14][15]

How Can a Team Start in 30 Days?

This is our suggested sequence, offered as practical opinion. It is not drawn from a cited program.

  1. Name the accountable executive and the five layer owners.
  2. Pick two to four products, including one with a boxed warning or a significant interaction profile.
  3. Write 20 to 40 prompts per product in the words patients and clinicians use. Include dosing, interaction, safety, and indication questions.
  4. Build the label verification set and record the label version.
  5. Run each prompt on each platform several times. Store every response and every cited URL.
  6. Review discrepancies with two reviewers, assign severity, and route S1 findings to Pharmacovigilance.

Report the first results with the sample size and dates attached. A small, well-documented sample is more useful to a skeptical executive than a large score with no method.

What Happens Next?

Three things are documented and one is not. FDA’s 2025 draft AI guidance and the 2026 FDA-EMA principles address AI used in drug development, and they continue to shape how regulators think about AI credibility. [14][15] FDA’s advertising enforcement is active. [10] Usage is growing in the sources we reviewed. [1][2]

What is not documented: any FDA position on consumer chatbot answers about marketed drugs. We found none. Any claim that a specific rule will arrive is a forecast, and we do not make one.

What Did Google Do When Health Summaries Were Challenged?

In January 2026, the Guardian reported that Google’s AI Overviews gave liver blood-test reference ranges that did not account for factors such as age, sex, ethnicity, or nationality. Google removed AI Overviews for the specific queries the Guardian named. [17] Google told the Guardian that an internal clinician team found “in many instances, the information was not inaccurate and was also supported by high quality websites.” [17] The example shows that platforms do respond to documented findings, and that platforms and reviewers can disagree about what counts as an error.

Limitations of This Analysis

  • We ran no original AI queries. The accuracy figures come from published studies with different methods, models, and dates.
  • Several cited studies cover health topics beyond drug labels. We noted where this applies. [8][9]
  • The 5W and Worldcom figures are company-reported, and each source states or implies its own limits. [5][6]
  • The OpenAI figures come from press coverage of the report. [2]
  • The ownership model, taxonomy, severity scale, and RACI table are our proposals. No source validates them.
  • This article is not legal advice. Companies should take regulatory and legal questions to their own counsel.

Key Takeaways

  • 32% of US adults used AI chatbots for health information in the past year (KFF, 2026). [1]
  • In the Erlangen study, 26.0% of 484 assessed Bing Copilot answers fell below 100% accuracy against reference drug data. [3]
  • Pharmacists rated 10 of 39 ChatGPT answers to real drug questions satisfactory. [4]
  • One tool scored 95% and another 40% on antipsychotic injectable dosing, so results differ by platform. [7]
  • The nih.gov domain supplied 51% of citations in the Worldcom report; the top corporate domain held 1.8%. [6]
  • We found no FDA guidance on chatbot answers about marketed drugs. The 2014 third-party misinformation guidance is a non-binding draft. [12]
  • Our proposal: one accountable executive, five layer owners, and a severity scale that routes S1 findings to Pharmacovigilance first.

FAQ

Is an inaccurate AI answer about my drug an adverse event?

No. An adverse event involves a patient and a drug. A wrong answer is a communication problem. If someone reports harm connected to chatbot advice, that report is information your company received, and Pharmacovigilance decides how to handle it under 21 CFR 314.80. [13]

Does FDA require companies to monitor what AI says about their drugs?

We found no such requirement. FDA’s 2014 draft guidance on third-party misinformation says firms should respond “if they choose to respond,” which makes correction voluntary. [12]

How many people use AI for health questions?

KFF found 32% of US adults used AI chatbots for health information in the past year. [1] OpenAI reported more than 40 million people send health prompts to ChatGPT daily. [2]

Which AI model is the most accurate for drug information?

No single answer holds across studies. In one study OpenEvidence scored highest among five chatbots and still scored below pharmacists. [7] In another, tool accuracy ranged from 95% to 40% on a dosing task. [7] Test each platform for your own products.

Do AI chatbots invent references?

Sometimes. Coverage of the Long Island University study reports nonexistent references in ChatGPT answers. [4] In the BMJ Open audit, chatbots returned 1,013 of 1,250 required references, and none produced a fully complete and accurate list. [8]

Can a company ask an AI platform to correct an answer?

In our sources, the one documented removal came after a press investigation. Google removed AI Overviews for named liver-test queries after the Guardian’s report. [17] We found no documented company-submission process in the sources we reviewed.

Why do corporate websites get so few AI citations?

The Worldcom report found corporate domains each held a small share, with AbbVie highest at 1.8%, while nih.gov held 51%. [6] The report does not state the reason. We can only say that AI systems in that sample drew heavily on public institutional sources.

Does a high AI visibility rank mean the answers about my brand are correct?

No. 5W’s index measures estimated citation share. [5] It does not measure whether answers match the label. Accuracy needs a separate check against the current prescribing information.

Who should lead an AI monitoring program?

Our proposal is one accountable executive with functional owners for each layer. That structure is our opinion. No public survey in our search shows which structure companies use today.

Do patients understand AI answers correctly?

Not reliably. In a randomized study of 1,298 UK adults, participants using LLMs identified relevant conditions in fewer than 35% of cases, although the models alone exceeded 90%. [9] The study did not test drug information.

References

  1. KFF. (2026, March 25). Poll: 1 in 3 adults are turning to AI chatbots for health information, equaling the share who use social media for health. https://www.kff.org/health-information-trust/poll-1-in-3-adults-are-turning-to-ai-chatbots-for-health-information-equaling-the-share-who-use-social-media-for-health/
  2. Medical Economics. (2026, January 7). 40 million people now use ChatGPT daily for health questions, OpenAI report finds [Coverage of OpenAI, AI as a Healthcare Ally, January 2026]. https://www.medicaleconomics.com/view/40-million-people-now-use-chatgpt-daily-for-health-questions-openai-report-finds
  3. Andrikyan, W., Sametinger, S. M., Kosfeld, F., Jung-Poppe, L., Fromm, M. F., Maas, R., & Nicolaus, H. F. (2025). Artificial intelligence-powered chatbots in search engines: A cross-sectional study on the quality and risks of drug information for patients. BMJ Quality & Safety, 34(2), 100-109. https://doi.org/10.1136/bmjqs-2024-017476
  4. U.S. Pharmacist. (2023, December 20). Pharmacists should be wary of ChatGPT information on drugs. https://uspharmacist.com/article/pharmacists-should-be-wary-of-chatgpt-information-on-drugs
  5. 5W Public Relations. (2026). Pharma/Rx AI Visibility Index 2026. https://www.5wpr.com/ai-visibility-index/pharma-rx-ai-visibility-index-2026/
  6. Nawrocki, L. (2026, July 8). The invisible race: How pharmaceutical brands win or lose in the age of AI search. Worldcom Group, Healthcare Monitor Pharmaceuticals 2026. https://worldcomgroup.com/insights/2026-healthcare-monitor-pharmaceutical-brands-ai-search/
  7. Pharmacy Times. (2026, July 13). AI and drug information: Where OpenEvidence stands out and where pharmacists still win [Summarizing Ipema et al., 2026, Journal of the American College of Clinical Pharmacy, 9(6), e70231, and Surbaugh & Moeller, 2026, American Journal of Health-System Pharmacy, https://doi.org/10.1093/ajhp/zxag089]. https://www.pharmacytimes.com/view/ai-and-drug-information-where-openevidence-stands-out-and-where-pharmacists-still-win
  8. Tiller, N. B., Marcon, A. R., Zenone, M., et al. (2026). Generative artificial intelligence-driven chatbots and medical misinformation: An accuracy, referencing and readability audit. BMJ Open, 16(4), e112695. https://doi.org/10.1136/bmjopen-2025-112695
  9. Bean, A. M., et al. (2026). Reliability of LLMs as medical assistants for the general public: A randomized preregistered study. Nature Medicine. https://pubmed.ncbi.nlm.nih.gov/41663592/ (results as summarized by News-Medical, 2026, February 10: https://www.news-medical.net/news/20260210/Large-language-models-excel-in-tests-yet-struggle-to-guide-real-patient-decisions.aspx)
  10. U.S. Food and Drug Administration. (2025, September 9). FDA launches crackdown on deceptive drug advertising. https://www.fda.gov/news-events/press-announcements/fda-launches-crackdown-deceptive-drug-advertising
  11. Covington & Burling LLP. (2025, October). FDA advertising and promotion enforcement activities update. https://www.cov.com/en/news-and-insights/insights/2025/10/fda-advertising-and-promotion-enforcement-activities-update
  12. U.S. Food and Drug Administration. (2014, June). Internet/social media platforms: Correcting independent third-party misinformation about prescription drugs and medical devices (Draft guidance, Docket No. FDA-2014-D-0447). https://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM401079.pdf
  13. Postmarketing reporting of adverse drug experiences, 21 C.F.R. § 314.80. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-D/part-314/subpart-B/section-314.80
  14. U.S. Food and Drug Administration. (2025, January 7). Considerations for the use of artificial intelligence to support regulatory decision-making for drug and biological products (Draft guidance). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
  15. Regulatory Affairs Professionals Society. (2026, January 14). EMA, FDA issue joint AI guiding principles for drug development. https://www.raps.org/resource/ema-fda-issue-joint-ai-guiding-principles-for-drug.html
  16. DrugPatentWatch. (n.d.). What AI knows about your drug vs. your competitors. https://www.drugpatentwatch.com/blog/what-ai-knows-about-your-drug-vs-your-competitors/
  17. Ha, A. (2026, January 11). Google removes AI Overviews for certain medical queries. TechCrunch. https://techcrunch.com/2026/01/11/google-removes-ai-overviews-for-certain-medical-queries
DrugChatter - Know what AI is saying about your drugs
Scroll to Top