When a patient or a prescriber asks an AI assistant about a branded medicine, the answer is composed from whatever the system retrieved or remembers. It may describe the brand correctly. It may also attach an indication the brand is not approved for, merge it with a competitor, quote an old dosage or leave out a safety warning. For a pharma company in India, that raises two questions at once: is the answer accurate, and what is the company allowed to do about it?
This article is for brand managers, medical affairs, regulatory affairs and digital leads at pharmaceutical companies. It is a teardown method, not a set of findings. Results depend on the molecule, the market and the day, so we describe how to run the audit and read it, and we do not report numbers. Everything here assumes your regulatory and medical teams review each step. Our pharma marketing guide covers the wider channel context.
Set the boundary before you start
Two rules shape the whole exercise.
First, the audit is observational. You are recording what AI systems say. You are not trying to make them say promotional things. Wording that would be a problem on your own page is no less a problem because an AI is repeating it.
Second, compliance work comes before publication work. In India, the Drugs and Magic Remedies (Objectionable Advertisements) Act, 1954 restricts advertisements of drugs for specified purposes and prohibits misleading drug advertisements. The Department of Pharmaceuticals' UCPMP 2024 portal describes the code as mandatory, sets ethical norms for company interactions with healthcare practitioners, and runs enforcement through association ethics committees with appeals to the Department. That portal text covers industry and practitioner interactions and does not address public advertising, so confirm with regulatory counsel how each rule applies to a given page and audience. Prescription brands generally need extra care in what is shown to the public. This article does not give legal advice.
Build the query set
Draft about fifty queries, split by who is asking. Keep wording natural and avoid anything resembling a real patient's circumstances.
| Group | Example pattern (use your own brand) | What you are testing |
|---|---|---|
| Brand name only | "What is [Brand]?" | Basic identity: molecule, class, manufacturer |
| Brand and indication | "Is [Brand] used for [condition]?" | Correctness of approved indications |
| Brand and safety | "[Brand] side effects", "[Brand] and pregnancy" | Presence and accuracy of warnings |
| Brand and dose | "[Brand] dose for adults" | Whether the answer offers dosing without prescriber context |
| Brand versus comparator | "[Brand] vs [competitor brand]" | Comparative claims and sources |
| Generic and brand | "[Molecule] brands in India" | Whether the brand appears, and how |
| Availability | "Where to buy [Brand]", "[Brand] price India" | Retail and pricing information, which changes |
| Professional framing | "Prescribing information [Brand]" | Whether the official document is reached |
Run each query on the assistants your audience uses, with search and browsing turned on and off if the product allows it, and repeat each three times because answers vary. Record the date, surface, mode and full text. Do not enter real patient details into any tool.
Score each answer on five checks
Have a pharmacist or medical affairs reviewer score the answers, not the marketing team. Use a simple sheet.
- Identity. Is the molecule, strength range, dosage form and manufacturer correct?
- Indications. Does the answer list only approved indications? Flag extra, missing and off-label ones separately, because off-label content can be a safety concern even when medically plausible.
- Safety context. Are contraindications and key warnings present, and is there any statement that could be read as reassurance beyond the label?
- Currency. Does the answer reflect the current label, or an older version, a withdrawn strength or a discontinued presentation?
- Sources. Which URLs were cited or clearly used? Classify each as company, regulator, hospital, medical reference, pharmacy retailer, forum, news or unknown.
Record severity as well. A wrong manufacturer name is minor. A wrong indication, a missing contraindication or a dosing statement presented as advice is a potential patient-safety issue and goes to medical and pharmacovigilance leads the same day.
Read the source pattern
The most useful output of the audit is usually not the errors. It is the list of pages the AI relied on. Group them.
- Retailer and e-pharmacy pages. Often carry short, template-generated descriptions that may be outdated. You may not control them, but you can supply accurate information through legitimate channels.
- Third-party health information sites. Check whether their text is consistent with the current label.
- Forums and community threads. These carry personal experience and rumour. You should not edit or astroturf them. Reddit and similar communities have their own rules and expect disclosure of affiliation.
- Company pages. Is your prescribing information reachable and indexable, or hidden behind a script, a login or a PDF with no text layer?
- Regulator pages. Are the relevant regulator pages cited for your brand, or are older documents dominant?
A hypothetical illustration: the audit finds that for a fictional brand "Brandex", three of four assistants describe an indication that was removed from the label several years ago, and every one cites the same retailer page. The fix is not another blog post. It is to correct the retailer's source information through the supply chain and to publish a clear, current label page that can be retrieved.
What a company can do, within promotion rules
Options, in rough order of safety:
- Publish current, accessible prescribing information. Put the approved label on a crawlable page, with the text in HTML, not only in an image or a PDF. Add the revision date. Make sure the page is clearly labelled for healthcare professionals if that is its audience, and follow your regulatory team's rules about access.
- Keep a medical information page that states facts only. Brand name, molecule, class, approved indications, key safety information and a route to the medical information team. No superlatives, no comparative claims, no outcome language.
- Use structured data conservatively. schema.org's Drug type includes
activeIngredient,prescriptionStatus,warning,manufacturerandbrand. Marking up facts that already appear in the label helps systems read them unambiguously. Do not use markup to add claims that the page text does not carry. - Correct third-party sources. Contact retailers, aggregators and reference sites with the current label and ask them to update. Keep a log of requests and replies.
- Use platform feedback channels for factual errors. Google documents a way to report problems with an AI Overview through the thumbs-down control. Other assistants have in-product feedback tools too. Give a plain factual description, cite the label and avoid promotional language. These channels do not guarantee a change.
- Escalate safety-significant errors internally first. If an answer states something that could cause harm, medical and regulatory teams decide on wording and routing.
What not to do: do not publish a page whose purpose is to be quoted by AI with claims you could not make elsewhere. Do not seed forums. Do not write comparative content that has not passed your normal review.




