How visible are pharma brands in AI answers?
Updated 2026-07-23
What we measured
How visible pharma brands in AI answers actually are is an empirical question, so we measured it. We took two prescription medicines with large patient populations, one a GLP-1 used for diabetes and weight management and one an immunology biologic, and built a fixed set of 24 questions covering the things people actually ask: side effects, how well it works, how it compares with alternatives, dosing, cost, what happens if you stop, and whether you can get it more cheaply somewhere else. Each question went to ChatGPT, Claude, Gemini and Perplexity with web search enabled, and each was asked three separate times per engine so we could measure consistency rather than catch a single lucky or unlucky draw. That produced 288 recorded answers and 2,321 citations, where a citation is a source the engine itself attached to its own answer. We then classified every cited domain by what kind of source it is: the manufacturer's own properties, regulators, peer-reviewed literature, medical institutions, consumer drug references, telehealth and compounding services, patient forums, litigation marketing, and news. The brands are anonymised here because the purpose is the pattern, not a scorecard on any one company. This page describes how AI systems cite sources and is general information, not medical advice.
The manufacturer's own site is usually not in the room
The most consistent finding across both medicines is that the company that makes the drug is a minority voice in the answer about it. Manufacturer-owned domains accounted for 8.0% of all cited sources for the GLP-1 brand and 19.6% for the immunology brand. On most individual questions no manufacturer or product-label source appeared at all, meaning the answer a patient reads was assembled entirely from third parties. The gap is not explained by the manufacturer lacking content, since both companies publish extensive product sites, prescribing information and HCP-facing material. It is explained by what the retrieval step reaches for: consumer drug references such as Drugs.com, GoodRx, Healthline and WebMD are structured as direct answers to exactly the questions people ask, while manufacturer sites are structured around the product and wrapped in regulated language that reads poorly as an extractable passage. The engine is not hostile to the label. It simply finds a cleaner answer somewhere else first.
Who fills the gap depends on the category
The two medicines had almost opposite risk profiles, which is the finding most likely to matter to a specific brand team. For the GLP-1, telehealth prescribers and compounding services made up 17.2% of all citations, outnumbering the manufacturer's own sources by 2.15×, while patient forums were negligible at 1.8%. For the immunology biologic the pattern inverted: telehealth was marginal at 3.1% while patient forums, chiefly Reddit, rose to 7.3% and were among the most-cited individual domains. Both categories also surfaced plaintiff-firm litigation marketing on safety questions, where a patient asking what the risks are can receive an answer partly grounded in injury-claim advertising. The practical consequence is that a generic AI visibility programme is close to useless here. One brand's problem is a commercial supply channel selling around it; the other's is unmoderated patient discussion being treated as evidence. Those require different responses, and you cannot tell which you have without measuring the source mix rather than a single visibility score.
The engines disagree with each other, sharply
Treating any one assistant as a proxy for AI visibility produces a badly distorted picture. On the identical question set, the rate at which each engine cited a manufacturer source ranged from 15% on ChatGPT to 64% on Perplexity for the GLP-1 brand, and from 26% on ChatGPT to 82% on Gemini for the immunology brand. Claude sat at 24% and 31% respectively. The ordering was stable across both medicines: Gemini and Perplexity reached for manufacturer and label sources far more readily than ChatGPT and Claude, which leaned more heavily on consumer references and their own synthesis. This is a retrieval-behaviour difference, not a quality judgement about the engines. But it means a brand team that checks ChatGPT and concludes it is invisible, or checks Perplexity and concludes it is fine, will be wrong in opposite directions. Any credible measurement has to run the same fixed question set across multiple engines and report them separately rather than averaging them into one number.
Asking once is not measuring
Answer engines are non-deterministic, so the same question asked twice can retrieve different sources and produce a different answer. When we repeated every question three times, the effect was real but bounded: mean agreement across samples was 96%, and only a small minority of questions flipped between citing a manufacturer source and not. Critically, single-run measurement was wrong in both directions. One brand scored lower on a single pass than it did across three, and the other scored higher, so the error is not a consistent bias you can correct for with a fudge factor. The useful output of repeat sampling is not just a more accurate headline number but the distinction it exposes between three states: questions where you are always cited, questions where you are never cited, and questions that flip between runs. The third group is contested ground where the engine is nearly indifferent between your source and someone else's, and it is usually the cheapest place to win. A single-sample report collapses that group into whichever side it happened to land on.
What to do if you own a pharma brand
Start by measuring the source mix rather than a visibility score, because the mix tells you what kind of problem you have. If consumer drug references dominate, the gap is structural: your content answers product questions in regulated prose while the engine wants a direct, self-contained answer to a patient's actual phrasing, and that is addressable with material that stays entirely within your existing approved claims. If telehealth and compounding sources dominate, the issue is commercial and legal as much as editorial, and it belongs in front of the people who handle channel and brand protection, not only the digital team. If patient forums dominate, you are looking at a medical-information and community-engagement question. Run the same fixed question set on a schedule so you are tracking direction rather than a snapshot, sample each question more than once so the numbers survive somebody re-running them in front of you, and separate the engines rather than averaging them. None of this requires making a claim about the medicine. It requires knowing what is already being said about it. If you own a pharma brand and want the same measurement run against your medicine, that work is described on our pharma AI visibility page.
Frequently asked questions
Do AI assistants cite the drug manufacturer's own website?
Sometimes, but far less often than most brand teams expect. Across the two medicines we measured, manufacturer-owned domains accounted for 8.0% of all cited sources on one brand and 19.6% on the other. On the majority of questions the answer was assembled entirely from third parties, most often consumer drug references such as Drugs.com, GoodRx and Healthline.
Do different AI engines give different answers about the same medicine?
Yes, and the gap is large. On the same fixed question set, Gemini cited a manufacturer source on 82% of questions for one brand while ChatGPT did so on 26%. Measuring a single engine and treating it as your AI visibility will misstate the picture substantially.
Is asking an AI engine a question once enough to measure visibility?
No. Answer engines are non-deterministic, so a single run is a draw rather than a measurement. When we repeated every question three times, one brand's score moved up and the other moved down compared with single-run measurement. Repeat sampling and a reported hit rate are what make the number defensible.
Is this medical advice?
No. This is general information about how AI systems cite sources. It describes the behaviour of answer engines, not the safety, efficacy or suitability of any medicine. Questions about a specific medicine belong with a qualified healthcare professional.