How ChatGPT, Claude, and Google AI Overviews Select Sources in Pharma GEO

Pharma marketers spent two decades optimizing for ten blue links. Increasingly, the “search result” a patient, caregiver, or HCP sees is a single synthesized paragraph with two or three footnotes attached. Which footnotes appear is not random, and it is not consistent across platforms. ChatGPT, Claude, and Google AI Overviews each run a different retrieval pipeline, weigh different authority signals, and — in a category as regulated as pharma — end up citing a strikingly different mix of clinical, government, editorial, and community sources for the same question.

Why Pharma GEO Is Not Generic GEO: The Regulatory Backdrop

Before comparing platforms, it is worth being explicit about the constraints that make pharma GEO a different discipline from GEO in any other category. Several apply directly to any content strategy aimed at AI answer engines.

• First, scope: direct-to-consumer (DTC) advertising of prescription drugs is fully legal in only two countries in the world, the United States and New Zealand.¹ Everywhere else, including the European Union, branded prescription-drug promotion to the general public is prohibited, and GEO efforts in those markets are necessarily limited to unbranded disease-awareness content rather than the branded citation competition described later in this article. Most of the platform-behavior data here comes from U.S.-market research and should be read with that scope in mind.
• Second, fair balance: 21 CFR Part 202 requires that any presentation of a drug’s benefits be accompanied by equally prominent risk information.² An AI answer engine that synthesizes a one-paragraph summary from a page has no obligation to preserve that balance, and Google’s AI Overviews have already drawn documented scrutiny for stripping context from health answers.⁹ FDA enforcement shows how quickly benefit-without-risk framing becomes a violation even in ordinary social media: in May 2025, the agency issued a warning letter to a pharmaceutical company because a CEO’s Instagram post promoted a prescription drug’s benefits without any accompanying risk information.⁷ A pharma GEO strategy cannot simply try to get benefit claims cited more often; content has to be structured so risk information survives summarization, or the citation itself becomes a liability.
• Third, promotional material submission: 21 CFR 314.81 requires that promotional labeling and advertising be submitted to FDA at the time of initial dissemination.³ This applies regardless of channel, which means GEO content — trial summaries, indication pages, comparison content — goes through the same Medical, Legal, and Regulatory (MLR) review and submission process as any other promotional material. There is no “AI content” exemption.
• Fourth, off-label questions: FDA’s guidance on responding to unsolicited requests for off-label information directs companies not to answer off-label clinical questions publicly, but instead to route them to a Medical Information channel.⁴ If a company’s content, or a company representative, encounters an off-label question inside a public forum, thread, or comment section that an AI platform might later ingest, the same redirection rule applies there as anywhere else.

• Fifth, adverse events: current FDA policy does not require companies to proactively monitor third-party websites for adverse events, but once a company becomes aware of a reportable event — including one described in a comment on the company’s own social post — standard postmarketing reporting timelines apply.^6,8 The Sprout Pharmaceuticals warning letter is also instructive here: FDA noted the promotional post had not been submitted to the agency at time of first use, underscoring that ordinary submission and monitoring obligations do not relax just because a channel is informal or fast-moving.^7,8 Any GEO-adjacent social monitoring a pharma company runs should be built primarily as a pharmacovigilance safeguard, not a citation-optimization exercise.

How Each Platform Actually Retrieves Its Sources

The starting point is architectural, not editorial.

• By OpenAI’s own documentation, ChatGPT produces inline citations only when its search feature is active for a given response; most responses answer from training data and carry no links at all.¹⁰ A large-scale analysis of citation data covering ChatGPT’s top-cited domains found Wikipedia accounts for nearly half (47.9%) of citations among ChatGPT’s leading sources, consistent with a documented pull toward broad-consensus references.¹²
• Claude, in most consumer contexts, does not browse by default and does not cite sources unless retrieval is configured; where it is enabled, platform comparisons consistently describe Claude as more conservative and more oriented toward a small number of structured, in-depth sources than toward broad consensus.¹² In life-sciences deployments specifically, Anthropic has documented connectors that give Claude access to bioRxiv, medRxiv, ChEMBL, and Open Targets — a research-database layer with no equivalent in ChatGPT’s consumer product.¹¹
• Perplexity performs live retrieval on every query and, per the same large-scale citation analysis, leans further on real-time community content than any other platform, with Reddit representing a distinctly larger share of its top citations than any other platform studied.¹²
• Google AI Overviews run a passage-level extraction system that is measurably distinct from classic organic ranking, and the size of that gap is itself unstable: in a large Ahrefs analysis of 863,000 keyword searches and 4 million AI Overview URLs, the share of AI Overview citations that also ranked in Google’s own top 10 fell from 76% to 38% within roughly seven months, a shift Ahrefs links in part to Google’s move to a newer underlying model in early 2026.^13,14

Government and Clinical Authority: Who Actually Trusts PubMed, NIH, and .gov

This is where the platforms diverge most sharply — and counterintuitively. Enterprise SEO research firm BrightEdge’s analysis of healthcare queries found ChatGPT draws 27% of its healthcare citations from government domains (CDC, NIH, FDA) and medical specialty associations, but only 1% from elite hospital systems.¹⁷

Google AI Overviews inverts that pattern almost exactly: 33% from elite hospital systems and academic medical centers, only 10% from government sources.¹⁷ Same YMYL category, opposite default definition of authority. On PubMed specifically, Perplexity’s real-time retrieval architecture is widely characterized in platform research as the most oriented toward primary clinical literature and named institutional sources, in contrast to consensus-style aggregation.^12,18

Claude‘s pattern is different again: rather than citing government domains at high volume, Anthropic’s documented life-sciences connectors plug Claude directly into research infrastructure — bioRxiv, medRxiv, ChEMBL, Open Targets — positioning it closer to a specialist research-database model than a general web-citation model, though this reflects deliberate product integrations rather than a measured default consumer citation mix.¹¹

The practical, compliance-aware read for pharma GEO: publishing trial data in a crawlable, MLR-approved format on your own domain — not only behind a journal paywall — has outsized value for ChatGPT and Perplexity’s government/primary-source leanings, provided the risk information is structured to survive summarization. Getting that same data represented on partner hospital-system and academic medical center pages has outsized value for Google AI Overviews, and because those pages are typically authored and reviewed by the institution rather than the manufacturer, they carry less direct fair-balance exposure for the sponsor.

Reddit and User-Generated Content: Influential, and Not a Channel This Article Recommends for Pharma

A large-scale analysis of AI platform citations — spanning ChatGPT, Google AI Overviews, and Perplexity — found Reddit to be the single most commonly cited domain across the platforms studied, with Perplexity showing the heaviest reliance: Reddit accounts for 46.7% of its top citations.¹² That same research found only 11% of domains are cited by both ChatGPT and Perplexity, underscoring how differently the platforms weigh community content .¹² Independent large-scale tracking by Ahrefs of Google’s own AI Overviews shows citation composition can shift dramatically within months of a single model or indexing change, which should caution against treating any single platform’s current community-content reliance as stable.^13,14 Some generic GEO playbooks respond to Reddit’s influence by recommending that brands seed community discussion directly, including through accounts that do not disclose a brand relationship.

This article does not extend that recommendation to pharma, and not only because of the volatility above. An undisclosed, brand-directed post about a prescription product would be promotional labeling the moment it references the product, which triggers the same fair-balance and FDA submission obligations as any other promotion — obligations a fast-moving, character-limited forum thread is poorly suited to satisfy.^2,3 It would also implicate FTC Endorsement Guide disclosure requirements if any paid or incentivized account were involved. Separately, the moment company personnel see a forum comment describing a side effect, adverse-event reporting obligations apply regardless of whether the company was actively monitoring for it.^6,8 Reddit is a real and growing input to AI-generated health answers, and pharma companies have good reason to monitor it — for misinformation and pharmacovigilance signals — but “monitor” and “seed for citation” are different activities with very different compliance profiles, and this article treats only the former as advisable for regulated brands.

For misinformation specifically, FDA’s current draft guidance gives companies a narrow, voluntary path to respond to false third-party claims about their approved products, including claims that appear on social platforms, but the response must be clearly identified, limited to the specific inaccuracy, and non-promotional in tone to qualify for the associated enforcement policy.^4,5 That is a materially different activity from proactive community engagement aimed at improving citation share.

Wikipedia, Journalism, and Video: The Remaining Layers

Wikipedia is close to foundational for ChatGPT as explained here above.¹² A separate large-scale comparison of Google’s own AI Mode against AI Overviews found Wikipedia’s citation share differs even within Google’s own products — 28.9% of AI Mode citations versus 18.1% of AI Overview citations — and that AI Mode cites Quora roughly 3.5 times more often than AI Overviews do, evidence that even a single company’s AI surfaces are not internally consistent.¹⁶

Claude’s citation behavior generally skews toward established, editorially-vetted, long-form sources rather than the fastest-moving content, consistent with broader academic findings that answer engines vary widely in citation behavior, accuracy, and freshness preference even when covering the same underlying query set.²²

That variability is confirmed by a separate Ahrefs analysis of 17 million citations across seven AI platforms, which found Google’s AI Overviews prefer distinctly older content (a median age near four years) while ChatGPT shows the strongest preference for recently published material of any platform studied.¹⁵ This cuts against the common GEO advice to simply “publish more recent content” as a universal tactic — for Google AI Overviews specifically, durable, well-established pages may outperform newly published ones.

On video, Google AI Overviews is the strongest outlier: In January 2026, YouTube was the single most-cited domain in AI Overviews overall, having grown 34% over six months in Ahrefs’ tracking, and a separate study found YouTube ranked first in AI Overview citations for health topics specifically while ranking only 11^th in the corresponding organic results for the same queries.^14,19

The GLP-1 Case Study: What Concentrated Citation Looks Like in Practice

The clearest pharma-specific evidence comes from independent trade-press reporting rather than a vendor’s own release. Fierce Pharma, reporting on citation-share research covering ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews, found Eli Lilly held the highest estimated AI citation share among pharmaceutical manufacturers at 12.5%, with Novo Nordisk second at 11.5% — both driven by consumer interest in their GLP-1 drugs (Wegovy, Ozempic, Zepbound, and Mounjaro).²⁰

The same reporting highlighted a structural disconnect: AbbVie and Johnson & Johnson are the largest television DTC advertisers in the U.S., yet neither leads in AI citation share, indicating that AI answer engines are not simply amplifying paid media weight.²⁰

Fierce Pharma’s reporting attributes durable citation share to peer-reviewed trial publication and sustained coverage in outlets like Reuters, STAT, and Fierce Pharma itself, rather than to advertising spend.²⁰ Because DTC advertising is a U.S.-specific phenomenon, this finding does not generalize to markets where branded prescription-drug promotion is prohibited.¹

The lesson generalizes well beyond GLP-1s: in any indication where trial data is the deciding factor, publishing — through the normal MLR and submission process — primary results in citable, crawlable form appears to matter more than advertising weight, and it works across multiple platforms simultaneously rather than requiring separate, platform-specific content.

Building a Compliant, Cross-Platform Pharma GEO Program

A GEO program built around one generic “get cited” checklist will underperform, and in pharma it can also create regulatory exposure. Six recommendations follow directly from the evidence above:

• 1. Route GEO content through the same MLR and FDA submission process as any other promotional material. Trial summaries, indication pages, and comparison content built to be citable by AI platforms are still promotional labeling under 21 CFR 314.81, with no separate exemption for content aimed at AI crawlers.³
• 2. Structure content so fair balance survives summarization. Because AI answer engines synthesize single paragraphs from source pages, risk information should be woven into the same sentences and sections as benefit claims, not isolated in a footer an AI summarizer is likely to drop.^2,7
• 3. Invest specifically in hospital and academic medical center relationships for Google AI Overviews visibility, since that platform’s health-citation mix runs opposite to ChatGPT’s, and institutionally authored content carries less direct fair-balance exposure for the sponsor.¹⁷
• 4. Do not attempt to seed or optimize Reddit citations for branded content. Treat community platforms as a pharmacovigilance and misinformation-monitoring surface instead, with any corrective communication following FDA’s narrow, non-promotional misinformation-response pathway.^4,5,6,8
• 5. Prioritize long-form, structurally precise content for Claude — clinical monographs, mechanism-of-action pages, technical whitepapers — consistent with its documented preference for depth over breadth, and route unsolicited off-label questions encountered anywhere to Medical Information rather than answering them in place.^4,11
• 6. Apply the GEO tactics from KDD ’24 as a compliance-safe baseline across all platforms. Controlled testing found that adding verifiable statistics, named-expert quotations, and explicit source citations each independently improved content’s visibility in generative-engine answers by roughly 30–40%, while keyword-stuffing did not transfer to this context.²³ Separate academic research on citation quality signals found structured, well-organized content with clear metadata is consistently associated with higher citation likelihood across engines.²¹ These tactics apply equally well to MLR-approved trial data and are structurally compliant, unlike community-seeding tactics.

Conclusion

There is no single AI citation algorithm to optimize for in pharma, and no single compliance shortcut either. ChatGPT and Perplexity lean toward government and primary-source content; Google AI Overviews leans toward hospital systems and, increasingly, video; Claude leans toward long-form technical depth; and Reddit sits underneath all of them at a scale most pharma marketers still underestimate, without being a channel this article recommends trying to influence directly.

The tactics that work broadly and hold up under FDA scrutiny — publishing MLR-approved primary trial data in citable form, structuring content so risk information survives summarization, and routing every GEO asset through the same regulatory process as any other promotional material — are the right starting point for any pharma GEO program. Everything after that should be platform-specific, evidence-based, compliance-reviewed, and revisited often.

References

1. Menkes D, Mintzes B, Lexchin J. Most high-income countries ban direct advertising of prescription drugs – why does NZ still allow it? The Conversation. Accessed July 5, 2026. https://theconversation.com/most-high-income-countries-ban-direct-advertising-of-prescription-drugs-why-does-nz-still-allow-it-231688
2. Code of Federal Regulations. 21 CFR Part 202 — Prescription Drug Advertising. eCFR. Accessed July 5, 2026. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-C/part-202
3. Code of Federal Regulations. 21 CFR § 314.81 — Other Postmarketing Reports. eCFR. Accessed July 5, 2026. https://www.ecfr.gov/current/title-21/chapter-I/subchapter-D/part-314
4. U.S. Food and Drug Administration. For Industry: Using Social Media. Center for Drug Evaluation and Research. Accessed July 5, 2026. https://www.fda.gov/about-fda/center-drug-evaluation-and-research-cder/industry-using-social-media
5. U.S. Food and Drug Administration. Addressing Misinformation About Medical Devices and Prescription Drugs: Questions and Answers (Revised Draft Guidance for Industry). Published July 8, 2024. Accessed July 5, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/addressing-misinformation-about-medical-devices-and-prescription-drugs-questions-and-answers
6. U.S. Food and Drug Administration. Postmarketing Adverse Event Reporting Compliance Program. Accessed July 5, 2026. https://www.fda.gov/drugs/surveillance/postmarketing-adverse-event-reporting-compliance-program
7. U.S. Food and Drug Administration, Office of Prescription Drug Promotion. Warning Letter: Sprout Pharmaceuticals, Inc. (709942). Issued May 29, 2025. Accessed July 5, 2026. https://www.fda.gov/inspections-compliance-enforcement-and-criminal-investigations/warning-letters/sprout-pharmaceuticals-inc-709942-05292025
8. Pharmaceutical Commerce. FDA’s Social Media Enforcement: Emerging Rules of Engagement. Published May 5, 2026. Accessed July 5, 2026. https://www.pharmaceuticalcommerce.com/view/fdas-social-media-enforcement-emerging-rules-of-engagement
9. 9. Search Engine Journal. The Guardian: Google AI Overviews Gave Misleading Health Advice. Published January 6, 2026. Accessed July 5, 2026. https://www.searchenginejournal.com/the-guardian-google-ai-overviews-gave-misleading-health-advice/564476/
10. OpenAI. ChatGPT Search. OpenAI Help Center. Accessed July 5, 2026. https://help.openai.com/en/articles/9237897-chatgpt-search
11. Anthropic. Advancing Claude in Healthcare and the Life Sciences. Accessed July 5, 2026. https://www.anthropic.com/news/healthcare-life-sciences
12. Profound. AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information. tryprofound.com. Accessed July 5, 2026. https://www.tryprofound.com/blog/ai-platform-citation-patterns
13. Ahrefs. 76% of AI Overview Citations Pull From the Top 10. Published July 21, 2025. Accessed July 5, 2026. https://ahrefs.com/blog/search-rankings-ai-citations/
14. Ahrefs. Update: 38% of AI Overview Citations Pull From The Top 10. Published March 2, 2026. Accessed July 5, 2026. https://ahrefs.com/blog/ai-overview-citations-top-10/
15. Ahrefs. New Study: AI Assistants Prefer to Cite “Fresher” Content (17 Million Citations Analyzed). Published April 27, 2026. Accessed July 5, 2026. https://ahrefs.com/blog/do-ai-assistants-prefer-to-cite-fresh-content/
16. Search Engine Journal. Google AI Mode & AI Overviews Cite Different URLs, Per Ahrefs Report. Published December 16, 2025. Accessed July 5, 2026. https://www.searchenginejournal.com/google-ai-mode-ai-overviews-cite-different-urls-per-ahrefs-report/563364/
17. BrightEdge. Healthcare AI Citations: How ChatGPT and Google Define Trust Differently. BrightEdge Research. Accessed July 5, 2026. https://www.brightedge.com/resources/weekly-ai-search-insights/healthcare-ai-citations-chatgpt-vs-google-trust
18. Search Engine Land. What Gets Cited Most in Health, Finance, and YMYL in AI Overviews: A Sector-by-Sector Analysis. Accessed July 5, 2026. https://searchengineland.com/guide/ai-overviews-ymyl
19. Search Engine Land. Google AI Overviews Cite YouTube Most Often for Health Topics: Study. Published January 16, 2026. Accessed July 5, 2026. https://searchengineland.com/google-ai-overviews-cite-youtube-health-topics-467628
20. 20. Fierce Pharma. Eli Lilly, Novo Nordisk Top AI Citation Share as New Report Questions DTC Spend Culture. Accessed July 5, 2026. https://www.fiercepharma.com/marketing/eli-lilly-novo-nordisk-top-ai-citation-share-new-report-questions-dtc-spend-culture
21. 21. Kumar A, et al. AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO16 Framework. arXiv:2509.10762. Preprint posted September 13, 2025. https://arxiv.org/abs/2509.10762
22. Venkit PN, Laban P, Zhou Y, Mao Y, Wu CS. Search Engines in an AI Era: The False Promise of Factual and Verifiable Source-Cited Responses. arXiv:2410.22349. Preprint posted October 2024. https://arxiv.org/pdf/2410.22349
23. Aggarwal P, Murahari V, Rajpurohit T, Kalyan A, Narasimhan K, Deshpande A. GEO: Generative Engine Optimization. In: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24). Association for Computing Machinery; 2024. doi:10.1145/3637528.3671900

Olivier Gryson, PharmD, MSc
25 years of experience in digital marketing in the pharmaceutical industry
Special focus on AI Search in Pharma Marketing