How to Use AI to Write Pharma Content That AI Search Systems Are More Likely to Cite

Should pharma content teams use AI to write content that gets cited by AI search systems? The question is commercially urgent and the answer, based on the current evidence, is: yes, under specific conditions — but those conditions are more demanding than most optimistic accounts acknowledge.

The most influential data on this question comes from a 16-month experiment by SE Ranking, published in March 2026. That study is frequently cited to support the case for AI-assisted content. But it is also frequently misread. This article sets out what the data actually shows, corrects the most common misreadings, and draws out the practical implications for pharma content teams operating under Google’s quality framework and EU AI Act obligations.

What the SE Ranking experiment actually measured

SE Ranking ran two distinct experiments, reported together in a single publication1. The two experiments must be read separately, because they tested fundamentally different conditions and produced opposite results.

Experiment A: 2,000 articles on 20 new domains

Between June and September 2024, SE Ranking published 2,000 fully AI-generated articles across 20 newly purchased domains, each targeting a different niche. The domains had no prior authority, no backlinks, and no brand recognition. No editorial input was applied after AI generation.1

Within the first 36 days, approximately 71% of pages were indexed by Google, a result SE Ranking described as notable for zero-authority domains1. Pages ranking in the top 100 reached a peak of 28% in the first month. Then, around 3 February 2025, that figure collapsed to 3%. By the six-month mark, cumulative impressions across all 20 sites totalled approximately 706,000 and clicks 1,062 — roughly one click for every two articles published2.

An important nuance: the article did not end there. Following the August 2025 Google spam update, pages ranking in the top 100 recovered partially, rising from 3% back to 20%1. After 16 months, the experiment had generated a cumulative total of approximately 1,092,000 impressions and 1,381 clicks across all 2,000 articles2. This is not “no recovery” — it is a very modest partial recovery. The honest characterization is that the content remained structurally weak, not that it was permanently penalized in all cases.

Experiment B: Six AI-assisted articles on the SE Ranking blog

In parallel, SE Ranking published six AI-assisted articles on its own established blog, with human editorial review, on an authoritative domain with an existing backlink profile. Between June 2024 and July 2025, these six articles received approximately 555,000 impressions and 2,300 clicks; three ranked in the organic top 10, and five triggered Google AI Overviews, with four cited as sources within those Overviews1.

Critical note on a common misreading: Some summaries of the SE Ranking study present the 555,000 impressions figure as the result of the six successful blog posts alone, implying a high per-article return. This reading is misleading. The 555,000 figure in the study’s summary encompasses data across both experiments combined. The six established-domain articles and the 2,000 new-domain articles are reported in aggregate in some sections of the original publication. Readers should treat per-article comparisons between the two experiments with care, and consult the primary source directly.

What the experiment does and does not prove

The variable that differentiated outcomes was not AI involvement. It was the combination of domain authority, editorial quality, and human oversight1,3. The experiment tested all three simultaneously on the new-domain sites and found the combined absence of these factors to be determinative. It did not isolate any one variable. The correct inference is: AI-assisted content on established domains, with substantive human editorial input, can perform well in organic and AI search. Unedited AI content on new domains, at scale, is likely to fail — particularly after Google’s algorithmic review cycles run.

What Google’s policy framework actually penalises

Google’s Search Central documentation is explicit: “Using generative AI tools or other similar tools to generate many pages without adding value for users may violate Google’s spam policy on scaled content abuse.”4 The operative criterion is adding value, not the production method. Google’s spam policies confirm that scaled content abuse applies “no matter how it’s created”5 — a human-written page that is thin, templated, and published in volume faces the same penalty as an AI-generated equivalent.

Google’s January 2025 Search Quality Rater Guidelines update reinforced this by directing human raters to assign the lowest quality rating to content that is AI-generated with little originality or added value6. Rater assessments inform long-run algorithm development. This does not mean AI content is categorically penalised; it means AI-generated content with no substantive human layer is treated as low-quality content — which it typically is.

The March 2026 Google core update specifically targeted scaled content abuse, and sites publishing large volumes of AI-generated pages without editorial oversight reported organic traffic drops of 50–80%7. This is consistent with the SE Ranking new-domain experiment results and reinforces the policy direction.

EU AI Act disclosure obligations: What Article 50 requires and what remains unsettled

Under Article 50(4) of Regulation (EU) 2024/1689 (the EU AI Act)8, applicable from 2 August 2026, deployers who publish AI-generated text with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated (Art. 50(4); Art. 113). The obligation does not apply where the content has undergone substantive human review or editorial control and where a natural or legal person holds editorial responsibility for the publication (Art. 50(4)).

Pharma disease education pages, therapeutic information, and HCP-facing clinical material fall within the scope of “matters of public interest” under this provision. For pharma content teams, this means either a clear AI disclosure is required, or the human review exemption must be genuinely satisfied — not as a formality, but through documented editorial responsibility by a qualified person.

The human review exemption in Article 50 is real: disclosure is not required where content has been subject to substantive human review and genuine editorial responsibility. However, the precise boundaries of this exemption are still under active regulatory development. The European Commission’s Code of Practice on Article 50 compliance was in its second draft as of March 2026, with a final version expected by June 20269. Additionally, the Digital Omnibus process introduced the possibility of a transitional extension to 2 December 2026 for AI systems placed on the market before the August deadline10. Pharma content teams should monitor these developments and not treat the legal landscape as fully settled.

What is settled: a tick-box review process will not satisfy the exemption. The Commission’s draft guidelines indicate that disclosure buried in terms and conditions or applied via vague labels will not meet the threshold. A practical compliance model for pharma content should include:

  • A named medical reviewer with documented credentials
  • Primary source verification against original peer-reviewed publications
  • MLR clearance with documented rationale
  • An explicit AI-assistance disclosure in the article footer (even where the human review exemption may apply, transparency is a trust signal)
  • Ongoing monitoring of the final Article 50 Code of Practice and Commission guidelines

What AI does well, and what it cannot do

AI drafting accelerates the structural and semantic work that makes content citable: definition-first framing, query alignment, consistent attribution formatting, and clear section structure. For pharma teams managing content across multiple therapeutic areas, this efficiency is materially significant.

What AI cannot do is verify. It cannot confirm that a clinical statistic traces to the correct primary source. It cannot check that a treatment guideline reflects the most current version. It cannot ensure that a mechanism-of-action description is consistent with the current approved label. These are not temporary limitations. They reflect the fact that general-purpose language models do not have access to specific, current, proprietary, or regulated information.

The human workload in an AI-assisted pharma content workflow is therefore not lighter than in traditional production. It is concentrated at the verification stage rather than the drafting stage. Specifically:

  • Every clinical statistic must be traced to its original peer-reviewed source, not to a secondary article that cited it
  • Treatment guidelines must be verified against the most recently published version from the relevant professional body
  • Mechanism-of-action descriptions must be checked against the current approved prescribing information
  • All sources should be explicitly linked within the published content
  • MLR review by a qualified medical professional, with documented sign-off, remains required regardless of drafting method

The pharma competitive position in AI search: a plausible hypothesis, not a proven advantage

A reasonable hypothesis, grounded in how large language models handle niche queries, is that pharma organisations are well-positioned to occupy AI citation slots in specialist therapeutic areas. LLMs are trained on large general-purpose datasets; when a patient or HCP asks a specific question about a rare disease mechanism or a second-line treatment decision, the pool of high-quality, well-sourced, citable content is relatively thin. A verified, expert-attributed page addressing a niche clinical question with appropriate sourcing may face limited competition for the citation position.

This remains a hypothesis. No published study has yet quantified citation rates for pharma content in AI search systems by domain type, therapeutic area, or content format. The inference is structurally plausible, but pharma teams should treat it as a working hypothesis to test — not a demonstrated strategic advantage. Systematic tracking of AI citation across ChatGPT, Perplexity, Google AI Overviews, and Microsoft Copilot, using a library of clinically relevant queries, is the appropriate way to generate evidence.

Conclusion

The SE Ranking experiment does not prove that AI content fails. It proves that unedited AI content on zero-authority domains, published at scale with no editorial investment, fails to sustain search visibility after initial indexation. This is not surprising, and it does not contradict Google’s stated position that quality — not production method — is the relevant criterion.

The correct lesson for pharma is: use AI as a drafting tool, apply rigorous primary-source verification to every claim, attribute content to named qualified reviewers, comply with Article 50 disclosure obligations as they are finalised, and publish on domains that have earned authority. The workload this requires is different from traditional content production, not lighter.

The organisations that will be well-positioned in AI-mediated healthcare search are those that use AI honestly, verify rigorously, disclose transparently, and build on established domain authority. The evidence currently available (as well as my own personal experience) supports that conclusion. It does not yet support stronger claims about competitive advantage, citation rates, or the degree of recovery possible from algorithmic penalties.

References

  1. Khromova, Y. (2026, March 19). How AI-generated content performs in search: Results from an experiment by SE Ranking. SE Ranking Blog. https://seranking.com/blog/ai-content-experiment/ Last accessed 16/06/2026
  2. Schwartz, B. (2026, March). How AI-generated content performs in Google Search: A 16-month experiment. Search Engine Land. https://searchengineland.com/ai-generated-content-google-search-experiment-472234. (Third-party analysis of SE Ranking experiment data, including the Sword and the Script summary by Babiak, B., 2026, April.) Last accessed 16/06/2026
  3. ALM Corp. (2026, March 24). AI-generated content and Google Search: What a 16-month experiment shows about rankings, indexing, and long-term SEO. https://almcorp.com/blog/ai-generated-content-google-search-rankings/ Last accessed 16/06/2026
  4. Google Search Central. (2025). Google Search’s guidance on generative AI content. Google for Developers. https://developers.google.com/search/docs/fundamentals/using-gen-ai-content Last accessed 16/06/2026
  5. Google Search Central. (2025). Spam policies for Google web Search. Google for Developers. https://developers.google.com/search/docs/essentials/spam-policies Last accessed 16/06/2026
  6. Search Atlas. (2026). Does Google penalize AI content? (Citing Google’s January 2025 Search Quality Rater Guidelines update.) https://searchatlas.com/blog/does-google-penalize-ai-content/. Last accessed 16/06/2026
  7. Digital Applied. (2026, March). Scaled content abuse: Google’s March 2026 AI page crackdown guide. https://www.digitalapplied.com/blog/scaled-content-abuse-google-march-update-ai-pages-decimated Last accessed 16/06/2026
  8. European Parliament & Council of the European Union. (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union, L 2024/1689. https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689 Last accessed 16/06/2026
  9. Jones Day. (2026, January 27). European Commission publishes draft Code of Practice on AI labelling and transparency. https://www.jonesday.com/en/insights/2026/01/european-commission-publishes-draft-code-of-practice-on-ai-labelling-and-transparency Last accessed 16/06/2026
  10. Across Legal. (2026). AI transparency in the EU: Code of Practice and Article 50 explained. (Notes on Digital Omnibus provisional agreement and potential deadline extension.) https://acrosslegal.com/en/transparencia-ia-reglamento/ Last accessed 16/06/2026

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

Frequently Asked Questions

No. Google penalises content that lacks originality, depth, or genuine value to users — regardless of how it was produced. Its scaled content abuse policy applies equally to AI-generated and human-written pages published in bulk without editorial investment. The production method is not a ranking signal; content quality, expertise, and usefulness are. (Regulation (EU) 2024/1689, Art. 50(4); Google Search Central, Spam Policies.)

It proved that unedited AI content on zero-authority domains, published at scale with no human editorial input, fails to sustain search visibility after initial indexation — not that AI content fails categorically. The same study found that AI-assisted articles published on SE Ranking’s established blog, with human editorial review, ranked in the organic top 10 and were cited in Google AI Overviews. The differentiating variable was domain authority and editorial quality, not AI involvement.

Partially. The article collapsed from 28% to 3% of pages ranking in the top 100 around February 2025. Following the August 2025 Google spam update, that figure recovered to approximately 20%. After 16 months, the sites had generated a cumulative total of roughly 1,092,000 impressions and 1,381 clicks across 2,000 articles — a very modest return, but not zero. “No recovery” is an overstatement of the data.

There is no regulatory prohibition on using AI to draft pharma content. All pharma content — regardless of drafting method — must pass MLR review, comply with applicable codes (including EFPIA and IFPMA guidelines), and accurately reflect the approved scientific evidence base. The drafting method is not itself a compliance variable; the output quality and the rigour of the review process are.

From 2 August 2026, yes — in most cases. Under Article 50(4) of Regulation (EU) 2024/1689, deployers who publish AI-generated text on matters of public interest must disclose that the text has been artificially generated. An exemption applies where the content has undergone substantive human review or editorial control and a named person holds editorial responsibility for the publication. Pharma disease education and HCP-facing clinical content fall within the scope of “matters of public interest.” Outside the EU, disclosure requirements vary by jurisdiction.

The exemption requires that the AI-generated content has undergone a genuine process of human review or editorial control, and that a natural or legal person holds editorial responsibility for the publication. A tick-box approval process will not satisfy it. In pharma terms, this means documented MLR review by a qualified medical professional with named sign-off — not a superficial read-through. The precise boundaries of the exemption are still being finalised through the European Commission’s Code of Practice on Article 50 compliance, with a final version expected in June 2026.

Every claim that could affect clinical understanding or patient safety must be verified by a qualified human reviewer. In practice this means: each clinical statistic traced to its original peer-reviewed source; treatment guidelines checked against the most recently published version from the relevant professional body; mechanism-of-action descriptions verified against the current approved prescribing information; all sources explicitly linked within the published content; and MLR sign-off with documented rationale.

It is a plausible hypothesis, not yet a demonstrated advantage. The reasoning is structurally sound: LLMs cite the most authoritative available answer to a specific query, and in niche therapeutic areas the volume of high-quality, well-sourced content is relatively thin. A verified, expert-attributed page addressing a specific clinical question may face limited competition for the citation position. However, no published study has yet quantified AI citation rates for pharma content by domain type or therapeutic area. Pharma teams should treat this as a hypothesis to test through systematic citation tracking, not a proven strategic advantage.

Traditional SEO metrics — rankings, organic traffic, click-through rate — are insufficient when AI Overviews and generative engine responses resolve queries without a click. Teams should track AI citation directly: systematically querying target prompts in ChatGPT, Perplexity, Google AI Overviews, and Microsoft Copilot to establish whether their content is cited and in what context. This requires a query library aligned to patient and HCP search intent, run regularly, with citation frequency documented over time.

In practice, the distinction Google’s systems apply is not between “AI-generated” and “AI-assisted” as labelled categories, but between content that demonstrates genuine expertise, originality, and usefulness versus content that does not. The SE Ranking experiment confirms this: AI-assisted content with substantive human editorial investment performed well; unedited AI content on new domains did not. The label matters less than the quality and depth of the human layer applied on top of the AI draft.

Follow the conversation on LinkedIn

I regularly share reflections on pharma marketing, search behavior, and the impact of AI on healthcare communication.

Follow me on LinkedIn

This article was written with the assistance of generative AI technology and reviewed for accuracy.