An ROI model for AI visibility in pharma.

Two facts now widely reported make the question worth modeling. More than 230 million people ask ChatGPT health questions every week, and 54% of HCPs report using generative AI to access scientific information. If a non-trivial share of treatment decisions is now informed by AI answers, that has revenue consequences for any brand whose visibility on those answers is uneven. The natural follow-up question is how big those consequences might be, and how they compare to the cost of doing something about them.

This article presents a simple framework for answering that, applied to a generic $300M specialty brand. The point is not the dollar values, which are entirely a function of the assumptions; the point is the structure. Every parameter is written out so the model can be re-run with different inputs, and several of the inputs are uncertain enough that we give a range rather than a point estimate.

1. The information channel that is now real

The reason a model like this is interesting in 2026 (and was not in 2023) is that AI has become a measurable share of how patients and clinicians look things up. On the patient side, OpenAI reports roughly 40 million daily users sending health-related prompts, and over 5% of all ChatGPT messages are healthcare-related. Pew Research found in April 2026 that 22% of US adults get health information from AI chatbots at least sometimes.

On the HCP side, Doximity's 2026 State of AI in Medicine survey of 3,151 US physicians shows clinical AI use rising from 47% to 63% in under a year, and the IQVIA/EPG survey above documents 54% of HCPs using generative AI in clinical contexts, rising to 75% among those born after 1990. Specialist clinical tools are growing on a similar trajectory: OpenEvidence went from 2.6 million monthly physician queries in 2024 to 18 million in December 2025.

The economics of how information is found are shifting in parallel. Gartner projected in 2024 that traditional search engine volume would drop 25% by 2026. Industry trackers report zero-click searches around 59% in the US and EU, and organic CTR collapsing by roughly 61% when an AI overview is present. Adobe's longitudinal analytics on US retail sites (a directional analogy, not a pharma measurement) show AI-referred visitors converting 38% worse than other channels in March 2025 and 42% better by March 2026, with 48% longer visits and 37% higher revenue per visit. Fewer touchpoints, each carrying more decision weight.

Taken together, this is enough to assume that some share of healthcare information moments has moved into the AI answer layer. It is not enough to know exactly how much. That uncertainty is the first thing the model has to handle.

2. Structural factors that may matter more in pharma

Every industry is exposed to AI-mediated search. A few structural factors plausibly make the exposure shaped differently for pharma:

Owned content is trusted less than third-party content. Cross-engine audit work on the PharmaGEO platform suggests LLMs assign manufacturer websites lower retrieval reliability than institutional sources: brand.com domains score around 53% and HCP-targeted brand sites around 68%, versus 80%+ for FDA's accessdata.fda.gov and 83% for NICE. That inverts the optimization target compared to traditional pharma digital.

MLR cycles are slow. Content that influences what an AI says about a product has to clear medical-legal-regulatory review. Two to four quarters from decision to compliant, localized content live is a reasonable working assumption for most large pharma organizations.

Training data is English- and US-skewed. When a French patient asks about treatments for myasthenia gravis, or a German clinician compares osteoporosis options, models often default to US-approved drugs and US guidelines. For European and global brands, language-localized AI visibility is closer to a different competitive surface than to a translation problem.

Absence does not produce silence. When authoritative brand-linked content is missing, models still produce an answer, drawn from whatever is available: older studies, patient forums, region-mismatched data, competitor-adjacent content. A 2026 audit published via BMJ rated 49.6% of chatbot health responses as problematic. The point is not that this is the default outcome, but that the failure mode is observable and worth pricing.

3. The model: at-risk envelope

The illustrative brand: $300M in annual revenue, of which 25% ($75M) comes from new-to-brand patients each year. We focus on the new-to-brand stream because it is the most sensitive to information at the moment of treatment choice. Established refill revenue is stickier and is excluded, which makes the model lower-bound by construction.

Three parameters drive the at-risk estimate:

1. AI-influenced decision share. The fraction of treatment-decision information moments influenced by AI answers. We model 12% in 2026, 20% in 2027, 30% in 2028, derived from the adoption data above with a discount for AI being one input among several.
2. Visibility-gap loss rate. Among AI-influenced moments, the share a poorly-represented brand loses. We model three scenarios: 10% (conservative), 25% (base), 40% (aggressive).
3. Persistence multiplier (×1.8). For chronic and semi-chronic therapy, a lost new-patient start represents a lost refill stream, not a single prescription. ×1.8 is a working assumption that should be tuned to the product's actual persistence curve.

Multiplied out:

A few properties of this envelope are worth flagging. The figure is invisible in standard dashboards because a prescription that was never written generates no signal: if the dynamic exists, it shows up as drift in new-patient share that is easily attributed to competitive pressure, access, or seasonality. The figure also grows over time even at a constant loss rate, because the AI-influenced share itself is growing. And the model only counts revenue: it does not price the separate scenario of inaccurate safety information being served at scale, which is a different (and harder) exposure category.

4. The model: program cost and return

What a GEO program does

Audit work across OpenAI, Perplexity and Gemini suggests a fairly stable set of high-impact actions:

1. Regulatory anchoring. Inline citations from brand digital properties to FDA/EMA labels and official documents. In audit settings this appears to lift LLM-assessed reliability scores by 15–25%.
2. An Evidence and Guidelines hub. A single MLR-reviewed page consolidating labels, guidelines (NICE, professional societies), and peer-reviewed evidence.
3. Structured data. Schema.org Drug, MedicalWebPage, FAQPage markup, so retrieval systems parse content as intended.
4. Third-party authority. Compliant enrichment of molecule and class Wikipedia pages, plus completeness on ClinicalTrials.gov and PubMed Central, which show up frequently in citation traces.
5. Localization. Native-language, market-specific content in priority markets. Language is consistently one of the largest sources of LLM output variation we observe.

Plus continuous measurement of how each major model represents the brand in branded and agnostic modes.

What it costs

For reference, $300K is roughly 2–3% of a typical mid-size brand marketing budget, and a small fraction of a single DTC campaign ($12M–$18M is a documented range).

What it returns

A program will not recover 100% of the at-risk value. Competitors are also acting, some answer patterns resist change, and many other factors influence outcomes. We model recovery rates of 30% (conservative), 50% (base), 65% (aggressive), and pair the lowest recovery rate with the most expensive program tier as a stress case:

The most informative number is the break-even. Under base-case assumptions, the program pays back if it protects roughly 4.3% of at-risk value. That threshold is low enough that the directional conclusion is reasonably robust to large parameter movement: the loss rate could be off by half, the AI-influenced share by a third, the recovery rate by 20 points, and the model still produces a positive net.

Two return categories sit outside the modeled figure. The first is offensive upside: the model counts only defended revenue, while categories where competitors have not invested (still many, especially in non-English markets) plausibly offer answer-share that exceeds market share. The second is risk and efficiency: earlier detection of misinformation, deflection of medical-information inquiries already correctly answered by AI, and a shared external reference for medical, digital and brand teams. Both are real; neither is quantified here, on purpose.

5. A theoretical note on time

One feature of this model is harder to capture in a table. AI visibility appears to be built from slow assets: guideline citations, registry completeness, peer-reviewed presence, structured content with accumulated trust signals. Models grounded on today's authority structure tend to reproduce it tomorrow, and observed answer patterns appear easier to maintain than to overturn.

If that holds, the temporal asymmetry differs from familiar paid media. Budget can typically buy back share of voice at any point. Time spent accumulating authority signals may not be substitutable by spend later. A brand starting 18 months later than a comparable competitor is not behind on budget; it is behind on signal density, and on the regulatory clock that mediates how fast new signal can be added (two to four quarters from decision to live, compliant, localized content).

This is a theoretical claim, not a measured one, and the strength of the effect is the second-largest source of uncertainty in the model after the visibility-gap loss rate. We flag it because, if it is even directionally right, it changes the cost of waiting more than it changes the cost of acting.

6. A pragmatic way to test the model

The most useful next step on a specific brand is to test the assumptions cheaply rather than commit to them expensively. A 90-day sequence:

1. Baseline (weeks 1–4). Measure how ChatGPT, Gemini, Claude and Perplexity (plus market-relevant models) represent the brand and the category today, in branded and agnostic modes, per market and per language. The baseline either confirms the model has bite for the brand, or it doesn't.
2. Triage (weeks 4–8). Rank observed gaps by revenue exposure: misrepresentations and safety inaccuracies first, then absence in agnostic disease-state answers, then sub-optimal sourcing.
3. Fix the foundations (weeks 6–12). Regulatory anchoring, the evidence hub, structured data. The owned-content playbook covers the operating detail.
4. Localize where it matters. Prioritize markets where revenue concentration overlaps with training-data bias.
5. Re-measure. Treat AI visibility as a position to be tracked, not a project to be closed.

The 90-day cycle is small enough to learn from and stop, if the model turns out to be wrong for a particular brand. That is the test the model is designed to support.

Methodology and assumptions

Known limitations. No published pharma-specific GEO ROI study exists to calibrate the loss and recovery rates. The Adobe conversion data is drawn from retail and e-commerce. The Gartner figure is a prediction, not a measurement. Zero-click statistics come from secondary industry aggregations. The break-even property is the main reason the directional conclusion is robust to imprecision in any single input. We expect every number here to be revised as better data becomes available.

If the framing is useful, the easiest way to pressure-test it on a specific brand is a small baseline audit. Request a sample report or browse the PharmaGEO Playbook.