8 key GEO findings every pharma brand needs.

Eight structural patterns emerge from the May 2026 PharmaGEO public index — a measurement of Answer Rate (AR) and Share of Voice (SOV) across four therapeutic areas, three AI engines, and three languages. These are not vendor-internal observations: every data point below traces to publicly observable brand scores. Taken together, they constitute the empirical case for why single-engine, single-language GEO measurement is insufficient in 2026.

Finding 1: Engine divergence is structural — 33-point gaps within a single TA

The index's most striking finding is the magnitude of cross-engine divergence on the same brand in the same therapeutic area. Adbry (lebrikizumab) in the atopic dermatitis TA holds an Answer Rate of 41.4% on OpenAI — a top-4 position. On Perplexity, the same brand in the same query week scores 8.2% — a rank-8 niche mention. The gap is 33.2 percentage points.

This is not a measurement artifact. It reflects the genuinely different retrieval architectures, training corpus emphases, and citation-weighting logic of the two platforms. OpenAI's blend of training data and retrieval gives more weight to clinical literature that has been indexed over time. Perplexity's live retrieval gives more weight to current web content, including the specialist hub and forum ecosystem. The same brand has a different content footprint in each of those source pools.

The strategic implication is direct: there is no single "AI visibility" for a pharma brand. There are at minimum six distinct visibility figures — one per engine-language combination. A brand that looks healthy on its primary measurement engine may be structurally absent on the engine its target HCPs actually use. Any GEO strategy that does not monitor across at least three engines is working from an incomplete picture.

Finding 2: Language reorders the leaderboard — Adtralza gains 35pp from English to French

In the atopic dermatitis TA, Adtralza (tralokinumab) holds an English-language AR of 13.8% on OpenAI (rank 9). Switch the query language to French and the same brand's AR moves to 48.8% (rank 4) — a 35-point gain. In Spanish: 51.9% (rank 4) — a 38.1-point gain.

The mechanism is geographic. Tralokinumab received earlier approval and broader initial clinical coverage in EU markets than in the US. The non-English internet — the source pool Perplexity and OpenAI both draw on for non-English queries — reflects that European content asymmetry. The LLM is not making an error. It is accurately reflecting the source pool in each language.

This pattern is not unique to Adtralza. The French top-10 in atopic dermatitis includes Ebglyss — lebrikizumab under its European brand name (Almirall/Lilly EU) — at rank 8 with an AR of 20.0%. Ebglyss does not appear in the English top-10 at all. The same molecule, under a different brand name, in a different geography, has a different AI identity. Brand teams with global portfolios that include EU-first approvals or dual-named molecules need language-level visibility tracking as a standard capability.

Finding 3: Source stacks differ by TA — NCCN monopoly in lung cancer vs FDA labels in obesity

A common assumption is that the same content strategy — publish in NEJM, earn society guideline mentions, build PMC presence — will work across TAs. The May 2026 index shows this is false. The citation source mix is fundamentally different by TA, and the winning strategy in one category does not transfer to another.

For an oncology brand, the path to AI visibility runs through NCCN guideline inclusion above everything else. NCCN pages alone account for approximately 85% of citation uses in lung cancer answers. For a metabolic brand launching in obesity, the FDA label is structurally the dominant citation before the brand team publishes a single web page — because engines treat regulatory labels as the ground-truth safety and indication source. The GEO task in those two categories is fundamentally different, and conflating them produces a wasted content strategy.

Finding 4: Gemini concentrates SOV at the top

Across TAs, Gemini exhibits a consistent pattern: it gives materially larger Share of Voice to category leaders than OpenAI does on the same query set. In atopic dermatitis, Dupixent holds 16.3% SOV on OpenAI and 23.5% on Gemini — a 7.2-point premium. In obesity, Wegovy holds 22.2% SOV on OpenAI and 34.3% on Gemini — a 12.1-point premium.

The hypothesis is retrieval architecture: Gemini surfaces fewer distinct products per response than OpenAI, amplifying a winner-take-most dynamic. An engine that returns three brands per answer gives the #1 brand a larger proportional slice than an engine that returns eight. For category leaders, Gemini is a favorable engine. For mid-tier brands trying to earn share in a competitive TA, Gemini is the hardest surface to displace the leader on, and brand strategy should account for that asymmetry.

Finding 5: Perplexity surfaces a longer tail

The inverse of Finding 4: Perplexity consistently surfaces more distinct brands per query set than OpenAI. In lung cancer, Perplexity shows 29 distinct products with non-zero SOV; the OpenAI public table caps at 10 visible products. In psoriasis, Perplexity surfaces 19 products; OpenAI's visible set is 10.

In atopic dermatitis, Perplexity's top-10 includes Protopic (AR 29.9%, rank 4), Elidel (AR 20.6%, rank 6), and Eucrisa (AR 13.4%, rank 7) — drugs from the early 2000s that OpenAI's top-10 omits in favor of newer JAK inhibitors. These older brands have decades of peer-reviewed literature and deep guideline mentions that Perplexity's retrieval architecture surfaces reliably. Perplexity is structurally the more "generic-aware" engine. For brands in categories with long legacy tails, Perplexity competitive analysis is not optional — the brands Perplexity names are not the same brands OpenAI names.

Finding 6: AR is not SOV — a 4x compression even for category leaders

A common measurement confusion in early GEO programs is treating Answer Rate (AR) as equivalent to Share of Voice (SOV). The index shows they are not the same metric, and conflating them leads to strategic overconfidence.

In atopic dermatitis, Dupixent holds an AR of 65.5% on OpenAI — it appears in nearly two-thirds of all relevant prompts. Its SOV on the same engine is 16.3%. The AR-to-SOV ratio is approximately 4:1. In obesity, Wegovy holds an AR of 93.2% — nearly universal mention. Its SOV is 22.2%. Again, roughly 4:1 compression.

The mechanism is straightforward: every AI answer that names one brand in a category also names competitors. A brand that appears in 90% of answers is still competing for surface area in each of those answers with two to eight other named brands. A brand team that optimizes for AR and declares success because its brand "always appears" may be missing the fact that its SOV is one-quarter of what the AR figure implies. Measuring both metrics separately, and tracking the ratio over time, is the minimum viable reporting standard.

Finding 7: Older drugs retain visibility through literature depth

Finding 5 hints at this, but it is worth stating as a standalone structural pattern: literature depth is a retrieval moat that newer brands cannot overcome quickly. In the atopic dermatitis TA, Protopic (tacrolimus) — approved in 2000 — holds a Perplexity AR of 29.9% (rank 4) despite being a generic-era calcineurin inhibitor in a category now dominated by biologics. Elidel (pimecrolimus, approved 2001) holds AR of 20.6% (rank 6). Eucrisa (crisaborole, 2016) holds 13.4% (rank 7).

These brands have one advantage newer entrants do not: two to three decades of peer-reviewed publications, guideline mentions, real-world evidence, and systematic review citations. Each new meta-analysis that includes them extends their citation footprint without any active effort from a brand team. For newer brands without five or more years of literature, this represents a structural disadvantage — a "novelty penalty" that takes years to close. The strategic implication for recent launches: prioritize PMC-indexed publication volume and guideline submission early, not once commercial launch is consolidated.

Finding 8: AI is already a pharmacovigilance surface — FDA labels lead obesity citations and off-label leakage is measurable

The final finding is the one most brand teams are not ready for. In the obesity TA, Perplexity's top-cited sources are the FDA label for Zepbound (36 citation uses, rank 1) and the FDA label for Wegovy (32 citation uses, rank 2). These labels contain boxed warnings, full contraindication lists, and REMS information where applicable. The AI answer layer is not delivering marketing copy. It is delivering safety-and-indication information directly from regulatory documents, framed by the engine.

More acutely: in obesity prompts, the May 2026 index records measurable Answer Rate for Ozempic and Mounjaro — both approved for Type 2 diabetes, not obesity — surfacing in AI answers to obesity queries. Perplexity's SOV chart for the obesity TA includes Ozempic at approximately 6% and Mounjaro at approximately 3%. These are not promotional mentions. They are the engine connecting same-molecule branded products across indication lines — a systemic AI behavior, not a brand-specific error. But each such mention constitutes an off-label indication surface reaching prescribers and patients outside any MLR-reviewed promotional channel.

Brand teams that have assigned GEO to the digital marketing function and kept pharmacovigilance and MLR separate have the wrong organizational structure for what the AI answer layer is actually doing. Monitoring AI answers for off-label brand surfacing is a regulatory and safety function, not a media one.

Want a real audit on your brand? Request a sample report or get the full PharmaGEO Playbook.