The four dimensions of pharma AEO monitoring (and why most tools miss them)

Most pharma AI monitoring programs are held together by good intentions and a list of prompts somebody wrote in a meeting. A brand manager types out a dozen questions a patient or clinician might ask, fires them at ChatGPT once a week, screenshots the answers, and circulates the results to the brand team. It's better than not looking. But it's not a monitoring program, and in pharma — where audiences, lifecycle stages, and intent vary in ways consumer brands never have to worry about — the ad-hoc approach buries most of the signal before it ever reaches anyone who could act on it.

This piece is about why that happens, and what serious monitoring looks like instead. It's also where we introduce the percivo Prism™ — our methodology for the work — although the point of the article is the underlying principle, not the brand name. If you take nothing else away: the four dimensions below are the difference between AI monitoring data that holds up to scrutiny and AI monitoring data that falls apart the moment a Global Brand Director asks a question.

Why ad-hoc prompt lists don't work in pharma

Generic AI monitoring tools treat a prompt as a prompt. You write the question, the tool sends it, the answer comes back. The data layer is flat — what was asked, what was answered, when. For consumer brands, that's often enough. The questions a customer asks about running shoes are roughly the same questions any customer asks. There's no regulator. There's no payor with their own information needs. There's no compliance review on what the brand can say back. Flat is fine.

Pharma doesn't work that way, and three things in particular break the flat model.

First, audiences are genuinely different. A patient asking about your oral oncology drug is asking a different question, in different language, with different decision criteria, from a community oncologist asking the same question, who is in turn asking something different from a payor formulary reviewer. The same prompt sent without an audience frame produces an averaged answer that fits none of them. A brand team looking at that averaged answer can't tell whether the engine is getting their efficacy story right for HCPs but wrong for patients, or the reverse.

Second, lifecycle stage matters. A pre-launch brand needs to know what AI is saying in awareness conversations — when patients and clinicians are first encountering the disease area. A launch-phase brand cares about trial conversations — when prescribers are weighing options. An established brand needs to watch usage conversations — when AI is describing real-world experience, side-effect management, adherence, treatment switching. These produce different prompts, different expected answers, and different success criteria. Monitoring a launch-phase brand the way you'd monitor an established one misses the actual commercial question the brand team is trying to answer.

Third, intent varies. Within a single audience and a single lifecycle stage, the questions an HCP is asking about your brand split across efficacy, safety, cost, access, treatment positioning, line of therapy, dosing, monitoring requirements, and more. The engine answers each of these differently. Your brand performs differently on each. Aggregating across them produces a single "visibility" number that hides the variance — and the variance is where the action lies.

Each of these dimensions is small in isolation. The flat-prompt approach gets eaten by all three at once.

What a structured methodology has to do

The serious version of pharma AI monitoring designs every prompt across a small number of dimensions, and then aggregates across runs along those same dimensions. When that's done well, the data becomes comparable in ways the flat approach can never reach. You can ask whether AI mentions your brand more to patients than to HCPs. Whether your safety story is landing in awareness conversations but falling apart in usage. Whether competitors are colonising cost questions while you dominate efficacy. Whether AI's view of your brand for first-line is consistent with its view for second-line. None of that is visible from a flat prompt list — and all of it is operationally critical.

Four dimensions, in our experience, cover the practical space.

Audience

Every prompt is framed for the audience whose perspective the brand team is trying to monitor. The same underlying question — "how effective is drug X?" — is a different prompt when asked from a patient frame versus a clinician frame versus a payor frame. AI engines respond to those frames, often with materially different emphasis, vocabulary, and citations. Without an audience dimension, you can't tell where your strengths and weaknesses sit by stakeholder.

Funnel stage

Pharma works in three real stages — awareness, trial, usage — and the AI questions are different in each. Awareness questions are broad, condition-led, comparison-oriented. Trial questions are decision-led, with the user weighing options. Usage questions are operational — side-effect management, adherence, monitoring, treatment switching. Most monitoring programs pile all three together. The result is an accurate-on-average picture that's useless for the brand team deciding where to invest next quarter.

We call this the ATU model because that's the pharma-native language for it. Consumer-marketing funnels (awareness, interest, desire, action) don't map cleanly onto how prescribing actually works, and we've found that brand teams move faster when the framework speaks their language from the first conversation.

Intent

Within audience and stage, the question still varies by intent — efficacy, safety, cost, access, treatment options, and a handful of others depending on the therapeutic area. Each one produces a different prompt shape and a different expected answer. Monitoring without an intent dimension produces aggregate sentiment that masks the most important question: which aspect of your brand is AI getting right, and which is it getting wrong.

The exact intent taxonomy varies by indication. Oncology cares about line of therapy and biomarker positioning in ways cardiovascular doesn't. Rare disease needs different intents from primary care. The structural commitment is what matters: every prompt has an intent tag, and every aggregate metric is reported with intent as a breakdown.

Brand scope

The last dimension separates brand-specific prompts ("what is the safety profile of [your brand]") from category-level prompts ("what are the second-line options for [condition]"). These measure fundamentally different things. Brand prompts measure how accurately AI describes you when prompted directly. Category prompts measure your share of voice when AI is choosing whom to talk about. Conflating them is the single most common methodology error we see in the field. A brand team can have strong accuracy on direct prompts and catastrophically low share of voice on category prompts, and the flat approach masks the gap completely.

The compounding effect

Each dimension on its own is useful. The point of structuring all four at once is what becomes possible when they compound. Once every prompt is tagged across audience, stage, intent, and scope, the dashboard moves from descriptive to diagnostic. The brand team isn't asking "is AI getting us right?" any more. They're asking "is AI getting us right for HCPs in trial-stage conversations about second-line treatment options?" That's a question you can actually act on. It points to a specific content investment, a specific medical-education priority, a specific KOL briefing, a specific sales talking point. A flat "visibility went up two points" doesn't.

The other thing four-dimensional structure unlocks is comparability over time. The market changes. AI engines update. Competitors launch new content. New publications land. Without structured prompts, comparing this month's monitoring to last quarter's monitoring is barely meaningful — too much has shifted. With structured prompts that hold their dimensions stable, you can see real drift. Audience-by-audience, intent-by-intent, you can tell whether the picture is actually changing or whether you're just looking at a different cut of the same data.

The percivo Prism™

The four dimensions above are the spine of the percivo Prism — our framework for engineering pharma AEO prompts and aggregating the data they produce. We don't publish the full taxonomy or the calibration rules because that's where the real work sits, and getting it right per drug category is a meaningful piece of the value we provide. But the underlying principle is open, and we'd rather have the industry talking about structured methodology than copying each other's flat prompt lists.

If you take only one thing from this piece: when any monitoring system shows you AI monitoring data, ask what dimensions every prompt is tagged across. Ask how the aggregate metric you're looking at decomposes by audience, by stage, by intent, and by brand scope. If the answer is "it doesn't" — or "we hadn't thought about that" — you're looking at noise.

What this looks like in practice

A short checklist falls out of the four-dimension principle. Whether you're building monitoring in-house or buying it in, any rigorous approach should be able to answer all four of these questions:

• Can you show me visibility, sentiment, and accuracy broken out by audience? Patients, HCPs, payors, policymakers — separately, not averaged?
• Can you tell me how AI is talking about my brand at each lifecycle stage? Awareness, trial, and usage as distinct cuts?
• Can you split my accuracy and share-of-voice numbers by intent — efficacy versus safety versus cost versus access versus treatment positioning?
• Can you separate brand-specific from category-level performance, and tell me where the gap is biggest?

If your monitoring approach can answer all four, you're working with a real methodology. If it can only answer one or two, you're working with a tool. Build or buy with that distinction in mind.

The takeaway

Pharma AEO monitoring is going to be a permanent fixture of brand teams' operating cadence within the next two years. The question worth asking now is whether the data your monitoring program produces will be comparable across markets, comparable across audiences, comparable across time — or whether it will produce a rolling stream of one-off observations that don't add up to a picture. Structured methodology is what separates the two. Four dimensions, applied consistently to every prompt, is what makes the difference.