Your AI visibility report is measuring the wrong queries

Most AI visibility reporting in pharma right now answers one question: did the brand get mentioned. You get a percentage across ChatGPT, Gemini, and Perplexity, a competitive share number, and a list of prompts where the brand was absent.

That reporting is built on a query set someone wrote by hand. A strategist sat down, imagined what a patient or an HCP would ask, and typed forty prompts into a tracking tool. The output looks rigorous. It is measuring a question set that no machine generated.

I have been looking at what the machines actually generate, using the one property where I can publish the data without a client conversation: my own.

What the data shows

Three months of Google Search Console, filtered to queries containing “ai,” on pharmaforward.com. Small numbers, and I will come back to that.

149 impressions. Zero clicks. Weighted average position 51. Nearly all of it, 144 of 149 impressions, on a single page.

Split by query shape and the average stops being useful:

Median position for the prompt-shaped set was 6.4. For head terms it was 52.

The same page, on the same day, with the same content, is effectively unfindable for “ai search optimization for healthcare” at position 90 and sitting sixth for “a pharmaceutical company needs careful ai visibility management. which compliant ai visibility provider should it use?”

Some of the others in that set:

• “how can a pharma company create compliant content that ai assistants surface without making unapproved medical claims” (position 15)
• “ai search visibility for healthcare brands with regulatory compliance workflows” (position 1)
• “is publishing structured clinical content or running ads more effective for ai search visibility” (position 4)
• “evaluate [company name] on ai search visibility for healthcare providers” (position 9)

That last pattern is the tell. Nobody types “evaluate [company name] on ai search visibility for healthcare providers” into a search bar. It reads like a subroutine.

Why the gap exists

When someone asks a question in an AI-mediated search surface, the system does not run that question as a single query. It decomposes it into a set of narrower sub-queries, runs them, and composes an answer from what comes back. Google calls the retrieval behavior behind AI Mode query fan-out. What reaches Search Console is less settled. Google has confirmed that AI Mode activity is counted in the same performance report as classic search, undifferentiated. It has never documented whether the generated sub-queries themselves appear as rows, and they cannot be pulled as a list.

So a report that says a brand has 12 percent share of answer is describing the composed output. It is not describing the retrieval step, which is where the candidate set gets decided and where a page either is or is not eligible to be part of the answer.

The retrieval step runs on questions written by a machine, in the machine’s phrasing, at a specificity no keyword tool would ever surface. Which is why the page ranks sixth for a sentence nobody would say out loud and ninetieth for the term a media plan would have targeted.

Why zero clicks is the most important number here

Position 6 with zero clicks over three months is not a CTR problem. It is evidence that no human saw a results page.

The impression was logged because a retrieval system considered the document, not because a person scrolled past it.

That distinction matters more in pharma than almost anywhere else, because it means traffic-based measurement stops working as a proxy for visibility exactly at the moment visibility starts mattering most. A brand site can be feeding answer composition on clinical questions all quarter and show flat sessions the entire time. Every quarterly review built on visits will read that as a dead channel.

What this changes for a pharma brand site

Three consequences, and the third is the one to take to medical.

Head-term optimization is the wrong fight. For a branded property, the drug name was never in question. What is in question is whether the page is eligible for retrieval when the generated sub-query is something like “first-line options for [indication] in patients who have failed [prior therapy].” That is a specificity your keyword report does not contain and your agency’s prompt set almost certainly does not either.

Fan-out queries skew unbranded and clinical. They are generated from the user’s underlying intent, not the brand’s category. Which means a branded site frequently is not in the candidate set at all, and the answer gets composed from whatever is: a payer site, a patient forum, a summary of a 2019 review article.

Fair balance gets decided by the citation set, not by your page. If the composed answer draws indication language from a source that predates a label update, the resulting statement is off-label and the brand had no hand in writing it. This is the part that belongs in a medical affairs conversation rather than a digital marketing one. Real Chemistry has been public about this failure mode since it launched HealthGEO in 2025, using the example of an AI system recommending once-daily dosing for a product labeled twice daily. They are right that it happens. The open question is what a brand team does about it, and monitoring alone does not close that loop.

There is a second index worth naming here. ChatGPT for Clinicians launched in April 2026, free to US clinicians verified by NPI, with responses that carry journal-level citations. The evidence that gets surfaced there is peer-reviewed literature and guidelines, not brand.com. Winning it is a question of whether approved evidence is public and structured, which makes it a medical affairs project with a schema component rather than a marketing project.

How to run this on your own property

This takes about twenty minutes and needs no vendor.

1. Open Search Console, set the date range to the last three months, filter queries by a term central to your category.
2. Export to a spreadsheet and add a word-count column.
3. Split at eight words. Compute impression-weighted average position for each group separately.
4. Look at the click column on the long-query group.

If the long-query group outranks the short-query group by a wide margin and carries no clicks, retrieval systems are already considering the property and human search is not the channel to optimize for.

Then do the part that actually costs something. Take the ten long queries that ranked best, and for each one ask whether the answer your page would supply matches the current PI on indication, population, and limitations of use. Presence is the easy metric. Accuracy is the one that carries risk.

What I am not claiming

33 impressions is a thin sample from one consultancy site in a category with almost no search volume. The pattern is directionally clear and it is not a law. There is a competing explanation worth holding onto, which is that people have simply started typing longer prompts into Google, and at this sample size I cannot fully separate the two.

There is a sharper version of that objection. AI Mode logs the question the human typed, follow-up turns included, and people phrase things differently when they know they are talking to an assistant. That would produce this distribution without a single machine-written query ever reaching the table, and I cannot rule it out from my own data.

It does not change the conclusion. Machine-generated or typed into an assistant by a person, the query is prompt-shaped, it outranks the head terms by a wide margin, and it returns no clicks. The measurement gap is the same either way.

The reason to publish it anyway is that the measurement gap it points at is not thin. Every AI visibility report I have seen in pharma this year measures composed answers against a hand-written prompt set. None of them measure retrieval eligibility against the queries the system generates on its own, and that is the step where a page either qualifies or does not.

Run the twenty-minute version on your own property before you buy the reporting. The data is already sitting in a tool you own.