Every approved module a pharma team ships has a cost attached to it. The medical reviewer's hours, the regulatory sign-off, the legal pass, the version history, the careful clause that survived three rounds of MLR. That cost buys a guarantee: the content is accurate, on-label, and defensible. What it does not buy is a guarantee that anyone, or any machine, will ever repeat it.
That second gap is the one that matters now. When a clinician asks ChatGPT what Varigel is indicated for, or asks Perplexity how it compares to the standard of care, or sees a Google AI Overview before any blue link, the answer is assembled from whatever the engine can retrieve and chooses to echo. Sometimes that is your approved content. Often it is not. The metric that captures the difference is AI Pickup: how much of your approved content the models actually echo back to the people asking. Pickup is the bridge between the MLR effort you already spent and the answer the HCP actually receives. When pickup is high, the review cycle paid off twice. When pickup is zero, the module is accurate, compliant, and invisible, which is to say the MLR effort was, from the audience's point of view, wasted.
What AI Pickup actually measures
Start with a precise definition, because "AI visibility" has become a catch-all that hides the thing that counts. AI Pickup measures the share of an answer that traces back to your approved content. Not whether your brand was mentioned. Not whether your domain was linked. Whether the substance the engine echoed, the indication phrasing, the efficacy framing, the safety language, the claim and its qualifier, matches a module your team approved.
That is a stricter test than a brand mention, and it is the right one for pharma. A brand can be mentioned in an answer that gets the indication subtly wrong. A domain can be cited as a source while the engine paraphrases a competitor's framing of your drug. Pickup asks the question the regulator and the medical team both care about: when the machine talks about our product, is it repeating what we approved, or something else? It splits cleanly into two halves. The surfacing half: approved content the engine retrieves and echoes. And the invisible half: approved content that exists, that cleared MLR, that is correct, and that the engine never reaches, so a third-party source wins the answer in its place.
The invisible half is the expensive one. It represents review hours converted into content that the audience, increasingly mediated by an answer engine, never sees. And the audience is moving. WebFX's analysis of more than 130,000 health-related queries found that AI Overviews now appear in 51 percent of healthcare searches, double the average across all industries. On the professional side, Doximity's 2026 State of AI in Medicine report found that 54 percent of surveyed physicians use AI in clinical practice, with literature search the single most common use case, rising to 35 percent of physicians surveyed in January 2026 from 22 percent in April 2025. The interface where your content gets repeated, or does not, is shifting from the ten blue links to a synthesized answer. Pickup measures how your approved content is faring in that shift.
Why pickup has to be measured per engine
The instinct is to ask "are we visible in AI" as if AI were one surface. It is not, and the data on this is now unambiguous. The five canonical engines, ChatGPT, Gemini, Perplexity, Google AI Overviews, and Claude, draw from different source pools, weight them differently, and echo content in structurally different ways. A module that surfaces strongly in one can be entirely invisible in another.
Profound's cross-platform citation analysis is the clearest illustration. ChatGPT leans on encyclopedic and established sources, with Wikipedia its single most cited source; Perplexity concentrates heavily on community discussion, with Reddit close to 47 percent of its top sources; and Google AI Overviews sits between the two with a more balanced distribution across professional and social platforms. The divergence is not a rounding error. An analysis of 680 million AI citations found that only 11 percent of the sources cited by ChatGPT are also cited by Perplexity. Two engines, asked the same question about your drug, are assembling their answers from almost entirely different material.
For pharma this has a hard operational consequence: a single aggregate "AI pickup" score is misleading. Your approved efficacy module might be echoed cleanly by ChatGPT because it pulled from a structured, authoritative page, while Perplexity reaches for a Reddit thread that frames the same drug through patient anecdote, and an AI Overview splits the difference. One number averages those into a fiction. The only honest measurement runs the same prioritized query set across each engine separately, then scores each answer for which half it falls into, surfacing or invisible, and which source carried it. As one practitioner guide to measuring AI share of voice puts it, the same piece of content can yield very different outcomes depending on the engine, because of structural differences in how each handles citation, personalization, and localization. Pickup is an engine-by-engine measurement or it is not a measurement at all.
Why approved content goes invisible, and what moves the needle
Approved content rarely loses the answer because it is wrong. It loses because the engine cannot retrieve it or cannot lift it. The clause that would make a perfect, citable answer, "indicated for adults with X who have not responded to Y," is rendered as an image inside a detail aid, locked in a slide in a CLM deck, or buried on page nine of a leave-behind. It is built to be presented by a representative, not parsed by a retriever. To a machine, content it cannot read is content that does not exist, and a plain-text forum post that says roughly the same thing wins the echo by default.
The good news is that the levers that move pickup are well documented and they happen to align with how pharma already builds content. The peer-reviewed GEO study from Princeton University and IIT Delhi tested content modifications across thousands of queries and found that the most effective methods, adding citations, quotations, and statistics, produced a relative improvement of 30 to 40 percent on source visibility, with the authors noting boosts of over 40 percent across various queries. Keyword stuffing and other traditional SEO tactics did nothing. The engines reward content that states a clear fact, attaches a credible source, and can be lifted as a self-contained passage.
That is precisely the shape of a well-built approved module: a clause-cited fact with its reference attached. A brand that adopted modular content for faster MLR review has already done most of the structuring work that pickup requires. The unlock is rendering those approved modules as clean, machine-legible content, sourced from the system of record so they stay in sync with what was cleared, rather than trapping them inside assets the retriever cannot reach. The content is already approved. Making it surface is a matter of form, not a new review.
A worked example: Varigel
Varigel is a fictional brand approved for one narrow indication, with a well-characterized efficacy claim and a contraindication for patients on a common comorbidity medication. The approved library is immaculate. Indication, efficacy claim, and contraindication each exist as a clause-cited, MLR-signed module in the system of record.
Measure pickup the only way that works: per engine. Take the questions HCPs actually ask, "what is Varigel indicated for," "how effective is Varigel," "who should not take Varigel," and run each across all five engines, scoring every answer as surfacing or invisible and recording the source that carried it.
The picture that comes back is the gap made concrete. On the indication, two engines surface the approved phrasing cleanly because a structured page carried it; two echo a third-party drug-information site that paraphrases the indication a shade loosely; one ChatGPT answer leans on an encyclopedic source that is close but not the approved wording. On the efficacy claim, the approved module is invisible across the board: every engine reaches for a press summary rather than the module with its reference attached, exactly the citation-and-statistic form the GEO study found gets echoed. On the contraindication, the most consequential of the three, only one engine surfaces it at all, and it does so from a patient forum rather than the brand, because the approved safety module lives only inside a CLM deck no retriever can reach.
Now pickup is not a worry, it is a ranked backlog with addresses. The efficacy module, invisible everywhere, gets published as structured text with its reference attached, so the approved claim enters the candidate pool the engines draw from. The contraindication module, which existed but was trapped in a deck, gets rendered as retrievable web content, so the safety statement competes with the forum instead of ceding the answer to it. The indication, already surfacing in two engines, gets watched for drift in the other three. Re-run the measurement in a few weeks and the surfacing-versus-invisible split moves. No new claims were written. The approved content that already cost MLR hours simply became reachable, and the team can finally see, per engine, whether the machine is echoing it.
Where this leaves you
The MLR effort behind your approved content is sunk cost the moment the module is signed. Pickup decides whether that cost ever returns. Content the engines echo to HCPs is review hours doing their job; content that exists but never surfaces is review hours the audience never benefits from. The split is invisible until you measure it, and you can only measure it engine by engine, because the engines do not agree on what to echo.
This is the work Juncture's Content Intelligence is built for. It reads your approved modules and claims from the system of record, scores AI Pickup as how much of that approved content the models actually echo, and shows the surfacing-versus-invisible split per engine so you can see which modules are doing their job and which cleared MLR only to disappear. Paired with Answer Monitor, which tracks what the engines tell HCPs against the same approved label, the loop closes: measure the echo, find the invisible modules, make them reachable, and turn approved content that was a cost back into content the machine repeats.
Sources
- WebFX, "AI Overviews in Healthcare: What Our Study of 130K+ Health Queries Reveals." webfx.com
- Doximity, "2026 State of AI in Medicine Report." doximity.com
- Profound, "AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information." tryprofound.com
- AuthorityTech (citing Averi's March 2026 analysis of 680M citations), "ChatGPT vs Perplexity: Only 11% of Cited Sources Overlap." authoritytech.io
- Single Grain, "Measuring Share of Voice Inside AI Answer Engines." singlegrain.com
- Aggarwal et al., "GEO: Generative Engine Optimization," Princeton University and IIT Delhi, KDD 2024. arxiv.org