When AI Recommends a CDMO, What Is It Actually Evaluating?

Claims in this piece are observable and methodological — no client or benchmark data used.

A pharmaceutical buyer asks an AI assistant: “Which CDMOs in India can manufacture sterile injectable products for regulated markets?” The answer comes back with several company names, descriptions of their capabilities, and reasons each appears relevant.

It can look as though the AI has evaluated those manufacturers and decided which are better. It hasn’t — not in the way a pharmaceutical buyer evaluates a supplier. It has not audited a facility, reviewed quality systems, assessed a technology-transfer programme, or decided whether a supplier can be trusted with a complex programme. What it has done is narrower: it has taken the information available to it, interpreted that information in the context of the question, and constructed an answer.

Most manufacturers, when they see such an answer, ask the wrong question about it. They ask: “Why is that company above us?” This article is about the more useful question — and it starts with noticing that an AI answer is a description, not a ranking, and should be read as one.

Three things that are not the same

When an AI assistant answers a buyer’s question about your company, three distinct things are in play, and confusing them is where manufacturers go wrong:

• The business itself — what your CDMO genuinely manufactures, the facilities it operates, the markets it serves, the capabilities it holds.
• The public record — what an outsider can actually establish from your website and other available sources.
• The AI representation — what an assistant can currently establish and communicate about you when a buyer asks a particular question.

Those three are not necessarily the same, and the gaps between them are the whole commercial issue. A manufacturer cannot easily change how a given model behaves. But the distance between the business and the public record — and therefore what the AI can establish — is something a manufacturer can observe and influence.

So the question isn’t “why did AI rank that company above us?” It is: “what could an AI-assisted buyer actually establish about us — and is that representation accurate?” That is something you can test.

Read the description, not the ranking

Here is the reframe that matters, and it is the point of this article.

Imagine an assistant returns three CDMOs for a sterile-injectable question. One is described simply as “a pharmaceutical manufacturer.” Another is described in terms of sterile fill-finish and injectable products. A third is described primarily around oral solid dosage.

The order of those names tells you almost nothing reliable — it can change with the prompt, the platform, or the day. But the descriptions tell you something stable and useful: the system has constructed a different representation of each organisation from the information it could find. That is where a commercial team should look.

Most manufacturers instinctively ask, “why aren’t we number one?” The more actionable question is: “what does the assistant believe we do, and what information is it using to explain why we are relevant?” The first question chases a position on a list that isn’t a real leaderboard. The second directs attention to the underlying public record — the thing you can actually improve.

An AI recommendation, in other words, is best read not as a score you are trying to climb, but as a readout of your own legibility — a description of what an outside system was able to establish about you. The ranking is noise. The description is signal.

Four questions to inspect the description

These are not a proprietary scoring system, and not claims about how any model works internally. They are four practical checks a commercial team can run on an AI answer about its own company:

• 1. Did it identify the right relevance? If the buyer asked about sterile injectables, does the assistant associate you with sterile injectables — or describe you as a general manufacturer? A specialist with genuine HPAPI containment or fill-finish experience can be described generically if the public record doesn’t connect that capability to the requirement.
• 2. Did it describe the capability specifically? Does it name the actual dosage form, technology, or therapeutic area — or retreat into “end-to-end solutions” and “comprehensive services”? Specificity gives an outside system something concrete to work with; generic language gives it little. (This is the distinction between capability and legibility: a company can hold a capability without communicating it clearly enough for an outsider to establish the connection.)
• 3. What evidence supports the description? Ask the assistant: “What public evidence supports that assessment?” Then inspect it. Does the source actually support the claim? Is it current? A fluent answer can still rest on weak or irrelevant information. The answer is not the evidence.
• 4. Is the representation accurate? Compare what the AI says with what your business actually does. Does it confuse an old capability with a current one, attribute a capability to the wrong site, confuse development with commercial manufacturing, or omit something you genuinely want buyers to understand? An AI response can be fluent, specific, and still wrong.

The objective across all four is not a more flattering answer. It is to make the accurate answer easier to establish.

The answer is not the evidence

This deserves to stand on its own, because it is what separates useful interpretation from dangerous over-reading: an AI-generated recommendation is an interpretation of available information, not evidence of manufacturing quality.

It does not establish GMP compliance, regulatory standing, manufacturing quality, technical suitability, capacity, technology-transfer competence, or programme fit. Those require appropriate evidence and human evaluation. This is why AI output belongs at the start of a supplier investigation, not its conclusion — the fuller version of this argument, from the buyer’s side, is in How Procurement Teams Can Use AI to Evaluate Manufacturing Partners, which draws the line between AI as a research lead and AI as a reference check.

For a manufacturer, the implication is symmetrical: a company appearing in an AI answer is not necessarily better; a company absent is not necessarily worse. The useful observation is narrower — what reasons for considering you could the system establish from what was available to it?

When the public record is fragmented

Why would a genuinely strong manufacturer be described vaguely? Usually because the public record is fragmented in ways that are invisible from inside the company. The website says “injectables.” An older PDF says “lyophilised injectables.” The LinkedIn description emphasises oral solids. A facility page doesn’t connect the site to either capability.

Internally, there is no confusion — the quality team knows the facility, the technical team knows the equipment. But an outside researcher, and an assistant working from public information, has none of that internal context. The resulting description may be incomplete or inconsistent, not because the capability is absent, but because the public record never made it legible. That coherence problem — and what a website needs to make establishable — is covered in Your Pharmaceutical Website Is No Longer Just a Brochure and its companion, Seven Trust Signals Every Pharmaceutical Website Should Demonstrate.

How to read your own description

The exercise follows directly from the reframe. Don’t start with “tell me about our company” — that produces a summary, not the situation a buyer creates. Start with a realistic buyer question:

Which CDMOs in India can manufacture sterile injectable products for regulated markets?

Run it, unchanged, through ChatGPT, Gemini, Perplexity, Claude and Copilot. For each, record the prompt, platform, and date; which manufacturers appear; how each is described; what capabilities are attributed; what evidence is offered; and where the answers differ. Then, for your own organisation, ask each assistant: “Why did you include [Company]?”, then “What public evidence supports that?”, then “What important capability is unclear or missing?”

Read the descriptions, not the order. The useful finding is never “we appeared” — it is whether the representation is accurate or inaccurate, specific or generic, complete or missing, supported or unsupported, consistent or contradictory across sources. Those observations tell you far more than a name’s position in a list, and — unlike the position — they point at something you can fix. Results vary by platform, prompt, and time, so this is reproducible observation, not a benchmark.

Where Emerivo fits

Emerivo is a specialist advisory for pharmaceutical manufacturing and research organisations — TPMs, CMOs, CDMOs and CROs. It does not serve buyers. Its focus is AI Visibility: helping these organisations understand how AI assistants find, interpret and represent them when buyers research potential partners.

Emerivo delivers this through the AI Discovery Audit™, which tests an organisation against realistic buyer prompts across ChatGPT, Gemini, Perplexity, Claude and Copilot and records what each platform says, gets wrong, or cannot establish. The purpose is not to promise rankings or recommendations — no credible organisation can guarantee those. It is to establish a baseline: what can an AI-assisted buyer currently establish about this organisation, and where is that representation inaccurate, incomplete, or unclear?

The practical takeaway

Don’t judge your AI Visibility by whether an assistant said something flattering, and don’t treat a recommendation as proof you are stronger than a competitor. The harder, more useful test is: when a real buyer asks a real supplier question, what can an AI assistant accurately establish about you?

Run the question. Read the descriptions, not the ranking. Ask what evidence supports them. Look for omissions, ambiguity, and inaccuracy. Then compare the representation with the business you actually run. The objective is not to manufacture a better answer — it is to make the accurate one easier to find, understand, and verify. That is the real commercial issue: not whether AI can qualify your CDMO, but whether it can accurately establish why a buyer should consider you, before your team ever gets the chance to explain it.

Frequently asked questions