Getting a Manufacturer Named in ChatGPT Supplier Searches

"Find me a CNC machining supplier that can hold plus or minus 0.0005 inch tolerance in Inconel." What happens next depends on a routing decision the buyer never sees, and the two branches have completely different implications for a supplier.

Two paths, one interface

The browsing path. The assistant issues searches, retrieves current pages, and assembles an answer from them. Candidates come from the live web, citations usually appear, and a supplier can influence the outcome by being retrievable and stating matchable attributes.

The memory path. The assistant answers from what it already holds. Candidates are whatever was prominent in its training corpus, typically large and well-known companies, citations are absent or reconstructed, and a manufacturer cannot add itself to a training corpus.

Which path runs varies by model version, interface, subscription tier, and how the question is phrased. A buyer asking a constrained question with a specification in it is more likely to trigger browsing than one asking for "the best contract manufacturers".

The practical consequence: work on the browsing path and treat the other as weather. It is the only one a supplier can affect, and constrained questions, which are the ones real buyers ask, favour it.

What the constrained question does

A prompt carrying a tolerance, an alloy, a certification, and a region is a filter with four terms. The assistant is looking for pages that state all four, and most supplier pages state none of them in matchable form.

That is the entire opportunity. A page saying "tight tolerances in exotic alloys" cannot satisfy any term. A page stating "plus or minus 0.0005 inch in Inconel 718, AS9100D certified, Ohio" satisfies all four and almost nothing competes for it.

The capability page structure is the mechanism, and it is the same one that serves classic specification queries. Nothing about this is separate work.

Geography behaves differently here

"Who are the best AS9100 certified contract manufacturers in the Midwest?" carries a regional constraint that classic search would resolve through local ranking signals. An assistant resolves it from what pages say.

A job shop typically serves a regional radius, and most say so vaguely or not at all. Stating the service region in text, alongside the facility location, is the cheapest way to become eligible for regional supplier prompts, and it is routinely omitted because it feels obvious to the company.

Naming the region in the words buyers use matters too. "Midwest", "Great Lakes", and a list of states are different strings, and an assistant matching a prompt saying Midwest against a page saying Ohio is doing more work than one matching a page that says both.

Why the same prompt gives different answers

Assistant responses vary between runs for the same prompt. Model updates, retrieval variation, personalisation, and session context all contribute, and none is under a supplier's control.

This matters for how measurement is read. A single run is an anecdote. A fixed prompt panel run monthly against a stable engine set is the only way to distinguish a change in visibility from ordinary variance, which is what the manual method is built around.

It also means a competitor appearing once is not evidence of anything, and a competitor appearing across most runs over several months is.

What to do about refusals

Assistants sometimes decline to recommend specific companies, offering criteria instead. This is a policy behaviour rather than a visibility failure, and it varies by phrasing.

There is no way to make a supplier appear in an answer that names no suppliers. The response is to be present in the answers that do name them, and to accept that a share of prompts will return criteria rather than candidates.

Worth noting for measurement: a refusal should be recorded as a distinct outcome rather than as an absence, because counting it as "not named" makes visibility look worse than it is and hides the trend.

What actually changes the browsing path

Four things, in the order they pay.

State the specifications in text. Tolerances, alloys, capacity, certification scope, and service region, in HTML rather than in a drawing or a PDF.

Make the identity resolvable. A supplier can be retrieved and still fail to be assembled into a company, which is covered in entity consistency and is invisible from outside.

One capability per page, so a constrained prompt finds a precise match rather than a general page.

Measure monthly against a fixed panel, because none of the above can be shown to have worked otherwise.

The stage-by-stage account of how the selection runs is in how LLMs choose suppliers.

Common questions

Can a company pay to appear in assistant answers?

Not in the organic answer. Advertising products exist around some assistant surfaces and are separate from the recommendation itself.

Does having a ChatGPT plugin or GPT help?

No. Those are tools a user invokes deliberately. They do not affect whether the company is named when someone asks an unrelated supplier question.

Is it worth optimising for one assistant?

No. The retrieval layer draws on broadly the same public corpus across engines, so the work that makes a supplier eligible in one makes it eligible in others. Measure across several rather than targeting one.

What if a competitor is always named and the company never is?

Check resolution before content. A competitor named consistently usually has a coherent entity across third-party sources, which is a different problem from having better pages.

How long does it take to appear after making changes?

Quarters rather than weeks. Engines rebuild their view of an entity on their own schedules and third-party sources update slowly.