How LLMs Choose Which Suppliers to Recommend

A buyer asks for an ISO 13485 injection molder for a Class II device. Three companies get named. Understanding why those three requires separating three things that happen in sequence and are usually discussed as one.

Three stages, not one ranking

Retrieval. The engine gathers documents plausibly relevant to the question. This resembles search and is the stage most familiar, but the unit is a passage rather than a page, and a long page contributes only the passages that match.

Entity resolution. The engine decides which mentions refer to the same company. A supplier appearing as "Acme Precision", "Acme Precision Machining LLC", and "Acme Precision Inc" across a site, a directory, and an association listing may resolve as one entity or as three. Three weak entities lose to one coherent one.

Attribute matching. The question carried constraints: a certification, a device class, a process. The engine checks which resolved entities have stated attributes satisfying them.

A supplier can pass the first and fail the second, which is invisible from the outside. The pages were retrieved and the company was never assembled.

Why third-party sources dominate the middle stage

Entity resolution is a cross-source problem by definition. One website asserting its own name is a single data point; resolution improves when independent sources agree.

That inverts the usual priority. Most search work treats the site as the thing to fix and third-party mentions as a bonus. For supplier discovery the site supplies the attributes and the third-party corpus supplies the confidence that the attributes belong to one identifiable company.

Practically: a certification listed on the site, in a registrar's directory, and in an association membership record is a stronger fact than the same certification stated three times on the site. A discrepancy between them is worse than an omission, because it forces the engine to choose and it may choose the stale one.

Constrained questions are the ones that matter

Unconstrained prompts return generic answers assembled from general sources. Constrained prompts, the kind buyers actually ask, filter hard on stated attributes.

"Find a CNC machining supplier that can hold plus or minus 0.0005 inch tolerance in Inconel" is answerable only from pages that state a tolerance and name the alloy. A page saying "tight tolerances in exotic materials" satisfies no part of it.

This is why capability page structure is the operative lever rather than a content preference. Process, materials, tolerances, capacity, and certification scope are the five constraint types buyers attach, and a page carrying them is eligible for questions a page of adjectives is not.

What does not decide it

Domain authority is not the constraint it is in classic search. On one measured certification query, a manufacturer at domain rating 29 outranked a directory at domain rating 66, and the ranking page was a general About page that happened to state the accreditation.

Volume of content does not help. A hundred thin pages produce a hundred weak passages. Retrieval selects passages, and a passage that states nothing specific loses regardless of how many neighbours it has.

Schema markup does not substitute for stated facts. Generative engines cite sources carrying explicit, extractable, attributed statements. Markup helps a machine parse a statement that exists; it does not create one.

What actually moves it

Four things, in order of return.

State the attributes in text. Every tolerance, alloy, capacity, and certification scope in HTML rather than in a drawing or a datasheet.

Make the identity consistent everywhere the company appears. Same legal name, same address, same certification claims, on the site and in every third-party listing.

Structure one capability per page, so a constrained question finds a precise match rather than a general page mentioning several things.

Measure whether it worked, using the manual method, because none of the above can be shown to have changed anything without a baseline.

The wider argument about how much of this is new work is in generative engine optimization for industrial companies, and the answer is mostly that it is not.

What cannot be known from outside

No engine publishes how it weights a supplier's own site against third-party corroboration, and it cannot be inferred from output alone: the same recommendation is consistent with several different weightings.

What is observable is which sources get cited alongside a recommendation, which is why the citation field in the measurement method is the most useful column and the most commonly skipped.

Common questions

Does being in training data matter more than being retrievable?

Retrievability is the part a supplier can affect. Training corpora are fixed at a point in time and a company cannot add itself to one, while retrieval operates over current sources.

Do assistants prefer directories?

Not on the evidence available. On the certification query measured, one directory appeared in eight organic positions and sat below a manufacturer with a fraction of its authority.

What if an engine states something wrong about the company?

Treat it as the highest-priority finding. An engine repeating a wrong certification or process is actively disqualifying the supplier, and the fix is usually in the third-party sources rather than on the site.

How long does a change take to show up?

Longer than a ranking change and less predictably, since engines rebuild their view of an entity on their own schedules and third-party sources update slowly. Quarters rather than weeks, as covered in AI search visibility.

Is there a way to be recommended without being findable in classic search?

Not reliably. The retrieval stage draws on broadly the same public corpus, so a page that classic search cannot reach is usually not available to an engine either.