How AI builds supplier shortlists

When a buyer asks an AI system who supplies something, the system runs searches, retrieves pages, and composes an answer from what those pages say. The pattern is called retrieval-augmented generation, and it is the reason your website still matters: the answer is built out of documents, and a company no document describes clearly cannot be in it.

One question usually becomes several searches — query fan-out. "Who supplies case packers for glass jars at low volume" may be broken into narrower queries about case packers, glass handling and low-volume lines, with the answer stitched together from all of them.

The companies that end up named are the ones described clearly enough, in enough of those results, to be worth including — not the ones with the best machines.

Source: Optimizing your website for generative AI features on Google Search, Google Search Central.

Why it does not just answer from memory

Because memory is out of date and cannot be checked.

A language model trained months ago has no reliable knowledge of who currently supplies what, which companies still trade, or what a manufacturer has added to its range since. Retrieval exists to fix that: the system goes and looks, then writes from what it found. It is also what makes citation possible at all — a system can point at a page it retrieved, but not at a memory.

Your equipment is not being judged. Your documentation is being read.

What actually happens, in order

The question is expanded. One buyer question becomes several narrower searches, because the original is usually too broad to answer directly.

Documents come back. Pages from the search index — your site, competitors' sites, trade registers, directories, association listings, technical articles.

They are weighed. Material that is clear, relevant and consistent with other material is more useful to build an answer from than material that contradicts itself or says nothing specific.

The answer is composed, and sometimes the sources are named. Whether an answer cites anything at all is Citation Presence, and how strong those sources are is Citation Authority — both defined on Prove.

What gets retrieved, and what does not

Only what is in the index and can be processed. A page that is not crawled, blocked, or too slow and broken to read is not a candidate, however good the company behind it.

Beyond that, the material that earns its place is specific. A page headed "Packaging solutions" that describes a philosophy is of little use to a system answering a question about glass jars at 40 units a minute. A page that states the application, the format, the rate and the constraint is directly usable, and gets used.

This is why thin, brochure-style sites fare badly here regardless of how the company is regarded in its market. There is simply nothing on them to build a sentence out of.

Why consistency matters more here than anywhere else

Because the system is comparing several sources about you at once.

If your own site, a directory entry and an association listing describe your company three different ways, there is no settled version to use — and an answer assembled under uncertainty tends to reach for a company it can describe without hedging. A competitor with a duller website and four agreeing descriptions is in a stronger position than you are.

That is also why corrections to old listings can change answers quickly. You are not competing for attention when you fix them — you are removing a contradiction, which is the subject of when AI describes your company wrongly.

Does it favour big companies?

Only where the question is broad.

A general question about packaging machinery manufacturers draws on decades of published material, and scale wins that. A specific question about an application with three genuine suppliers draws on whatever exists about that application — and if you are the only company that has written it down properly, the retrieval step has nowhere else to go.

Which is the whole argument for depth over breadth in this work, and the reason we look at narrow questions first when we measure.

What this means for a supplier

Four things follow directly, and none of them is exotic.

Be retrievable: crawlable, indexed, quick, readable. Be specific: write the application, the format, the rate, the constraint. Be consistent: one description of your company everywhere it appears. And be corroborated: make sure something other than your own website says so.

The last is the slowest, and it is what separates a company that is named from one that is merely present.

What this article cannot tell you

It cannot tell you how any particular system weighs what it retrieves. None of them publishes that, they differ from each other, and they change. What is described here is the shape of the process, not a formula.

It cannot tell you which sources matter most in your category. That varies by industry and country and is a question for measurement rather than for argument.

And it cannot promise the process stays like this. AI systems are probabilistic and change continuously. What has been true throughout is narrower and more useful: answers are built from material that can be found, and a company nothing describes is a company nothing can name.

What to do next

Read your own pages the way a retrieval step would. Take the five questions your buyers ask most often before they call, and ask whether any page you own states the answer plainly enough to be lifted out and quoted.

Then check what the systems currently return. You can do that yourself in an afternoon, or have it done properly — agreed questions, all six systems, your named competitors alongside you and every source recorded — with the AI Visibility Audit. Either way, the useful part is seeing which documents the answers are actually built from, and whether any of them are yours.