You do the work buyers ask AI for. It sends them to a competitor

The RFP is still where the award happens. What changed is who gets on the list that receives it. Procurement and sourcing engineers now run an early research step before the RFP goes out, and a growing share of that step happens inside an assistant. If the model does not surface you when they describe the job, you are not on the list that gets asked to quote. The referral still works when it happens. The problem is the buyer who has no referral and starts by asking AI who can do the work.

Yes, and that is the usual shape of it. Most manufacturers we work with have a capable marketing person and someone who owns the site, and neither has had reason to learn how retrieval and AI answers work yet. We work through your team, not around them. We do the parts that need the specialized knowledge, entity structure, capability content, and measurement, and we hand back the reasoning so your people can carry it. The goal is that your team ends up able to run this themselves, not dependent on us to keep it running.

Usually, yes. A shop that runs coating, fabrication, and assembly tends to get tied to whichever line has the most public footprint, the one with the most pages, the most mentions, the clearest story. The model builds its picture of you from what it can read, and if eleven of your twelve processes live in a capabilities PDF or a single line on an overview page, they carry almost no weight. You know you do all twelve. The model knows the one you happened to write the most about.

We build a set of capability queries the way your buyers actually phrase them, the process plus the material plus the spec, and run them across the assistants on a schedule. Then we track whether you surface, for which capabilities, and against whom. The measure is not traffic. It is presence in the answer for the work you do. Here is how we measure what the model knows you can do.

It is one of the most common ones. A capabilities PDF, an equipment list, a tolerance table, a certifications sheet, these hold the exact detail a buyer's question turns on, and they are the part of your site retrieval reads least well. The fix is not to delete the PDFs. It is to make the same facts exist in text the model can parse, on pages built around the capability rather than buried in a download. The spec sheet stays for the human. The plain-text version is what gets you retrieved.

Small shops often do better here. Retrieval rewards a clear, specific match to a specific question, and a shop that does two processes well for one kind of part is easier to tie to those processes than a large supplier who does everything for everyone. Scale buys footprint in traditional search. It matters less in a system looking for the cleanest answer to a narrow capability question. If you are the obvious answer to "who does X for Y," size is not what decides it.

Structural fixes show up in retrieval within weeks, once the capability lives in text the model can read and the pages are built around the work rather than the company. Entity accuracy, the model reliably tying every line to your shop, moves over months, because it depends on third-party sources catching up to what you actually do. Watch presence first, whether you surface at all for a capability, then accuracy, whether the answer describes everything you offer rather than one line.