How AI Builds a Vendor Shortlist for an RFQ
What has to be true, in writing, on the open web, before ChatGPT, Claude, Gemini or Perplexity puts your company's name in an answer about a technical capability.
A procurement engineer looking for a five-axis titanium CNC shop used to start with a supplier directory: filter by process, filter by material, scroll through certifications in a sidebar. Increasingly, they start with a chat window instead, type the actual technical requirement, and get one paragraph back that names two or three companies. The AI already did something like that comparison, well or badly, and handed over a shortlist before the buyer opened a directory at all.
That paragraph deserves more scrutiny than most companies give it. It did not come from an engine that scored vendors against a spec sheet. It came from a language model predicting which company names are likely to appear next to a description of that capability, based on everything it has read about the industry. Understanding that mechanism, not vague advice about "AI visibility," is what tells you whether your company has any real chance of showing up in it.
What has to exist before a model can name you
For a model to write your company's name into an answer about five-axis titanium machining with AS9100D certification, that combination needs to exist somewhere in text the model has read: your name, that process, that material, that certification, close enough together to survive however the model compresses or retrieves what it read. A page that says "we offer precision CNC machining services" does not do this. It may be true, but it is not specific enough to be pulled up for a specific question, and the model has no way to know your general CNC page covers five-axis titanium work unless it says so, in words, somewhere a crawler could reach.
This is the part that catches manufacturers off guard. Plenty of shops that genuinely do the exact work being asked about have never written the exact words down anywhere public. The capability lives in a quote given to one customer years ago, in a machine list nobody put on the website, in an AS9100D certificate sitting in a binder in the quality manager's office. From a model's point of view, that capability does not exist, because nothing describing it exists in a form it could ever have read.
One mention is a claim, several are a signal
Even when the capability is written down somewhere, a single occurrence is a weak signal. A one-line claim on a company's own homepage is exactly the kind of thing a marketing department writes whether or not it holds up under scrutiny, and a model trained on enough marketing copy learns, implicitly, to discount it a little. What moves a name into an answer more reliably is the same capability described independently in more than one place: a project page on the manufacturer's own site, a listing in a trade directory that requires the certification to register, a mention in an industry publication, a forum thread where an engineer names who they used for a similar job. None of these needs to be large, but they need to be independent and agree. Ten restatements of the same unverified claim are still one thin signal wearing ten coats: what separates a strong signal from a weak one is provenance, not repetition.
Stale signals mislead a model exactly like missing ones
A certification page written a while ago and never revisited is a liability, not an asset. If AS9100D lapsed, or the five-axis machine was sold and replaced with three-axis capacity, and nobody updated the website, a model can still confidently name that company for work it can no longer do, because the old text is still the most recent version it has read. This fails the buyer just as badly as if the company had never been mentioned, only later, after the RFQ has gone out. And it is not something you fix once: a model's picture of you is built from whatever was written the last time something, somewhere, described you.
Why the narrow question is easier to win than the broad one
"CNC machining services" is a question with an enormous number of legitimate answers, most of them generic enough that a model has no strong reason to prefer one company's name over another's. It defaults to whichever names appear most often across the widest range of generic content, which rewards whoever has published the most, not whoever is actually best suited to the job. "Five-axis titanium CNC with AS9100D certification" is a different question: far fewer companies can honestly claim all three parts at once, which shrinks the field enough that a single well-documented page can plausibly be the strongest signal in the entire set.
This runs against how most marketing content gets written, aimed at appealing to as many searchers as possible, which is exactly what makes it forgettable to a system trying to match a narrow question. The page that wins a narrow query is written for the narrow query: process, tolerance, material, certification, in the words an engineer would use, not the words a brochure would use to sound impressive to everyone at once. Imagine a contract manufacturer that does exactly this work, five-axis titanium parts to tight tolerances, fully AS9100D certified, but has only ever described itself online as a "full-service precision machining partner." It could be the best-qualified vendor in the country for that RFQ and still never appear in an answer to it, because nothing it published gives a model a reason to connect the two.
A check you can run yourself this afternoon
Pick the single most specific real capability your company has, the kind of thing that would appear on an RFQ, not a homepage headline. Phrase it the way a buyer or engineer sourcing it would type it, not the way your marketing team would describe it: "who makes five-axis titanium parts to AS9100D" rather than "who is a leading precision manufacturing partner." Run that exact question in ChatGPT, Claude, Gemini and Perplexity, each on its own, and read who gets named.
If you are named, notice how: has the model described the capability accurately, or guessed and gotten a detail wrong? If you are not named, look at who is, and ask honestly whether they have published the thing you were just asked about more clearly than you have. Either way, be honest about what this told you: one capability, phrased one way, on four platforms, on one day. These are probabilistic systems, so the same question run again tomorrow can return a different list of names for reasons unrelated to anything you changed. One run is an anecdote. Knowing where you actually stand means asking the questions that matter to your business, worded the way your real buyers word them, repeatedly, across every platform and market you sell into. That is a measurement problem, and it is the specific one PSentry is built to track: it runs your prompt set across all four platforms on a schedule and shows which capabilities get you named, and which quietly do not.
What this does not tell you
None of this means documenting a capability precisely guarantees a model will name you next time someone asks. It does not, and any tool that promises it is selling something it cannot deliver. The mechanism above is a diagnosis, not a lever you can pull with certainty. If the capability is not written down anywhere independent of you, that is a fixable gap. If it is written down and current and you are still not named, the honest question is who is winning that query and what they have published that you have not, not a vague resolution to do more "AI optimization." A measurement tool can tell you which of these is happening; it cannot write the pages or promise a model will change its mind, because nobody outside the companies that build these models can promise that.
Frequently Asked Questions
Does the exact wording of the query change who gets named?
Yes, more than most people expect. Models respond partly to how closely a question resembles the way the underlying documents describe the same thing, so a question built from an engineer's vocabulary, tolerances, certification codes, process names, tends to retrieve differently than one built from marketing vocabulary like "partner" or "solution."
If our company is not named, does that mean the model does not know we exist?
Not necessarily. It can mean the model knows you exist in general, has read your generic marketing copy, and simply has no text connecting you to that capability. Usually that is a documentation gap for one thing, not a verdict on the whole company.
Do third-party directories help more than our own website?
Differently, not necessarily more. A directory listing is useful mainly because it is independent of you, but most directories only capture a broad category, not the exact combination an RFQ needs, so they rarely replace writing the capability out in full sentences somewhere.
Can we submit our company so an AI recommends it?
No. None of the major platforms has a submission form or paid placement for this. The only way in is how everything else enters a model's view of the world: something, somewhere, describes the capability in detail, corroborated independently enough times to become part of what the model has read.
How often should we re-run this kind of check?
Running it once tells you about one day. The same question can return different vendor names on different runs, and certifications and capabilities change over time, so a single check has a short shelf life. Checking repeatedly catches both a gap that opens and a stale claim that lingers after the capability has changed.