The Supplier Search That Happens Behind the Login

PSentry cannot see this, and nor can anything else sold from the outside. When a buyer at a large manufacturer asks the sourcing assistant inside their procurement suite for suppliers of a machined component, the question, the answer and every name in it stay behind a corporate login. No crawler reaches it, no public API returns it, and no monitoring tool, ours included, has the faintest idea what was said. It seems only fair to begin there, since the rest of this piece concerns a channel we do not measure.

It is still worth understanding. Supplier discovery was never confined to search engines and trade directories, and some of it now happens inside the source-to-pay platforms that large buyers already run their purchasing through. The vendors of those platforms have spent the past year or so putting agents into them.

What has actually been announced

Two announcements are enough to establish the direction without turning this into a product review. On 18 November 2025 Coupa announced a set of agents that included a Navi Supplier Discovery Agent, described as using natural-language queries to help sourcing managers find suppliers that fit. A few weeks earlier, on 6 October 2025, it had announced an agreement to acquire Scoutbee, which it called an AI-powered supplier discovery platform and network. In May 2026, at its Sapphire conference, SAP presented a family of Joule Assistants for procurement, among them a Sourcing Assistant said to manage the sourcing life cycle from drafting RFPs and bids to recommending negotiation strategies, and a Supplier Management Assistant for oversight of the supply base.

What neither announcement spells out, at least in the public material, is precisely which data the agent consults when it puts a supplier forward. That matters, and it is where honest writing on the subject has to slow down. What follows is the mechanism as far as it can be reasoned from how these platforms are built. It is not a description of any vendor's internals, which are theirs to document.

What an agent inside the platform has to work with

A public assistant answering on the open web draws on what it absorbed in training and on whatever pages it fetches while answering. An agent inside a procurement suite starts from somewhere else entirely: the data the platform already holds.

• The buyer's own supplier records. Suppliers already onboarded, the categories they are approved for, their qualification status, past sourcing events, contracts, spend. An agent grounded in these will, reasonably enough, lean towards companies the buyer already knows how to pay.
• The platform's supplier network. Both suites run networks in which suppliers keep a profile. SAP describes its Business Network Discovery as matching the postings buyers create against supplier profiles that set out the products and services offered and the locations served. A profile is what a matching process can read about you. Nothing else is.
• Catalogue content. Where a supplier publishes a catalogue into a network, item descriptions, part numbers and prices become records the platform holds, rather than text on a website that something would have to go and find.
• Whatever external data the vendor chooses to add. Coupa buying a supplier discovery platform with a network of its own suggests an interest in reaching beyond a customer's existing supplier base. How much outside data reaches the agent, and from where, is for Coupa to say, and we will not guess on its behalf.

Why this is a different problem from the open web

The contrast with a public assistant is sharper than it first appears.

On the open web, whether you are named depends on what a great many independent pages say about you, in the language of the question, weighed in ways nobody outside the model's maker can inspect. It is diffuse, it is slow, and your own site is one voice among many.

Inside a platform, the evidence is narrower and far more literal. The agent is reading structured records: a category code, a country served, a certificate with an expiry date, a catalogue line. Structured data is unforgiving in a particular way. A website that describes precision machining in fluent prose can still be matched to a five-axis titanium enquiry, provided it says so somewhere. A network profile filed under the wrong commodity category, or under none, is not a weak signal. If the agent narrows its candidates by category first, as a conventional supplier search does, that record may never get as far as the clever part.

Currency matters for the same reason. A certificate that has lapsed in the records, while the renewal sits in the quality manager's inbox, tells a qualification check exactly what it says. Nobody reads the covering email.

There is also an asymmetry that exporters in particular ought to notice. The buyer's own records carry the incumbents. An agent that draws on past events and existing contracts starts from the suppliers a buyer already uses, and in a foreign market that is usually somebody local rather than you. The network profile is the one part of that data set a newcomer can actually write.

What a mid-sized exporter can check

None of this needs a tool, ours or anyone else's. It needs an afternoon and the logins you very likely already hold, if a large customer has ever onboarded you through one of these platforms.

• Which networks you are on, and how many times. A customer that buys through Ariba or Coupa will have invited you to a supplier account at some point. It is not unusual for a company to have several, opened by different people for different customers, none of them maintained by anyone in particular.
• The profile, read as a matching process would read it. Product and service categories at the most specific level offered, not the broadest one that happens to be true. The countries and regions served, which for an exporter should be the markets you actually deliver to. A company description that names processes, materials and standards, rather than calling you a trusted partner.
• Certificates and their dates. Quality and sector certifications, and anything a qualification questionnaire is likely to ask for. The current version uploaded, the expired one removed, the next renewal noted somewhere a person will see it.
• Catalogue data, if you publish one. Descriptions written in the words a buyer would search with, part numbers that match your datasheets, units stated properly. A catalogue built for invoicing convenience is frequently useless for discovery.
• Contacts. A matched opportunity routed to someone who left three years ago is not an opportunity.

Language deserves a line of its own. Where a profile field is free text, it is worth asking which language the buyers you want will be working in, and whether a description written only in your home language says anything useful to them.

Where measurement stops

Which brings us back to the limit. PSentry measures what ChatGPT, Claude, Gemini and Perplexity say about a brand, language by language and market by market, 2 scans a month. That is the public half of discovery: the engineer who asks an open assistant before anyone has raised a sourcing event. It says nothing about what an agent inside a buyer's procurement suite proposes, and there is no honest way to claim otherwise from the outside.

The two halves are not unrelated. A buyer who first hears your name from a public assistant may go and look for you in the network; a buyer whose platform proposes you may well check you on the open web before sending an invitation to bid. But they are observed, and maintained, separately. The public half can be measured. The private half can only be kept in good order, which is less exciting and, for a supplier who wants to be found, rather more within reach.