This one is for suppliers. If you buy rather than sell, the mirror image of this guide is what to let an AI agent do in procurement.
Forrester expects 20% of B2B sellers to be forced into agent-led quote negotiations during 2026, and at least one in five to answer buyers' AI agents with counteroffers delivered through agents of their own (Forrester, 2026 B2B Predictions, October 2025). The advice on offer about this is written for two other audiences. Marketers are told to earn brand mentions in AI search. Retailers are told to adopt checkout protocols. Neither describes what happens when a procurement agent goes looking for a foundry, a machine shop, or a distributor.
Having built one of those agents, I can describe it. This guide covers how a buyer's agent actually finds and qualifies a supplier, the six things a supplier's website has to expose for that to work, and the popular advice you can safely ignore.
How does a buyer's AI agent actually find suppliers?
In three passes: match the brief, confirm the company is real and current, and find an address it can act on. If the third pass fails, the first two don't matter. That last sentence is the whole guide, and the rest is detail.
Here's how it works in the system we ship. A buyer describes a need in plain language. A specialist agent that owns the supplier catalog researches the web for candidates that fit the brief by industry, region, capabilities and certifications, and checks each one against the buyer's existing suppliers so nobody gets added twice. Then it pulls a reachable email address from each candidate's site. Suppliers without a public-facing contact are dropped, because an RFQ nobody opens is a wasted RFQ slot. The surviving records go to a second agent that drafts and sends the request. Periodically the first agent re-checks existing records and flags the ones that have gone stale. The mechanics are documented in how Buyer24's agents work.
The agent, in other words, is doing what a diligent junior buyer would do with a search engine and an afternoon, except it does it in minutes and it never skips the step where it checks whether it can actually reach you. That is the step most supplier websites fail.
| What the agent is looking for | Where supplier sites usually put it | What the agent sees |
|---|---|---|
| Capabilities (processes, materials, tolerances) | A PDF brochure, an image slider | Nothing, or a filename |
| Region served | A map graphic in the footer | An image |
| Certifications | A scanned certificate as a JPEG | An image |
| A contact it can write to | A web form with a CAPTCHA | No address |
| Product or service names | Different wording on every page | Weak or conflicting matches |
| Whether you're still trading | A news page last updated in 2022 | A company that may not answer |
The gap agents are being sent to close is large. Businesses map only about 60% of their supplier networks and just 18% report full end-to-end visibility (QIMA, 2026 Global Sourcing Survey), which means buyers increasingly rely on agents to find suppliers they don't already know. A supplier the agent can't read might as well not exist for that buyer.
Why does a missing email address end the conversation?
Because the agent's next action is to send you a request for quote, and a form is not an address. An agent that can't reach you removes you from the shortlist before any human has seen it, and it does so without malice: there is simply nothing for it to do next.
The common failure modes are all reasonable decisions made for humans. A contact form with a CAPTCHA, chosen to stop spam. A "request a quote" flow that requires creating an account. A phone number as the only contact, because the sales team prefers to talk. An email address that exists but is rendered as an image to defeat scrapers. Each of these stops an agent as effectively as it stops a bot, and the agent is the one carrying a purchase order.
The fix is unglamorous. A plain-text email address that a person reads, on a page that loads without JavaScript, ideally repeated in your Organization schema as a `contactPoint`. That's the whole protocol for quote-to-drawing procurement in 2026, and it is worth noting that the site you're reading this on publishes its own contact address exactly that way.
What has to be written in plain text?
Capabilities, industries, regions served, and certifications, as HTML text on pages a crawler can index, using the words a buyer's brief would use. Not the words your brand agency chose.
The vocabulary point matters more than it looks. A buyer's brief says "aluminum die casting, A380, 5,000 units per year, ISO 9001". A site that says "precision solutions for demanding applications" doesn't match it, however good the work is. "Die casting, sand casting, CNC finishing, alloys A356 and A380" does. The agent isn't judging your copywriting; it's matching terms.
Certifications need the number and the scope, not the badge. "ISO 9001:2015, scope: aluminum sand and permanent mold castings" is a match; a scanned certificate is a picture of one. Regions served need to be text, because "we ship nationwide" and a shaded map image are not read the same way. Where you're willing to state them, lot sizes and typical lead times help, since buyers' briefs carry both.
If you want to know what the agent is trying to match against, read the other side of the transaction: our guide to what to include in an aluminum casting RFQ is literally the list of terms a well-written brief contains. Your site should answer it.
Can the agent even read your site?
Only if the HTML arrives with the content in it and the crawler is allowed in. Three things produce an empty page to an agent, and none of them show up in a browser: JavaScript-only rendering without prerendering, content behind a login, and a robots.txt that blocks AI user agents wholesale.
Test it the way the agent experiences it. Fetch your capabilities page with a command-line tool rather than a browser and read what comes back. If the capabilities aren't in that response, the agent doesn't see them either. Then check robots.txt: blocking every AI crawler was a defensible instinct in 2024 and is a way of hiding from your own buyers in 2026. Allow the known agents explicitly, or at least don't disallow them.
Keep PDFs, but as supplements. A brochure is fine as a download; it fails as the only place a capability is written down. The same problem exists on the buyer's side, where quotes arrive as PDFs and prices sit in email bodies, which is why buyers' teams struggle with data that isn't ready for AI. Suppliers who publish in text are easier to source from and easier to compare, and both help you win.
Do you need llms.txt, schema, or a protocol?
Mostly no, in roughly that order, and the popular advice has this backwards. Ahrefs analysed all 137,210 domains in its web analytics that received traffic in May 2026: about 28% had published an llms.txt file, which the authors treat as an upper bound because their customers skew technical, and 97% of those files received no requests at all that month. The few that were fetched were fetched by SEO audit tools, not by AI systems (Ahrefs, June 2026).
So: llms.txt is harmless, cheap and currently unread. Publish one if you like; we have one at /llms.txt as hygiene. Don't expect it to do anything, and don't let it displace the contact page on your to-do list.
Schema markup is worth doing, for a narrower reason than the marketing suggests. Organization schema with your legal name, address, `contactPoint` and `sameAs` links clarifies which entity you are, which matters when your company name appears in three variants across directories. Product or Service schema helps if you sell stable SKUs. Ignore the claim that complete markup makes you "6.4 times more likely to be selected" by agents; it circulates without a traceable source.
Protocols are where suppliers most often buy something they don't need. The Agentic Commerce Protocol from OpenAI and Stripe standardizes checkout between an agent and a merchant. Google's Universal Commerce Protocol covers catalog lookup, cart and checkout. The Model Context Protocol connects agents to tools and data. All three are real and all three are about retail or catalog transactions. For a supplier quoting to a drawing, none of them is where discovery happens today: the agent finds you on the open web and emails you. If you distribute stocked SKUs and buyers' agents start asking for live catalog access, MCP is the one to watch. Until then, the protocol is your inbox.
What does "current" mean to an agent?
It means the record still checks out when the agent re-visits it. Agents don't find you once; they re-audit. A site whose most recent news item is from 2022, whose certificate expired last year, and whose company name differs between the site, the LinkedIn page and the directory listing reads as a company that may not reply. It gets flagged, and flagged records get deprioritized.
The maintenance list is short. Dated content should carry a recent date or none. Certifications should show validity. The company name, address and email should be identical everywhere they appear. Dead links from old directory listings should redirect to a live page rather than a 404. None of this is new advice, but it used to be about human trust, and now it's about staying in a record.
How do you get onto a buyer's network if there's no directory?
You don't apply; you get asked. On a referral-only network there is nothing to submit to, and that is a feature rather than an omission. When a buyer's brief matches your site, the agent drafts a request and sends it to the address it found, and that request is your invitation. Your job is to be readable when it happens and to answer when it arrives.
There are two ways in, and both are worth understanding. The first is the one this guide is about: a buyer you've never met needs what you make, their agent finds you, and an RFQ lands in your inbox. The second is a buyer you already work with adding you as their supplier directly. Neither involves a listing fee, a profile to optimize, or a directory to climb. Open marketplaces fill up with directory spam and buyers stop trusting them, which is why buyers' networks are increasingly built by referral instead; the supplier side of ours is described at buyer24.ai/suppliers.
There's a symmetry here worth noticing. Suppliers triage incoming requests and decline to quote based on the buyer's clarity, volume, history and relationship (Production Machining). Buyers' agents now triage suppliers the same way, on readability and reachability. Both sides are asking the same question: is this one worth my time? Once the first request arrives, the task becomes answering fast and completely, which is the subject of how to improve your quote win rate and, when it doesn't go your way, why suppliers lose quotes.
The one-hour checklist
Contact first, because it's the drop rule. Then the text an agent has to match. Then whether it can read any of it. Everything below can be tested from a laptop in an hour, and most of it fixed the same afternoon.
| Check | Five-minute test | Pass looks like |
|---|---|---|
| Plain-text email on a crawlable page | View the contact page's source; search for "@" | An address in the HTML, not an image or a form |
| Organization schema with `contactPoint` | Paste the homepage into a schema validator | Name, address, email, `sameAs` present |
| Capabilities in HTML text | Fetch the capabilities page without JavaScript | Processes and materials readable in the raw response |
| Buyer vocabulary | Compare your terms with a real RFQ you received | Same nouns: process, alloy, tolerance class, certification |
| Certifications with number and scope | Search the page for the standard's number | "ISO 9001:2015, scope: …" as text |
| Regions served as text | Search for your states or countries | Words, not a map image |
| AI crawlers not blocked | Read robots.txt | No blanket disallow for AI user agents |
| Dated content current | Look at the newest date on the site | This year, or no dates at all |
| Consistent identity | Compare name and address across site, LinkedIn, directories | Identical strings |
| PDFs as supplements only | Find any capability that exists only in a PDF | None |
If you fix only one row, fix the first. If you fix two, add the third. The agent that will find you is patient about everything except an address it can't use.
The cost of doing this is an afternoon. The cost of not doing it is invisible by construction: you never see the RFQs that were drafted for someone else because your site couldn't be read. Procurement is early to agents, at 9% adoption against 35% in software development (HBR, 2026, 385 organizations), which means being readable now is cheap positioning for a shift that has barely started.
FAQ
How do AI buying agents find suppliers?
In three passes: they search the web for companies matching a brief by industry, region, capability and certification; they check the company is real and current; and they look for an email address they can send a request to. A supplier that fails the third pass is dropped from the shortlist, because the agent's next action is to send an RFQ.
Do I need llms.txt to be found by AI agents?
No. Ahrefs found that of 137,210 domains analysed in May 2026, about 28% had an llms.txt file and 97% of those files received no requests at all, with the few requests coming from SEO audit tools. It's harmless to publish one, but a plain-text contact address and capabilities written as HTML text matter far more.
Why isn't my company showing up to AI agents?
Usually one of three reasons: your capabilities are in PDFs or images rather than text, your only contact is a form or a phone number, or your pages render only with JavaScript so a crawler receives an empty shell. Fetch your own capabilities page without a browser and read what comes back; that is what the agent sees.
Do I need to join a marketplace to get RFQs from agents?
No. Buyers' agents search the open web, and referral-based procurement networks have no directory to apply to; the RFQ itself is the invitation. Marketplaces and paid directories may bring traffic, but the agent that emails you a request found you because your site was readable, not because you were listed.
What should be on a supplier website for AI procurement?
A plain-text email address on a crawlable page; capabilities, materials, industries and regions written as HTML text in the vocabulary buyers use; certifications with their number and scope; Organization schema with a contact point; a robots.txt that doesn't block AI crawlers; and dates and company details that are current and consistent across the web.
Key takeaways
- A buyer's AI agent finds suppliers in three passes: match the brief, confirm the company is current, find an address it can act on. Fail the third and the first two don't count.
- The single most common reason a supplier is dropped is a missing public email address. A form is not an address to an agent.
- Capabilities, certifications and regions have to exist as HTML text in the vocabulary buyers' briefs use; PDFs and images are invisible.
- Fetch your own page without a browser. If the content isn't in the response, the agent doesn't see it. Don't block AI crawlers in robots.txt.
- llms.txt is currently unread: 97% of files got zero requests in May 2026 (Ahrefs). Schema helps with entity clarity. Retail checkout protocols are not where industrial discovery happens.
- There is no directory to apply to on a referral network. Be readable when a brief matches, and answer when the RFQ arrives.
- Forrester expects 20% of B2B sellers to face agent-led quote negotiations in 2026 (Forrester). Readability now is cheap positioning for a shift that has barely started.