When your B2B buyer is an AI agent: how agentic vendor discovery works

Published June 14, 2026 · Last reviewed June 14, 2026

Almost everything written about agentic commerce is about retail. McKinsey’s widely-cited forecast, that AI agents could mediate 3 to 5 trillion dollars of consumer commerce globally by 2030 and roughly 1 trillion in the US, is a projection about shopping carts: groceries, electronics, apparel, simple purchases an agent can complete end to end.¹ That story is real, but it is not the change most B2B companies will feel first.

The change they will feel first is quieter and already underway: their buyers are sending agents to do the vendor research. Forrester’s 2026 survey of B2B buyers found that 94 percent now use AI somewhere in the purchase process, up from 89 percent the year before, and that twice as many buyers named generative AI as their single most meaningful research source than named any other source, ahead of vendor websites, product experts, and sales reps.² The shortlist your sales team competes to get on is increasingly assembled by a machine.

This article is about how that B2B discovery actually works: how an agent builds a shortlist, what it needs from a vendor that a static page cannot provide, the agent-to-agent layer that almost no coverage touches, and an honest read on where the agentic stack is still wobbling. Here is how it works in detail.

Agentic commerce is written for retail, and B2B is the gap

Read the headline forecasts closely and they are all about the consumer cart. McKinsey’s 1-trillion-dollar US figure is agentic retail revenue, about 30 percent of projected B2C across retail categories. The whole genre, the retail-media decks and the checkout-API launches, assumes a transaction an agent can finish on its own: pick the item, compare a few, pay.

B2B vendor selection breaks every one of those assumptions. The purchase is considered, not impulse. A buying group decides, not one person, and Forrester notes those groups are getting larger as buyers question AI output. Fit matters more than price: whether a tool supports SSO, offers regional data residency, integrates with the stack already in place, or scales to a 40-person team. None of that fits a product feed, and none of it closes in a checkout flow.

So the retail playbook does not map. What does map is the part of the funnel that comes before any transaction: discovery, comparison, and qualification. That is exactly where B2B agents are active now, and it is the slice the agentic-commerce coverage mostly skips. McKinsey’s own framing of AI search as the “new front door to the internet” reported that about half of consumers already use AI-powered search for buying decisions.¹ In B2B the front door is the same, but what waits behind it is a research problem, not a cart.

How an agent actually builds a B2B shortlist

Picture the request the way the agent receives it: “find three vendors that do contract lifecycle management, support SSO and EU data residency, and fit a 50-person legal team.” The agent does not browse the way a person does. It decomposes that into requirements, then gathers evidence against each one from wherever it can: AI search engines, third-party directories and review sites, and the vendor sites themselves.

A growing share of that work happens before anyone loads your homepage. The same agentic browsers now visiting sites on a buyer’s behalf, ChatGPT Atlas, Perplexity Comet, Gemini in Chrome, are the tools doing the gathering, and B2B buyers increasingly research vendors inside AI engines before they ever click through. The agent reads each candidate in milliseconds, extracts what it can parse, scores it against the requirements, and narrows the field to the two or three it will surface to its human or the buying group.

The blunt consequence is that your site either participated in that comparison or it did not, and the buyer will never know which. If the agent could not extract a clear answer on data residency, you are not the vendor that “failed” the question; you are the vendor that silently dropped off the list. This is a different failure mode from a lost demo. There is no form fill to retarget and no bounce to analyze, because the evaluation left no trace in your analytics at all. The first signal you get is a pipeline that quietly thins.

What the agent needs that a static page cannot give

Follow the agent past the marketing copy and it has specific, often novel questions. Not “what does this product do,” which it can read, but “does this do the particular thing my buyer needs, under my buyer’s constraints.” A static site offers two ways to answer that, and both fail an agent.

The first is the contact form. Humans already abandon those at high rates; Zuko’s benchmarking across 93 million form sessions puts average completion at just over half, around 52 percent, with the rest abandoned.³ An agent acting for a buyer either cannot complete a form or should not, because a fabricated lead helps no one. The second path is “request a demo and we will get back to you.” That loses on timing. A 2011 Harvard Business Review study found that answering a web lead within an hour made a meaningful conversation with a decision-maker roughly seven times likelier than waiting a day.⁴ That study assumed a human with human patience. A buyer’s agent queries several vendors in parallel and builds its shortlist from whoever answered, in seconds, not hours.

Being machine-readable helps the first step and only the first step. Publishing clean structure and an llms.txt file lets the agent extract your stated facts, and it is cheap hygiene worth doing. It does not answer a fit question the agent did not find pre-written on a page, and the honest evidence is that readability is a floor, not a lever: a SE Ranking analysis across roughly 300,000 domains found no measurable lift in AI citations from llms.txt alone.⁵ The open gap is not legibility. It is the ability to hold a short, specific exchange and qualify on the spot.

The agent-to-agent layer almost nobody covers

Most coverage of agentic discovery stops at “make sure agents can read your site.” The step it skips is the one that actually closes the gap: an agent on your side of the connection that the buyer’s agent can talk to. Not a chatbot waiting for a human to arrive first, but a representative that answers another agent’s questions directly, qualifies whether the buyer fits, and routes the conversation to a person only when it is worth a person’s time.

A small number of vendors are building toward this. Salespeak ships a discovery endpoint at /.well-known/mcp that any agent can find, a neutral example of the same direction.⁶ Tobira’s Site Agent sits on that side of the line with its own emphasis. Its job is captured in five verbs: Answer, Qualify, Fetch, Route, Intro. Answer the visiting agent’s questions, qualify fit, Fetch a missing answer by querying other agents on the network when the site itself cannot respond, route a warm lead to the owner, and make a human introduction only after both sides consent. To be precise about what that is and is not: it makes a site agent-addressable and networked. It does not make a site agent-readable; the readable layer, structure and markup and llms.txt, is separate work that no conversational agent does for you.

Two design choices matter for B2B trust. The representative is tied to a human-readable @handle, so the buyer’s agent is talking to an identity that maps to a real company and person, not an anonymous endpoint. And contact is exchanged only when both sides agree, with standing built from conversation history rather than a self-asserted badge: Tobira expresses that as a credibility level on a 0-to-5 scale, not a trust score. The point of the networked layer is not to win the agent. It is to be present, answerable, and qualifiable at the exact moment the shortlist gets built.

The honest part: discovery is maturing faster than checkout

It would be easy to extend this into “and then the agents transact,” but the evidence says slow down. OpenAI wound down its Instant Checkout feature in March 2026 after weak adoption — only around 30 Shopify merchants were ever available and fewer went fully live — and shifted to routing buyers back to retailers’ own sites.⁷ The payment rails for full agent-to-agent commerce, AP2, x402, and their siblings, are real but early, and the transactional end of the stack is the least settled part of it.

That stumble is not a reason to dismiss agentic buying. The useful reading is the split it exposes: research and comparison through agents kept growing while checkout-in-chat faltered. The layer where agents gather information and qualify vendors is maturing years ahead of the layer where they pay. For B2B that split is convenient, because the “transaction” was never a cart anyway. It is a signed contract that a human and a buying group approve. The agent’s job in B2B is discovery and qualification, the part that is working, not autonomous purchase, the part that is not.

So the practical priority is narrow. Do the readable floor cheaply and without inflated expectations, then ask the question worth taking to a pipeline review: when a qualified buyer’s agent reaches you with a real fit question, what answers it, and how fast? If the honest answer is a form and a follow-up email, you are running a 2011 response path against a visitor that decides in seconds. Close that gap deliberately, with a human-facing answering layer or an agent that also serves the visiting agents, rather than by default.

What to remember

• Agentic commerce forecasts are retail. McKinsey’s 3-to-5-trillion-dollar global and roughly 1-trillion-dollar US figures by 2030 are consumer cart projections, and B2B vendor discovery is a separate, less-covered slice.
• B2B buyers already research with AI: Forrester puts AI use in the 2026 purchase process at 94 percent (up from 89), with generative AI named the single most meaningful research source twice as often as any other.
• An agent builds a shortlist by decomposing a need into requirements and gathering evidence from AI engines, third parties, and vendor sites, usually before any human loads your page. You either made the list or silently dropped off it.
• A static page cannot serve that agent. Forms get abandoned (Zuko: only about 52 percent of started forms complete) and “we will get back to you” loses a race measured in seconds, not the hours the 2011 HBR speed-to-lead finding assumed.
• The uncovered layer is agent-to-agent: a representative on your side that answers, qualifies, and routes, tied to a human-readable identity with consent. Fix the answering and qualification layer before betting on agent checkout, which is still the wobbly part (OpenAI wound down Instant Checkout in March 2026).

FAQ

What does agentic vendor discovery mean? It is the process by which an AI agent, acting for a buyer, finds and shortlists vendors with little or no human browsing. The buyer states a need, the agent decomposes it into requirements, queries AI engines, vendor sites, and third-party sources, extracts what it can, and presents a short list of options. In B2B this matters because the comparison often happens before any human loads a vendor page.

How do AI agents discover B2B vendors? Through a mix of AI search engines, direct visits to vendor sites, and third-party sources like directories and review sites. Forrester found 94 percent of B2B buyers use AI somewhere in the 2026 purchase process, and twice as many named generative AI as their single most meaningful research source than any other source. The agent reads in milliseconds, does not fill forms, and ranks whoever gave it usable answers.

Can you sell to an AI agent? You can sell through one. A B2B agent rarely buys outright; it qualifies fit and assembles a shortlist for a human or a buying group to decide on. So the goal is not to close the agent but to answer its questions well enough to make the shortlist: handle specific fit questions, qualify the buyer, and route a worthwhile conversation to a person, fast enough that the agent does not move on.

Is agentic commerce only for retail? No, but most coverage is. McKinsey projects AI agents could mediate 3 to 5 trillion dollars of consumer commerce globally by 2030, roughly 1 trillion in the US, and those figures are retail. B2B vendor discovery is a separate, less-covered slice: considered purchases, multiple stakeholders, and fit questions that a product feed and a checkout API were never built to answer.

What is the difference between an agent-readable and an agent-addressable website? Agent-readiness is about how agents read your site: llms.txt, clean markup, markdown mirrors. An agent-addressable site adds the other half: a representative agent that visiting agents and humans can find and talk to, which answers questions, qualifies fit, and routes worthwhile conversations to a person, tied to a human-readable name so a real person is reached only with consent. They are complementary layers, not substitutes.