What Is Agentic Commerce? A Definition (and the Two Layers Underneath It)

Why this is happening now

Until recently, AI models were confined to answering questions. Ask a chatbot to recommend running shoes and it would give you an opinion, formed from its training data, and stop there — the actual purchase still ran through a human, a browser, a storefront.

That boundary is dissolving. Agents are moving from advising to acting: browsing live sites, comparing real-time pricing and stock, and — via emerging protocols such as MCP, UCP, Mastercard's Agent Pay and Visa's Intelligent Commerce — authenticating and completing checkout without a person in the loop for each step. The infrastructure for machine-executed transactions is maturing fast enough that this is no longer a thought experiment. It's a live buying channel, and it operates on entirely different rules to the ones brands have spent two decades optimising for.

The two layers of agentic commerce

Agentic commerce splits cleanly into two layers, and most of the confusion brands run into comes from treating them as one thing.

Layer 1: Discovery

This is the agent forming an opinion. When asked an open-ended question — 'what's a good pair of trail runners for wet weather?' — the agent draws on its training data and whatever it can retrieve in the moment to build a shortlist.

This layer is probabilistic: it runs on sentiment, content, reviews, PR, and the general shape of what the model has learned about a category and the brands in it. This is the discovery layer covered under AEO — a brand's job here is to exist favourably in the model's reasoning at all.

Layer 2: Transaction

This is the agent acting on that shortlist. Once a brand has been surfaced, the agent needs to verify it — is this actually in stock, in this size, at this price, right now — and, where checkout is agent-executed, complete the purchase.

This layer is deterministic: it runs on structured product feeds, accessible APIs, and machine-readable trust signals the agent can check without ambiguity. This is the transaction layer covered under ACO — a brand's job here is to be verifiable and executable, not just mentioned.

The two layers only work together. A brand can dominate discovery — sit at the top of every relevant shortlist — and still lose the sale at the transaction layer if the agent can't verify live inventory or complete checkout cleanly. When that happens, the agent doesn't wait. It moves to whichever competitor's infrastructure it can actually transact with. Discovery gets a brand considered; transaction gets it chosen.

What breaks in the old model

Most commerce infrastructure was built for a human buyer, and much of it quietly assumes a human is still there. Session-based checkout flows, CAPTCHAs, product pages built as unstructured HTML rather than structured data, trust signals designed to be read rather than verified — none of it was built with a non-human buyer in mind.

An agent can't reliably parse a product description written for a person, can't click through a CAPTCHA meant to keep bots out, and has no way to verify a claim that only exists as marketing copy rather than structured, checkable data. A brand can be technically 'online' and still be functionally invisible, and worse, un-transactable, to an agent.

What agentic commerce requires of a brand

At a minimum, this means:

• Structured, live product data — feeds an agent can query directly, not scrape and guess at
• An accessible API or feed layer — the actual handshake point an agent transacts through
• Machine-readable trust and inventory signals — stock, price, and authenticity an agent can verify without a human checking behind it
• A genuine discovery presence — none of the above matters if the brand never makes the shortlist in the first place

The closing thought

Agentic commerce isn't a new marketing channel to be bolted onto an existing strategy — it's a new buyer, with its own requirements, sitting between the brand and the person who used to make the decision directly. Brands that treat it as a channel will spend their effort on content and sentiment, win the discovery layer, and still lose the sale — because the infrastructure work needed to actually transact with a machine buyer was never built.