Agentic commerce: when AI buys on your behalf.

What is agentic commerce?

Agentic commerce refers to the set of commercial transactions in which an autonomous software agent, an AI assistant, acts on behalf of a human buyer: it searches for a product, compares offers, applies constraints (budget, deadline, conditions), then triggers the order and the payment. The human gives the intent and the limits; the agent carries the execution.

It should be distinguished from two neighboring notions. Predictive AI recommends (recommendation engines). Generative AI produces content (descriptions, answers). Agentic commerce adds a third capability: acting, that is, completing an operation with a third party. It is the shift from self-service commerce (a human browses) to delegated commerce (an agent transacts).

How does agentic commerce work?

An agentic transaction follows four steps: discovery (the agent finds the machine-readable offer), cart (it assembles and validates the contents), checkout (it structures the order with the merchant) and payment (it triggers an authorized settlement). Each step relies on a different technical layer, which the Universal Commerce Protocol framework breaks down into four layers: semantic, agent, transaction, governance.

A key point for merchants: in most current models, the merchant remains the seller of record. The agent does not become the merchant; it relays a purchase intent and a controlled payment token. Control of the catalog, the price and the customer relationship stays on the merchant side. Established

The two modes: assisted and delegated

Most confusion about agentic commerce comes from collapsing two very different things into one word. The specification separates them explicitly, and the dividing line is a single clause.

UCP's Checkout capability requires an order to be finalized manually by the user through a trusted UI, unless the AP2 Mandates extension is supported. Established Everything follows from that. In assisted commerce, the agent researches, compares and assembles the cart, then a human approves. In delegated commerce, the buyer signs an intent mandate up front carrying limits such as a price ceiling and a time window, and the agent executes once those conditions are met, with nobody at the screen.

Almost every agentic purchase you can make today is the assisted kind. The delegated kind is what changes the economics of replenishment, procurement and subscription, and it depends on AP2 mandates being in place on both sides. Reading an adoption claim without asking which of the two modes it refers to is how most market estimates become meaningless.

What a transaction actually looks like

A checkout session is a state machine, not a form submission. The status field tells the agent what is expected next: incomplete, requires_escalation, ready_for_complete, complete_in_progress, completed or canceled. Errors are separately graded by severity: recoverable, requires buyer input, requires review, or unrecoverable. Established

requires_escalation is the state worth understanding, because it is where the real world lives. Age verification, a signature threshold, a supporting document, a pricing exception, a restricted delivery zone: these are not failures of the model, they are a defined hand-off back to a human. A merchant that cannot express its conditions ends up invisible to agents not because its products are wrong, but because its rules are unreadable.

Two related guarantees are written into the same capability. The Merchant of Record does not change, and the capability introduces no PCI DSS compliance obligation for card payments. Eligibility is handled through claims that must be resolved before a transaction can complete, which is the mechanism behind contract pricing and geographic restrictions.

Source: UCP specification, Checkout capability and specification overview, version 2026-04-08.

It already runs at platform scale

Agentic commerce is often discussed as a forecast. Part of it is already infrastructure. Shopify exposes two catalog interfaces implementing UCP's Catalog capability: a Global Catalog at catalog.shopify.com/api/ucp/mcp that searches across every eligible merchant on the platform, and a Storefront Catalog at {storeDomain}/api/ucp/mcp scoped to a single store. Both expose the same three operations: search_catalog, lookup_catalog and get_product. Authentication uses an agent profile, with no API key required. Established

Two details tell you how seriously the platform takes it. Global Catalog results are clustered by Universal Product ID, so the same product from different merchants resolves as one entity, which is exactly what comparison shopping needs. And Shopify already sells promoted placements inside those results, with commission on attributed purchases. The advertising layer arrived at the same time as the open standard, which is worth noting before treating agent discovery as a neutral channel.

Source: Shopify, About Catalogs.

The protocols of agentic commerce

Several standards are emerging separately, each covering one layer. Here is the reading grid, with their backer and public announcement date.

These protocols are not competitors layer by layer: they stack. A merchant can be discoverable via MCP, expose a transaction via UCP or ACP, and accept a payment authorized by AP2. Only UCP and ACP genuinely contest the same transaction layer. For the payment layer, see agentic payments and AP2. Emerging

Concrete examples in 2026

• ChatGPT and OpenAI: OpenAI offered direct purchase via Instant Checkout (ACP), scaled back in 2026, and now relies on Visa for agent payments. The merchant remains the seller of record. Established
• Google, AI Mode and the Gemini app: Google and Shopify are rolling out native checkout via the UCP protocol, announced at NRF in January 2026, with Walmart, Target, Etsy, Wayfair and 20+ partners. Established
• Universal Cart and beyond: at Google Marketing Live (May 2026), Google expanded UCP to a multi-retailer cart (Universal Cart), Affirm/Klarna buy-now-pay-later inside Google Pay and Shopping ads on YouTube, with hotel booking and food delivery verticals announced. Details on the UCP reference page. Established

Generalization (multi-country, multi-vertical, multi-agent) remains a trajectory in progress, not a given. Prospective

What agentic commerce changes for a merchant

To stay purchasable by an agent, the challenge is no longer only to rank well for a human, but to be readable, negotiable and actionable by a machine: structured product data, explicit policies (delivery, return, availability) and an agent-usable endpoint. The readiness audit methodology assesses this state layer by layer, and the use cases by business type illustrate the concrete trade-offs.

Stakes and limits

Three issues remain open. Fraud and authorization: proving that an agent acts on a human's mandate (the purpose of AP2's cryptographic mandates). Governance and neutrality: preventing a single platform from capturing the transaction layer. Traceability: keeping verifiable proof of purchase intent. See the market timeline analysis.

How to tell whether any of this reaches you

Until 2026 the honest answer to "are AI engines recommending my products" was that nobody could measure it. Both major engines now publish the data, and they do not count the same thing.

On 3 June 2026, Google launched generative AI performance reports in Search Console, covering AI Overviews and AI Mode in Search as well as Discover. They report impressions, the pages that appeared, plus country, device and date breakdowns down to hourly granularity. The rollout is deliberately limited to a subset of sites. Established

In February 2026, Microsoft opened AI Performance in Bing Webmaster Tools, covering Copilot, Bing's AI answers and selected partner surfaces. It reports total citations, average cited pages, page-level citation counts, and grounding queries, the phrases the model actually used to retrieve the content. Microsoft explicitly frames the release as an early step toward Generative Engine Optimization tooling. Established

The asymmetry matters more than either metric alone. An impression tells you that you were shown. A citation tells you that you were used, and a grounding query tells you what question you were used to answer. That last signal is the closest thing to a keyword report the generative era has produced, and it is the one worth instrumenting first. See GEO and catalog readability.

Sources: Google Search Central, 3 June 2026, and Bing Webmaster Blog, February 2026.

The agentic commerce timeline

The stack came together in under two years: MCP in late 2024 (tool access), A2A in spring 2025 (agents to each other), ACP and AP2 in autumn 2025 (transaction and payment), then UCP in January 2026 (transaction, on the Google and Shopify side). This acceleration is what makes the topic actionable now rather than in three years. Established

Frequently asked questions