# How Agentic Commerce Works

The agent shopping loop step by step: intent, discovery, evaluation, checkout, payment authorization, and fulfillment, with who does what.

Updated: 2026-10-04

> **In short:** An agentic purchase runs through six stages: capturing intent, discovering products, evaluating options, checking out, authorizing payment, and handling fulfillment and after-sales. Each stage has competing technical approaches, from structured APIs and open protocols to an agent operating a website the way a person would. The table near the end summarizes who does what.

## The shopping loop at a glance

1. **Intent capture.** The user states a goal; the agent turns it into constraints.
2. **Discovery.** The agent finds candidate products from catalogs, feeds, search, or browsing.
3. **Evaluation and comparison.** The agent checks fit, price, availability, reviews, and policies.
4. **Checkout.** The agent assembles a cart and submits an order through a protocol, a browser, or a handoff.
5. **Payment and authorization.** A credential is presented with proof the user approved this purchase.
6. **Fulfillment and post-purchase.** Shipping, tracking, returns, and support.

The loop is not always linear. Agents often return to discovery after evaluation fails, and some stages (price monitoring, reordering) run in the background over days.

## 1. Intent capture

The user describes what they want in natural language ("a carry-on that fits Ryanair's free size, under 100 euros"), sometimes with an image. The agent converts this into structured constraints: category, dimensions, budget, delivery date, brand preferences.

Key practices:

- **Clarifying questions.** Shopping-focused modes such as ChatGPT's shopping research (launched November 2025) ask about budget, use, and recipient before searching.
- **Memory and context.** Agents tied to an account (for example Alexa for Shopping, which draws on Amazon order history) can infer sizes, past brands, and addresses.
- **Mandates.** For delegated purchases, the intent is recorded as an authorization with limits. In [AP2](/wiki/ap2) this is an *open* mandate: a signed record of what the user asked for and the bounds (price cap, merchants, expiry). AP2 v0.2 (April 2026) calls these open Checkout and Payment Mandates; the original spec called them intent mandates.

## 2. Discovery

The agent assembles a candidate set. There are four broad sources, and most agents mix them:

| Source | How it works | Strengths | Weaknesses |
|---|---|---|---|
| Merchant product feeds | Merchants submit structured catalogs (title, GTIN, price, availability, attributes) to a platform such as Google Merchant Center or an AI assistant's merchant program | Accurate, current if maintained | Coverage limited to merchants who submit |
| Catalog APIs and aggregators | Agents query a multi-merchant product database through an API or [MCP](/wiki/mcp) server | Broad, normalized, fast | Depends on the aggregator's freshness and coverage |
| Web search and indexes | Agents use a search engine's index and crawl product pages | Very broad coverage | Unstructured, prices and stock may be stale |
| Live browsing | The agent loads merchant pages directly | Current information | Slow, brittle, may be blocked |

Protocols increasingly formalize discovery. [UCP](/wiki/ucp) includes catalog search and lookup capabilities, and Shopify exposes merchant catalogs to AI channels through its Catalog and MCP servers. The quality of this layer is covered on [product data](/wiki/product-data).

## 3. Evaluation and comparison

The agent narrows candidates against the user's constraints. Typical checks:

- Attribute fit (size, compatibility, materials).
- Total price including shipping, tax, and discounts.
- Availability and delivery date to the user's location.
- Reviews and ratings, sometimes summarized.
- Merchant policies: returns window, warranty.

Evaluation quality depends heavily on structured data. An agent cannot reliably compare "fits under an airline seat" if dimensions are buried in an image. This is why Google added conversational attributes such as compatibility and common questions to Merchant Center in January 2026, and why [agent optimization](/wiki/agent-optimization) focuses on machine-readable facts.

At this stage the agent usually presents a shortlist to the user. How much it decides alone is a question of [autonomy level](/wiki/levels-of-autonomy).

## 4. Checkout

Checkout is where approaches diverge most.

### Native protocol checkout

The agent calls a merchant's (or platform's) checkout API using a shared protocol. The merchant returns a priced cart with tax and shipping; the agent confirms; the merchant creates the order and remains merchant of record.

- [ACP](/wiki/acp) (OpenAI and Stripe, September 2025) defines checkout session endpoints and a delegated payment step.
- [UCP](/wiki/ucp) (Google, Shopify, and partners, January 2026) defines capability discovery, cart, checkout, and order management, with REST, MCP, and A2A bindings.
- Platform-specific implementations exist: Shopify announced WebMCP checkout tools for browser-based agents in September 2026.

Advantages: reliable, fast, and the merchant controls pricing and policies. Disadvantage: the merchant or its platform must integrate.

### Browser automation

The agent operates the merchant's website like a person: clicking, filling forms, entering payment details. OpenAI's Operator (January 2025), Perplexity's Comet browser, and Amazon's Buy for Me (which purchases on other brands' sites) work this way in whole or part.

Advantages: works on any site without merchant cooperation. Disadvantages: brittle when pages change, slower, and often in tension with merchants' terms. Amazon has sued Perplexity and blocked Meta's Muse over agent access to its store. Identity schemes such as [Web Bot Auth](/wiki/web-bot-auth) and Visa's Trusted Agent Protocol let agents identify themselves cryptographically so merchants can distinguish them from malicious bots.

### Handoff

The agent prepares the selection, sometimes a pre-filled cart, then sends the user to the merchant's own checkout or app. After ending its Instant Checkout feature in March 2026, OpenAI moved most ChatGPT purchases to this model, including retailer apps inside ChatGPT. Handoff preserves the merchant's checkout and conversion data at the cost of a less seamless experience.

## 5. Payment and authorization

The core problem: the merchant and the card issuer need confidence that a real customer approved this specific purchase, even though software submitted it. Approaches:

- **Delegated or shared payment tokens.** The agent never sees the card number. In ACP, a payment provider issues a token scoped to one merchant and amount; Stripe's version is the Shared Payment Token.
- **Network agentic tokens.** Visa Intelligent Commerce and Mastercard Agent Pay issue tokenized card credentials bound to a specific agent, with spending controls set by the user.
- **Single-use virtual cards.** Stripe's Link agent wallet (announced April 2026) issues a one-time card per agent task.
- **Mandates.** AP2 uses signed verifiable credentials. Since v0.2 (April 2026) there are two types, a Checkout Mandate (what is being bought) and a Payment Mandate (how it is paid), each moving from an *open* stage (the user's constraints) to a *closed* stage (one specific, approved transaction). Version 0.1 used intent and cart mandates for the same idea. These records give issuers and merchants evidence in disputes.
- **Agent-native rails.** [x402](/wiki/x402) lets agents pay over HTTP, usually in stablecoins, mainly for APIs and digital goods.

Authentication may include passkeys or issuer step-up at the moment of approval. Risk systems are adapting: in September 2026 Mastercard began testing a score estimating whether a transaction was agent-initiated. See [payments and trust](/wiki/payments-and-trust).

## 6. Fulfillment and post-purchase

Once the order exists, physical fulfillment is usually unchanged: the merchant ships. What changes is the communication path.

- **Order status.** Protocols such as UCP include order webhooks so the agent can report tracking to the user.
- **Returns and cancellations.** The agent may initiate a return through a protocol endpoint or direct the user to the merchant. Policies need to be machine-readable for this to work.
- **Disputes.** If the agent bought the wrong item, the question is whether the user, the agent provider, the merchant, or the issuer absorbs the cost. Mandates and agent registration programs exist partly to answer this. See [consumer protection and liability](/wiki/consumer-protection-and-liability).
- **Customer relationship.** Merchants generally remain merchant of record and receive the customer's details, but the agent sits between them and the shopper.

## Who does what

| Stage | User | Agent | Merchant / platform | Payment provider / network |
|---|---|---|---|---|
| Intent capture | States goal and limits | Clarifies, records constraints or mandate | — | May register mandate or spending controls |
| Discovery | — | Queries feeds, APIs, search, or pages | Publishes structured catalog and inventory | — |
| Evaluation | Reviews shortlist (if required) | Compares fit, price, policies | Exposes accurate attributes, reviews, policies | — |
| Checkout | Approves cart (or pre-authorized) | Builds cart, calls checkout or operates site | Prices cart, calculates tax and shipping, creates order | — |
| Payment | Authenticates when required | Presents token or credential | Charges via its processor | Issues scoped token, verifies agent, scores risk |
| Fulfillment | Receives goods | Reports status, initiates returns | Ships, handles returns and support | Handles disputes and chargebacks |

## Sources

- [Agentic Commerce Protocol — agenticcommerce.dev](https://www.agenticcommerce.dev/)
- [Universal Commerce Protocol — ucp.dev](https://ucp.dev/)
- [New tech and tools for retailers to succeed in an agentic shopping era — Google](https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/)
- [Google launches payments protocol for AI commerce — Digital Commerce 360](https://www.digitalcommerce360.com/2025/09/19/google-ai-payments-protocol-ap2/)
- [Shopify opens checkout to browser-based AI agents — TechCrunch](https://techcrunch.com/2026/09/28/shopify-opens-checkout-to-browser-based-ai-agents/)
- [OpenAI Ends Instant Checkout in ChatGPT — The Keyword](https://www.thekeyword.co/news/openai-chatgpt-instant-checkout-scrapped)
- [Stripe builds out the economic infrastructure for AI with 288 launches — Stripe](https://stripe.com/newsroom/news/sessions-2026)
- [Mastercard Deploys New Risk Insights to Support Agentic Payments — PYMNTS](https://www.pymnts.com/mastercard/2026/mastercard-deploys-new-risk-insights-to-support-agentic-payments/)
- [Visa's Trusted Agent Protocol sets new standard — Help Net Security](https://www.helpnetsecurity.com/2025/10/14/visa-trusted-agent-protocol/)
- [OpenAI debuts ChatGPT Shopping Research — SiliconANGLE](https://siliconangle.com/2025/11/24/openai-debuts-chatgpt-shopping-research-tool-personalized-product-discovery/)
- [Amazon pilots 'Buy for Me' AI shopping tool — Digital Commerce 360](https://www.digitalcommerce360.com/2025/04/08/amazon-buy-for-me-agentic-ai-shopping-tool/)
- [Introducing x402 — Coinbase](https://www.coinbase.com/developer-platform/discover/launches/x402)