How AI Checkout Actually Works

Last updated on July 1, 2026

AI assistants are becoming a place people shop. A shopper can describe what they want, compare options, and move toward buying without ever opening a category page.

If you sell online, that raises three practical questions:

  1. How does buying through an AI assistant actually work today?
  2. Can you connect your store to ChatGPT, Gemini, or Claude right now?
  3. If you run a US store, what are the concrete steps and links?

This article answers those three first, in plain terms, then covers the payment and protocol details underneath. Status labels are marked as of July 2026, because this space moves quickly.

What buying through an AI assistant looks like today

Picture a shopper typing into ChatGPT: “comfortable black walking shoes under $150, in stock in EU 42.” The assistant reads product data from stores that made their catalog available, shows a few matching options, and the shopper moves toward buying. From there, checkout follows one of two paths.

Shopper Asks for a product
AI assistant Recommends options

Then checkout splits into two paths

In-chat checkout Shopper pays inside the assistant
  1. Shopper Confirms and pays in chat
  2. Merchant Validates cart, tax, shipping
  3. Payment processor Charges the card
  4. Merchant Creates the order
  5. AI assistant Shows order status
Storefront handoff Shopper pays on the store
  1. Shopper Taps buy, opens the store
  2. Store checkout Store checkout opens
  3. Shopper Pays with existing methods
  4. Merchant Creates the order

The assistant starts the buying journey. The merchant checkout and payment stack still complete the transaction.

The two paths differ in where the shopper pays:

  • Native in-chat checkout: for a small set of approved merchants, the shopper pays without leaving ChatGPT. This is OpenAI’s Instant Checkout, and it is gated to approved partners.
  • Storefront handoff: for most merchants, tapping “buy” opens the store’s own checkout in an in-app browser (or a new browser tab on desktop), with the merchant’s branding and payment methods. This is how Shopify describes its Agentic Storefronts channel completing a ChatGPT sale.

Both patterns keep the merchant as the seller, and both are here to stay. The assistant helps the shopper decide and prepare the order; the store still owns payment, fulfillment, and support. That is the mental model for everything below.

Can you connect your store right now?

Short answer: partly, and it depends on where you already sell. Here is the state as of July 2026.

SurfaceWhat works todayWho can join nowHow the sale completes
ChatGPT (native Instant Checkout)Live in the US, single-item, limited to a small set of approved merchantsApply at chatgpt.com/merchants (approved partners)Inside ChatGPT
ChatGPT (storefront handoff)Shopify says eligible products can surface via Agentic Storefronts; Etsy listings were live at the Instant Checkout launchShopify and Etsy merchants, through their platformOn your own store checkout
Google AI Mode and GeminiAnnounced; rolling out to eligible US retailersEligible US retailers, via Google’s Universal Commerce ProtocolUniversal Cart / your checkout
ClaudeNo native checkoutAnyone, but you build the toolsWherever your tool flow sends the shopper

The verdicts:

  • On Shopify or Etsy: Shopify says products can be surfaced in ChatGPT through its Agentic Storefronts channel, and Etsy listings were part of the Instant Checkout launch. Confirm it is enabled for your store, then focus on making the catalog and checkout ready.
  • On another platform: native ChatGPT checkout is application-only through chatgpt.com/merchants, and approvals are limited today. You can still prepare your product feed now.
  • For Google’s AI shopping: Google has announced checkout on its AI surfaces and is rolling it out to eligible US retailers through the Universal Commerce Protocol. Treat it as announced and rolling out, not a guaranteed live path for every retailer yet.
  • For Claude: there is no store to “connect.” Claude can only shop through tools a developer builds. Treat it as a build-your-own path, not a channel you sign up for.

US store: what to do this week

The right steps depend on your platform. None of them require building a payment protocol.

If you sell on Shopify or Etsy
  1. Confirm your products are discoverable. Shopify says eligible products can be surfaced in ChatGPT through its Agentic Storefronts channel; check whether it is enabled in your admin.
  2. Clean the catalog an agent will read: clear titles, real variants (size, color, SKU), current inventory, accurate price, honest images, readable shipping and return policies.
  3. Test your checkout in a mobile in-app browser. ChatGPT completes many sales there, so a checkout that only works well on your full desktop site will lose orders.
  4. Nothing to build for payments. Your existing processor and checkout do the work.
If you sell on another platform or custom stack
  1. Apply for native ChatGPT checkout at chatgpt.com/merchants. Access is limited to approved partners today, so treat it as a queue, not a switch.
  2. Prepare a structured product feed now (file upload or API), following OpenAI’s commerce onboarding guide. This is useful even before approval.
  3. For Google’s AI shopping surfaces, keep your Google Merchant Center feed complete and check eligibility for the Universal Commerce Protocol.
  4. Make sure your own checkout can be reached and completed from a link, so a storefront-handoff flow works cleanly.

Either way, the near-term win is not a new payment rail. It is a catalog and checkout that software can read and a shopper can finish on a phone.

Who gets paid, who is liable, and what it costs

This is where teams overthink AI commerce. The answer is stable across OpenAI, Shopify, Google, Visa, and Mastercard.

  • You stay merchant of record. The order, payment, refunds, and support run on your existing systems. OpenAI states this directly in its Instant Checkout launch, and Shopify keeps the merchant as merchant of record for Agentic Storefronts.
  • You keep your payment processor. The assistant does not become your acquirer. It passes a scoped credential or a signed authorization, and your processor charges the card as usual.
  • The assistant never sees raw card data. Payment moves through scoped tokens (a credential limited to one merchant, cart, and amount) or signed mandates that record what the shopper approved.
  • Cost depends on where the sale completes. OpenAI says merchants pay a small fee on completed in-chat purchases; some reports put the native Instant Checkout fee at around 4%. A sale completed on your own store checkout adds no fee beyond your standard processing, per Shopify’s description of Agentic Storefronts.

The durable principle: a language model can tolerate ambiguity (“this jacket looks right”), but a payment cannot (“this amount looks close”). So payment is deliberately kept exact and outside the model, held by the store and its processor.

How the payment actually moves (the technical half)

Two checkout patterns cover almost everything today.

Native in-chat checkout runs on OpenAI’s Agentic Commerce Protocol, co-developed with Stripe. The merchant exposes checkout-session endpoints (create, update, complete, cancel, retrieve) and accepts a scoped payment token, so the agent can complete a purchase without holding card credentials. Payment stays with the merchant’s processor.

Storefront handoff skips the in-chat payment step. The agent surfaces the product, then opens the merchant’s own checkout in a browser. All the store’s existing payment methods, including wallets and bank payment, work unchanged. This is simpler to support and is how most stores participate today.

If bank payment (Pay-by-Bank) is your specific interest, the deeper build guide is in the agentic commerce protocols article. For the catalog and product-data layer that makes any of this work, see how to make your store readable by AI shopping agents.

The protocols, in one place

Four names come up constantly. They sit at different layers, and mixing them up is the main source of confusion.

NameOwnerLayer it solves
ACP (Agentic Commerce Protocol)OpenAI, StripeProduct, cart, checkout, and delegated payment for ChatGPT
UCP (Universal Commerce Protocol)GoogleProduct and checkout state for Google’s AI shopping surfaces
AP2 (Agent Payments Protocol)GoogleProving what the user authorized, through signed mandates, across payment methods
MCP (Model Context Protocol)AnthropicHow an assistant calls external tools; the path for Claude and custom agents

None of these turn the assistant into a payment processor. They standardize how the store and the assistant exchange product, checkout, and authorization data. Support a protocol shape and you are ready for a channel; you still need the channel, the merchant, and the processor to be live.

What to build regardless of platform

Whichever assistant wins in your market, the same commerce primitives are what make you ready.

Build these first
  • Structured product and variant data (title, SKU, price, image, stock, policies).
  • Server-side cart and checkout-session APIs, not UI that only a browser can drive.
  • Explicit user confirmation before any payment.
  • Payment authorization through your current processor, wallet, or bank provider.
  • Webhooks and status polling so the order is created only after verified payment.
  • A checkout that completes cleanly in a mobile in-app browser.

If you are building the payment product, not selling on it

Some readers are not merchants; they are building a checkout or payment product: a Pay-by-Bank provider, a wallet, or an agentic-checkout layer. For that reader, the near-term opportunity is not native distribution inside ChatGPT or Gemini, which is gated and not self-serve today. It is agent-ready checkout primitives: hosted checkout sessions, signed intent, a clear authorization handoff, webhooks, status APIs, and refund and reconciliation handling that an assistant can call and a merchant can trust. Build those, and the product is ready to plug into each assistant as its native channel opens. The full build is covered in the agentic commerce protocols guide.

The durable takeaway

AI commerce does not replace your store. It changes which surface the shopper uses first. The website stops being the only storefront, and your catalog, checkout, and order APIs become storefront infrastructure too. The stores that win will not be the ones with the most protocols. They will be the ones a shopper can find in a chat and finish paying in one tap.

Frequently Asked Questions

Can you buy things directly inside ChatGPT?

Yes, in two ways. A small set of approved merchants support native Instant Checkout, where the purchase completes inside ChatGPT. Far more merchants, such as many selling on Shopify or Etsy, can appear in ChatGPT and complete the purchase on their own store checkout through an in-app browser. Both keep the merchant as merchant of record.

Can I connect my store to ChatGPT, Gemini, and Claude right now?

If you sell on Shopify or Etsy, those platforms can surface your products in ChatGPT through their own integrations, so you likely do not need a separate application; confirm it is enabled for your store. Other merchants apply to OpenAI at chatgpt.com/merchants (approved partners, US). Google has announced AI-surface checkout and is rolling it out to eligible US retailers through the Universal Commerce Protocol. Claude has no native checkout; it can only act through tools a developer builds with MCP.

Does the AI assistant become my payment processor?

No. The assistant helps the shopper decide and prepare the order. The merchant stays merchant of record and charges through the existing payment processor. Payment moves through scoped tokens or signed mandates so the assistant never handles raw card credentials.

What does it cost to sell through ChatGPT?

OpenAI says merchants pay a small fee on completed in-chat purchases; some reports put the native Instant Checkout fee at around 4%. A sale completed on your own store checkout adds no fee beyond your standard payment processing. In both cases you keep your existing processor.

Is this the same as SEO for my products?

No. Search ranking sends a person to a page to interpret it. AI shopping asks software to understand your product, pick the right variant, verify availability, and hand off to checkout. That depends on structured catalog data and reachable checkout, not meta tags.

Pay-by-Bank and Agentic Commerce: What ACP, AP2, MCP, and UCP Actually Enable

Can Pay-by-Bank Work With Agentic Commerce Protocols?

Native AI checkout distribution is still gated. Pay-by-Bank providers can still build hosted sessions, bank authorization handoff, webhooks, status APIs, and AP2-ready consent records now.

How to Make Your Store Visible to AI Shopping Agents

How to Make Your Store Visible to AI Shopping Agents

AI shopping visibility starts with distribution paths, but it only works when your catalog, variants, inventory, policies, and checkout can be read by agents without guessing.

Multilingual Voice AI Agents and Code-Switching: The Engineering Guide for Real-Time ASR and TTS

The Code-Switching Gap: Where Multilingual Voice AI Loses Callers Mid-Sentence

Hinglish and Spanglish callers do not speak one language per call. Here is how to build an ASR-to-TTS pipeline that follows them across the switch instead of breaking on it.

Lowest-Latency Voice AI Agents: The Engineering Budget From Microphone to Speaker

The Core Latency Budget: Every Millisecond Between Microphone and Speaker

Streaming is not an answer. Here is the full turn-gap budget broken into twelve components, each in milliseconds, with the techniques that actually move the number.

Voice Prompt Engineering for AI Agents: Why Text Prompts Break in Real-Time Audio

Voice-Specific Prompt Engineering: Why Text Prompts Break in Real-Time Audio

A prompt that works in ChatGPT reads markup aloud, says 'two thousand five' for a year, and talks over the caller. Here is the prompt-engineering playbook for streaming voice agents.

Voice AI Telephony Stack: SIP, WebRTC, PSTN, and IVR Replacement

The Telephony Layer Under Your Voice Agent (And When It Breaks)

Carriers, codecs, SBCs, and caller-ID rules sit invisible beneath every voice agent until an outage or a spam label surfaces them. Here is what runs down there and which decisions actually matter.

AI Voice Agents for Personal Injury Intake: Solving the Missed-Call Problem

AI Voice Agents for Personal Injury Law Firms: How to Automate Intake Calls

AI voice agents handle personal injury intake 24/7 with attorney-level qualification. Technical deep-dive covering architecture, bilingual support, compliance, and real production results.

Building AI That Actually Understands Legal Documents: RAG Architecture for 500-Page Contracts

Building AI That Understands Legal Documents (Not Just Reads Them)

Engineering perspective on legal document AI: difference between text ingestion and contextual reasoning, RAG architecture for massive contracts, and how production systems handle legal complexity.