For twenty years, ecommerce has been designed for human eyes: photography, urgency banners, checkout flows engineered click by click. In 2026 a second kind of customer is arriving — one that reads your product data instead of your photography, compares thirty options in a second, and can't be persuaded by a countdown timer. AI assistants now complete purchases on users' behalf, and the payments industry has spent the last year standardizing how.
If you sell online — or build for people who do — this is the guide to what's actually happening: the three protocols, what changed in the first year of agent checkout, and the pragmatic preparation worth doing now. (I build ecommerce AI systems; this landscape is where several client conversations now start.)
How an agent buys something
A full agentic purchase has three layers, and a different standard emerged for each:
Discovery and carts: UCP
The Universal Commerce Protocol, backed by Google and Shopify among others, gives agents a standard way to read catalogs, check availability and pricing, and assemble carts — machine-readable commerce, replacing the scraping and guesswork agents otherwise rely on. (I covered UCP alongside the broader protocol stack in the agent protocols guide.)
Checkout: ACP
The Agentic Commerce Protocol, developed by OpenAI with Stripe and published openly, defines checkout execution: the agent collects the buyer's payment selection inside the assistant, passes the merchant a narrowly scoped payment token (not raw card details), and the merchant charges it through their own payment provider. Crucially, the merchant remains the merchant of record — the assistant is a channel, not a middleman acquiring your customer.
This is the protocol behind ChatGPT's in-chat checkout, which rolled out with Etsy sellers and expanded across 2026 — with OpenAI adjusting the experience along the way (in-chat purchasing has been reshaped around commerce apps and partners as usage data came in). The volatility is the point to internalize: the channel is real, and its exact shape is still moving.
Payment authorization: AP2
The Agent Payments Protocol, initiated by Google and now governed under an industry body with dozens of payments partners (card networks, PSPs, banks), answers the question every merchant and issuer asks first: how do I know a human actually authorized this? AP2 uses signed, verifiable mandates — cryptographic records of the user's intent ("buy running shoes under $120"), the approved cart, and the payment authorization — creating an audit trail that settles disputes about what the user asked the agent to do. It's designed to work alongside ACP and existing payment rails.
The stack, in one line: UCP finds and carts it, ACP checks out, AP2 proves a human said yes.
What agent buyers change for merchants
- Structured data becomes your storefront. Agents rank products on machine-readable attributes, price, availability, policies, and reviews. A catalog with accurate schema.org markup, complete attributes, and honest availability data is now conversion optimization — for a buyer that never sees your hero image.
- Persuasion mechanics stop working; trust signals start. Urgency banners, exit popups, and dark patterns don't move an agent — clear return policies, transparent pricing (agents notice fees appearing at checkout), and consistent data do.
- Your AI-assistant presence is a sales channel. How ChatGPT, Gemini, and Claude describe your products when asked for recommendations is already influencing purchases upstream of any protocol. This is the commerce face of the same visibility work as being citable in AI answers generally.
- New fraud and support questions. Which orders came from agents? What happens when a customer says "my agent bought the wrong thing"? AP2's mandates exist precisely for this, but merchant policies and support flows need to catch up.
What's worth doing now — and what isn't
Worth doing in 2026, for most merchants:
- Get product data agent-grade: complete structured markup, accurate feeds, clean availability and pricing. This pays off in ordinary SEO and AI-assistant visibility even before protocol traffic matters.
- Make checkout API-reachable. If completing a purchase requires your UI, you're invisible to ACP-style flows. Shopify and the major platforms are shipping this layer for their merchants; custom stores need it designed.
- Monitor your AI-assistant presence: ask the major assistants what they recommend in your category and how they describe your products; treat wrong answers as a data-quality bug to fix at the source.
- Decide your agent policy deliberately — welcome agent buyers, require verification, or gate them — rather than discovering your stance during a dispute.
Not worth doing yet, for most: deep custom protocol implementations ahead of your platform's support, or re-platforming for a channel whose volume in your category is still small. The protocols are young, the UX is being redesigned in public, and early-mover advantage here mostly accrues to categories where agent buying is already natural — digital goods, replenishment, commodity comparison shopping.
The conversational middle ground
Between "human clicks through your store" and "autonomous agent completes checkout" sits the layer that's already mainstream: conversational shopping on your own site — an assistant that understands needs, remembers preferences, and presents the right products. Gem AI does exactly this for Amazon product discovery, and EmbedChat AI gives merchants a store-integrated assistant with per-merchant knowledge and live product tools. That layer is buildable today, converts today, and produces the structured, conversational product data that agent channels will consume tomorrow.
If you're deciding how your store should meet AI buyers — or want the conversational layer first — that's a conversation worth having this year, not after the protocols settle.