Last week, OpenAI launched a redesigned shopping experience in ChatGPT. 900 million weekly users can now browse products visually, compare options side-by-side, and get real-time pricing — all inside the conversation.
The protocol powering it is called ACP (Agentic Commerce Protocol), an open standard co-developed by OpenAI and Stripe under Apache 2.0. And the technical implementation is worth a closer look.
What ACP actually is
ACP defines a common language for AI agents and merchants to coordinate transactions. Think of it as an API contract between ChatGPT (the buyer's agent) and a merchant's product catalog.
Key architectural principle: OpenAI is not the merchant of record. Merchants retain full control over products, pricing, payments and fulfillment. ChatGPT is just the conversational intermediary.
The protocol lives on GitHub with 1,300+ stars and 192 forks:
github.com/agentic-commerce-protocol/agentic-commerce-protocol
The pivot nobody expected
OpenAI originally launched ACP in September 2025 with "Instant Checkout" — buy directly inside the chat. Etsy was the first integration.
It didn't work. Out of Shopify's millions of merchants, only about 12 activated checkout. Users browsed and compared, but went to the retailer's site to pay.
OpenAI acknowledged it in March 2026: "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide."
The new focus is product discovery: visual browsing, image-based search, comparison tables, budget filtering — then redirect to the merchant's site for checkout.
Technical integration: how it works
Feed specification
Merchants push product data to an OpenAI endpoint via encrypted HTTPS. The feed spec:
| Parameter | Value |
|---|---|
| Formats | CSV, TSV (recommended), XML, JSON |
| Encoding | UTF-8 |
| Max size | 10 GB |
| Compression | gzip |
| Update frequency | Every 15 minutes |
Required fields per product:
-
id— unique identifier -
title— max 150 chars -
description— max 5,000 chars -
price— with ISO 4217 currency code -
availability— in stock / out of stock / preorder -
image_url— at least one high-res image
Recommended: GTIN/UPC/MPN, reviews, rich media, shipping options, performance signals.
Three APIs
Feeds API → Upload/manage full product catalogs
Products API → Individual product upserts
Promotions API → Manage promotions (API-only)
For existing Stripe merchants, integration can take as little as one line of code.
Delegated payments
The payment flow uses single-use, time-bound, amount-restricted tokens. OpenAI prepares a delegated payment request; the payment service provider (Stripe, PayPal, Checkout.com) handles tokenization. PCI compliant by design.
Versioning
ACP uses date-based versioning (YYYY-MM-DD). The latest stable spec is 2026-01-30, which added extensions, discounts, and payment handlers.
spec/2025-09-29/ → Initial release
spec/2025-12-12/ → Fulfillment enhancements
spec/2026-01-16/ → Capability negotiation
spec/2026-01-30/ → Extensions, discounts, payment handlers (latest)
ACP vs. Google UCP vs. Amazon
Three ecosystems are forming:
ACP (OpenAI + Stripe): Open standard, Apache 2.0. Feeds update every 15 minutes. 900M weekly users. Focused on conversational product discovery.
UCP (Google): Open standard, launched January 2026. Backed by Shopify, Walmart, Visa, Mastercard. 50 billion indexed products. Covers the full purchase journey.
Amazon (Rufus/Alexa+): Closed ecosystem. 300M Rufus users, 60% higher conversion than standard Amazon flow. Blocks OpenAI crawlers. Removed 600M products from ChatGPT results.
The key difference between ACP and Google Shopping feeds: ACP feeds are designed for AI reasoning, not indexing. Every field can become an argument the agent uses to explain why a product is relevant — not just that it exists.
The data that matters
- Agentic traffic converts at 15-30% (Q1 2026 data) — 5-10x over traditional e-commerce
- AI-generated recommendations convert 4.4x better than traditional search (McKinsey)
- ChatGPT drives 20%+ of referral traffic to Walmart
- But: less than 0.2% of total e-commerce sessions come from ChatGPT
The volume is small. The conversion is extraordinary. The year-over-year growth is +805% (Adobe, Black Friday 2025).
What this means if you build e-commerce
If you're building or maintaining e-commerce systems, ACP changes the optimization target.
1. Feed quality > ad spend
Products surface based on what the AI can parse and verify. Incomplete feeds = invisible products. Every missing attribute is a lost recommendation opportunity. A product without a GTIN, without proper category depth (up to 5 levels), or with a generic "Great product, buy now!" description will never be recommended when a user asks for something specific.
2. Descriptions for NLP, not SEO
Attribute-rich descriptions that an AI can reason about beat keyword-stuffed copy. When someone asks "waterproof running shoes for flat feet under $120," the AI needs to find "waterproof," "arch support for flat feet," and the price in your product data. If those attributes are buried in marketing copy instead of structured fields, you lose.
3. Schema markup matters more than ever
Product, Offer, AggregateRating with real data. The AI cross-references your structured data against user queries. Complete schema is no longer optional — it's the foundation of agentic visibility.
4. Multi-protocol support
Prepare to maintain feeds for both ACP (OpenAI) and UCP (Google). Shopify already offers dual support. The formats are similar enough that a single product data pipeline can serve both, but the update frequencies differ (15 minutes for ACP vs. 24 hours default for UCP).
5. Attribution is fundamentally broken
The customer behavior that drives the sale is invisible to existing analytics. A user asks ChatGPT for a recommendation, gets three product suggestions, visits one site, and buys. Your analytics sees a direct visit or an unknown referrer. No impression. No click. No session.
Server-side webhook tracking from ACP/UCP is the path forward. Start building this infrastructure now — you'll need 18-24 months of data before it becomes reliable.
Getting started
- Merchant portal: chatgpt.com/merchants
- Developer docs: developers.openai.com/commerce
- Feed spec: developers.openai.com/commerce/specs/feed
- GitHub: github.com/agentic-commerce-protocol
- Stripe integration: docs.stripe.com/agentic-commerce
Currently US-only for merchant onboarding. Product discovery is available globally for all ChatGPT users (Free, Go, Plus, Pro).
We wrote a detailed analysis with market data, competitive comparisons, and practical recommendations:
zoopa.es/en/digital-marketing-en/agentic-commerce-protocol-acp-chatgpt-shopping/
What's your take on agentic commerce? Are you already preparing your product feeds for AI agents, or waiting to see how ACP vs. UCP plays out?
Top comments (0)