Agentic Commerce: When an AI Agent Is the One Buying From You
Straight answer: agentic commerce is selling where discovery, choice and increasingly payment all happen inside an AI interface — ChatGPT, Google's AI surfaces, an assistant — without the customer ever opening your site. Through 2025-2026 it moved from demo to infrastructure: ACP (OpenAI + Stripe) and the Universal Commerce Protocol (Google + Shopify, announced at NRF in January 2026) are open standards already live. Being present doesn't require rebuilding your site: it requires clean, complete product feeds, structured data and APIs. A messy catalog simply doesn't exist in this channel.
TL;DR
- The channel is real: Shopify reports orders from AI-powered search up 15x year over year through 2025.
- 45% of consumers already use AI somewhere in the buying journey (IBM IBV, January 2026). Morgan Stanley expects agents to capture 10-20% of US e-commerce by 2030.
- Two open protocols: ACP (OpenAI/Stripe, checkout inside ChatGPT) and UCP (Google/Shopify, discovery via
/.well-known/ucp, transports REST, MCP, A2A). - The lever isn't content — it's data hygiene. Fresh feeds, valid
Product/Offerschema, real prices and availability. - On Shopify you're one toggle away. On a custom platform, it's a development project.
What actually changed
For twenty years the game was getting people to your site. First with ten blue links, then with GEO — becoming the source the AI cites. Agentic commerce takes the next, uncomfortable step: the user never arrives. They ask the assistant "find me a waterproof hiking jacket under €200, delivered by Friday", and the assistant compares, picks and in some cases completes the purchase.
At that moment you have no homepage, no banner, no product page with beautiful photography. You have a row in a feed and APIs that respond. All your brand work collapses into one question: is your data good enough to get you selected?
The two protocols that matter
| ACP — Agentic Commerce Protocol | UCP — Universal Commerce Protocol | |
|---|---|---|
| Who | OpenAI + Stripe | Google + Shopify (announced at NRF, January 2026) |
| Where the user buys | Inside ChatGPT (Instant Checkout, live since September 2025) | Inside Google's AI surfaces, native or merchant iframe checkout |
| Data foundation | Dedicated product feed (.jsonl.gz, .csv.gz, .xml.gz) |
Google Merchant Center, 170+ attributes available |
| Discovery | Merchant onboarding | /.well-known/ucp with declared capabilities |
| Transports | REST | REST, MCP, A2A, embedded |
| Payment | Stripe Shared Payment Token: the merchant never sees the card | Native or embedded checkout |
| Cost | Order commission (public sources point to around 4%) on top of processing fees | Being defined |
The interesting part, for someone who works with agents: UCP uses MCP as one of its transports. The protocol becoming the standard for connecting agents to business systems is also a rail for commerce. That's no coincidence — more on that here.
What you need in order (by priority)
1. The product feed, updated daily
This is item one and nobody takes it seriously enough. ACP asks for compressed files with mandatory fields: item_id, title (max 150 chars), description (max 5,000), price with ISO 4217 currency, availability, image URLs, brand, eligibility flags for search and checkout. Daily updates accepted. UCP starts from Merchant Center and rewards whoever fills in as many of the 170+ attributes as possible.
In practice: if your feed has truncated titles, supplier-copied descriptions, prices out of sync with the site, or stock updated "when we remember", you're invisible in the agentic channel. Worse: if the feed price isn't the real one, checkout fails on a customer who had already decided to buy. That's the most expensive kind of error there is.
2. Real structured data on product pages
Valid Product, Offer, AggregateRating schema, rendered server-side (many AI crawlers don't execute JavaScript). Price and availability in the schema must match what's on screen. This isn't old-school SEO: it's how a machine understands what you sell without trusting your layout.
3. Catalog and order APIs that hold up
To reach agentic checkout you need stable endpoints: create checkout session, update (shipping, variants), retrieve state, complete, cancel — with complete, consistent JSON responses. If your APIs answer in three seconds today and change shape every release, this channel will break on you.
4. Machine-readable policies
Shipping, returns, delivery times, availability by region. If an agent has to promise "delivered by Friday" and your policy lives only in a prose HTML page, the agent will pick the competitor who exposes that fact in a structured way.
What it costs in practice
Almost entirely a function of where you're hosted.
- Shopify: near-zero effort. Shopify stores are included in both protocols' programs, and for ACP you enable the sales channel from admin.
- WooCommerce and mainstream platforms: configuration work on feeds and schema, plus plugins. Days, not months.
- Custom or headless platform: a real project. Feed endpoints, checkout API, delegated payment integration. Weeks of development, plus data maintenance that never ends.
My take: this channel punishes sloppiness
I have a fairly firm opinion here. Agentic commerce doesn't reward the best copywriter, the prettiest brand, or the biggest ad budget. It rewards whoever has their data in order — which is exactly why I like it.
For years marketing could paper over back-office mess with good communication: incomplete catalogs, approximate stock levels, badly written product pages. An AI agent isn't moved by a great campaign: it reads fields, compares, and if a field is missing it drops you without telling you. It's the most meritocratic and most merciless channel we've had in a while.
The good news: the work required is the same work you already needed for GEO, for your Google Shopping feed, and for making an AI assistant on your store actually useful — clean, structured, true data. Do it once, use it across four channels.
FAQ
Do I need to join these protocols right now?
Not necessarily the protocols — but yes, right now, the data hygiene. Complete and fresh feed, valid structured data, consistent prices and availability: those are prerequisites for both protocols and you need them anyway for Google Shopping and AI search. Formal agentic-checkout onboarding can wait until you have numbers.
Does agentic commerce steal my customer relationship?
Partly yes, and that's the serious reason many brands are cautious: if the transaction happens inside ChatGPT, data and post-sale relationship pass through an intermediary. It's the marketplace trade-off again. The sensible answer isn't to boycott the channel but to hold a position in it while you keep building the direct one.
How much is it worth in Europe today?
Little in absolute terms, a lot in growth. The strong numbers are still US-centric, but most assistants Europeans use are the same ones. The right time to fix your data is while the channel is small: doing it in a rush when it matters is like redoing your SEO at Christmas.
Isn't a chatbot on my site enough?
Different things. A site chatbot converts people who already reached you; agentic commerce is about people who never will. You need both, and they share one foundation: a catalog a machine can read without ambiguity.
If you want to know how ready your catalog is for this channel, let's run a check together. It's the same groundwork that makes ShopBrain work well on your store — clean data first, everything else after.
Related articles
- Chatbots and the AI Act: What Really Lands on 2 August 2026From 2 August 2026 Article 50 of the EU AI Act applies: your chatbot must disclose it is an AI and generated content must be marked. Fines up to €15M or 3% of global turnover.
- llms.txt Does (Almost) Nothing: The Data, and Why I Recommended It97% of sites with a valid llms.txt got zero requests for the file. Google says no Search system reads it. Here is what to do instead — and who actually does use it.
- HTTP QUERY: The First New Method in 16 Years (RFC 10008)In June 2026 the IETF published RFC 10008: QUERY, an HTTP method that is safe and idempotent yet carries a body. First new verb since PATCH in 2010 — but do not ship it tomorrow.