-
Step 1 of 6The shopper sets a task
I’m looking for whole-bean espresso coffee under CZK 400. It needs to arrive by Friday.
The shopper doesn’t type into the store’s search box but talks to an agent. The agent pulls the criteria out of the sentence: product type, brew method, price and delivery date.
What you control: Nothing. This happens outside your store, before the agent knows you exist.
Technical detail
An agent is a model with access to tools. The platform decides which tools: ChatGPT, Copilot, Gemini, Meta or an agent running in the browser. Store eligibility, product discovery and direct checkout have different requirements. Google uses Merchant Center; ChatGPT and other channels can use Shopify Catalog.
-
Step 2 of 6The agent searches the catalog
I found three espresso coffees under CZK 400. Let me check stock and delivery.
- Demo RoasteryBrazil Cerrado, 250 gCZK 329
- Another storeEthiopia Sidamo, 250 gCZK 389
- Another storeColombia Huila, 250 gCZK 359
In this example the agent searches structured product data. Shopify Catalog serves channels such as ChatGPT and Copilot; custom agents can use Shopify’s catalog interfaces.
For Google, product discovery is powered by Google Merchant Center, for example through the Google & YouTube channel. Discovery and direct checkout availability are separate.
What you control: Titles, categories, variants and product properties. The agent can’t filter on what isn’t in the data.
Technical detail
Shopify offers two catalogues: Global Catalog searches across Shopify stores, Storefront Catalog within a single store. Agents connect through MCP servers that follow the Universal Commerce Protocol (UCP).
search_catalog({
query: "whole bean espresso",
max_price: 400, currency: "CZK"
})
Simplified illustration, not a real API call.
-
Step 3 of 6The agent checks the details
Demo Roastery: in stock, delivered by Thursday, 14-day returns. I recommend it.
Demo Roastery doesn’t state a return policy, so I can’t confirm it. I recommend a coffee from another store.
Before recommending anything, the agent reads the product detail and facts about the store: shipping, returns, terms. In Agentic Storefronts the store’s knowledge base serves this too.
What you control: Completed store policies and the knowledge base. Shopify requires these policies for participation in the ChatGPT channel and for supported channels’ direct checkout. Missing policies do not automatically hide a site from every form of AI discovery.
Technical detail
A tool such as get_product returns the product detail. The store knowledge base (Fall ’26 Edition) collects the policies and FAQ answers the merchant approves, and Shopify suggests what to add.
get_product({ id: "brazil-cerrado-250" })
→ stock, shipping, returns
Simplified illustration, not a real API call.
-
Step 4 of 6The agent builds a cart
I’ve added Brazil Cerrado, 250 g. CZK 329 plus shipping.
This simulated purchase from Demo Roastery does not continue because its return policy could not be verified. Complete the store’s data first.
The cart lives on Shopify’s side, not in the agent’s head. Prices, discounts, taxes and shipping are calculated by the store’s own rules, as in any other purchase.
What you control: Your pricing, discount and shipping rules. Since 9 Oct 2026 Shopify Catalog also shows compare-at prices to agents; you can turn the sharing off in the admin.
Technical detail
Cart MCP builds and updates the cart, including lines, localisation and buyer context. Checkout MCP then turns the same cart into a checkout.
create_cart({
lines: [{ variant: "…-250g", qty: 1 }],
country: selectedCountry
})
Simplified illustration, not a real API call.
-
Step 5 of 6Payment
If the agent recommends this store, you complete the purchase in its checkout. We do not show direct checkout for this route. The walkthrough does not guarantee discovery in any country.
Direct checkout is available in this example. In practice it depends on channel and store eligibility, and Google is rolling it out gradually. Review the order and total before confirming.
A compatible agent can work with the checkout open in the browser. WebMCP is experimental; availability depends on the browser, agent and checkout. The buyer authorises the order.
This simulated purchase from Demo Roastery does not continue because its return policy could not be verified. Complete the store’s data first.
Where the shopper pays depends on the channel. ChatGPT sends them to the store’s checkout; Copilot, Google and Meta can take payment inside the conversation. A browser agent fills in the checkout in the open tab, but a person has to confirm the order.
What you control: Channel access under Sales channels > Agentic. For channels with direct checkout, you can turn it off. ChatGPT always refers shoppers to the store’s checkout.
Technical detail
Server-side agents use Checkout MCP. An agent running in the buyer’s browser can, since 28 Sep 2026, use Checkout WebMCP: the get_checkout, update_checkout and complete_checkout tools run inside checkout and submit only after the buyer confirms. 3-D Secure and similar steps always hand control back to the buyer.
WebMCP is still a proposed standard. Chrome is testing it in an origin trial from version 149, so this is an experiment, not a finished channel.
-
Step 6 of 6After the purchase
You confirmed the order in this simulation. A real purchase is followed by an order confirmation and shipping updates.
The order has not been confirmed. Return to Payment and confirm the purchase in this simulation.
This simulated purchase from Demo Roastery does not continue because its return policy could not be verified. Complete the store’s data first.
After a real purchase is completed, its order appears in the Shopify admin. An agent may then check delivery or returns, depending on its tools and permissions. This offline walkthrough creates no real order.
What you control: Fulfilment, tracking and returns. What the shopper experiences after the purchase is still yours.
Technical detail
Order MCP returns the current order state through get_order, and webhooks report fulfilment, returns, refunds and edits. According to Shopify, Shop MCP has no webhooks yet, so agents have to poll for status.
// Illustrative response for an already completed order
get_order({ id: "…" })
→ shipped, tracking number
Simplified illustration, not a real API call.