Agentic commerce
for a Shopify catalogue.

Agentic commerce — an AI system finding, comparing and buying products on a person’s behalf — reaches a Shopify merchant through three kinds of thing: a platform programme such as Shopify’s own sales channel, a product feed a shopping surface ingests, and a protocol for the transaction itself. They overlap, they change quickly, and their owners describe them in different terms. This guide keeps them apart, and it is dated: every page cited was read on 2026-09-11.

2025-12-10 — Winter ’26 Edition

Shopify introduced Agentic Storefronts as a way for merchants to sell in AI conversations, set up once in the Shopify admin. It said Shopify Catalog “infers categories, extracts attributes, consolidates variants” and keeps price and inventory current, and it listed ChatGPT, Perplexity and Microsoft Copilot as available platforms, with others “coming soon”.

2026-06-18 — How agentic commerce works

“Agentic Storefronts is a sales channel in the Shopify admin that connects a merchant's products to AI shopping platforms.”Shopify — How agentic commerce works (2026-06-18)

The June post gives status by platform. ChatGPT: “Live for millions of Shopify merchants who sell to US buyers, regardless of store location.” Microsoft Copilot: live for Catalog merchants selling to US buyers. AI Mode in Google Search and the Gemini app: “Available to select brands selling to US buyers with native checkout powered by UCP, with broader rollout underway.” Perplexity, listed in December, is not mentioned in the June post. We do not infer why.

2026-06-17 — Spring ’26 Edition for developers

Shopify described UCP as “The standard Shopify co-developed with Google”, positioned its Catalog API as the discovery layer, and said developers can build agentic experiences without approval through the Developer Dashboard. In its June 2026 material Shopify defines UCP as “an open standard that defines how AI agents transact with merchants—covering cart creation, checkout, payment, and post-purchase—across any platform.”

What this means for a catalogue

In Rank Sniper’s reading, not Shopify’s: Catalog works from what you have published. Inferring a category or extracting an attribute is inference; a product type you set, an attribute you state and variants you have named are facts nothing has to infer. The product-data guide covers each field.

Product feed specification

Feeds are accepted as JSONL, or as Google-compatible TSV or CSV, gzipped if you like. Required fields: id, title (up to 150 characters), description (up to 5,000), link, image_link, availability (in_stock, out_of_stock, preorder or backorder), price, brand, and gtin or mpn. Three flags control where a product may appear: is_eligible_search (true by default), is_eligible_checkout and is_ads_eligible.

Agentic Checkout specification

OpenAI describes it as letting merchants “run end-to-end checkout flows inside ChatGPT while keeping orders, payments, and compliance on their existing commerce stack”. It defines five REST endpoints on /checkout_sessions. The page states no status and no dates, so this guide does not describe who can use it or since when.

ACP

agenticcommerce.dev describes ACP as “An open standard for programmatic commerce flows between buyers, AI agents, and businesses”, developed by Stripe and OpenAI, and “open source and community-designed under the Apache 2.0 license”. The page carries no version or date.

Instant Checkout

OpenAI’s announcement and help pages for Instant Checkout could not be read when this guide was written — both returned an access error — so its current status is not confirmed here on an OpenAI page. CNBC reported on 2026-03-24 that OpenAI was revamping the shopping experience in ChatGPT after Instant Checkout. Treat that as a news report, not documentation, and check OpenAI’s own pages before relying on either.

Four documented user agents

OpenAI’s crawler page lists four: OAI-SearchBot, GPTBot, ChatGPT-User and OAI-AdsBot, and notes that “Each setting is independent of the others.” It lists no separate token for an agent product; any other name in your logs claiming to be OpenAI is not one the documentation gives.

2026-01-11 — the announcement

“A new open standard for agentic commerce that works across the entire shopping journey — from discovery and buying to post-purchase support.”Google — agentic commerce tools and protocol (2026-01-11)

Google named Shopify, Etsy, Wayfair, Target and Walmart as co-developers and listed more than twenty endorsers, among them Adyen, American Express, Best Buy, Flipkart, Macy’s, Mastercard, Stripe, The Home Depot, Visa and Zalando. It said UCP “will soon power a new checkout feature on eligible Google product listings in AI Mode in Search and the Gemini app” for eligible U.S. retailers.

Merchant Center — the checkout feature

Google’s Merchant Center help page describes a checkout button on eligible listings in AI Mode and Gemini, with the merchant remaining the seller of record. It lists eligibility in the US, Canada and Australia, says the feature is “available for select merchants at this time” through an early access programme with an interest form, and requires the native_commerce(checkout_eligibility) attribute. It gives no launch date. The January announcement said U.S. retailers; the help page, read in September 2026, lists three countries. Both are given as their pages state them.

ucp.dev

The protocol’s own site states an Apache 2.0 licence, REST and JSON-RPC, and support for AP2, A2A and MCP. It carries “Copyright 2026 UCP Authors” and no version or date.

Google’s shopping crawler

Storebot-Google crawls for “All surfaces of Google Shopping” and, like Google’s other common crawlers, honours robots.txt. It is not on the list Rank Sniper checks; the crawler index shows which agents are.

Policy access belongs on the list too. Google’s merchant-listing markup recommends shippingDetails and hasMerchantReturnPolicy, and an agent that cannot find your shipping regions or returns window cannot answer the questions a shopper asks before buying. Rank Sniper’s composite gives shipping details or a return policy in product markup 20 of its 100-point schema sub-score. The structured-data guide shows where each of these fields goes in the markup.

OpenAI’s own sentence explains the reasoning several vendors share: “Because these actions are initiated by a user, robots.txt rules may not apply.” A user-initiated fetch is a person using a tool to read your public page, and several vendors treat it that way. Anthropic is the exception among those documented here: its article says its bots honour robots.txt and states no exception for Claude-User.

Two practical conclusions. First, robots.txt expresses a preference to crawlers; RFC 9309 itself says the protocol “is not a substitute for valid content security measures.” What a user-initiated fetcher receives is your public page — the same raw HTML any reader gets — so the structure of that page decides whether the agent can answer. Second, Rank Sniper’s scan reads a rule for ChatGPT-User like any other AI agent on its list, and counts a full block toward the capped penalty — 6 points each, capped at 30 — even though OpenAI says such rules may not apply to it. The score reflects the policy you wrote, not what the fetcher does with it. The other fetchers above are not checked.

Protocol participation is not on Rank Sniper’s release table at all — not as Planned, not as anything. If that changes, it will appear there first. Nothing on this page claims that a better score gets a product shown, chosen or bought by any agent. For AI shopping as the journal has covered it, see the invisible-channel field note.