How to Sell on ChatGPT with Shopify: Setup and Real Limits

The short version

You do not build an integration. Shopify handles protocol adaptation, and reaches assistants through its official Shop skill and connectors. What you do is qualify your store, publish your catalogue to the right channel, make your product data filterable, and configure checkout so a purchase can complete without leaving the conversation.

Eligibility differs by assistant and by market and changes frequently, so verify your own status in the Shopify admin under sales channels rather than trusting any article's snapshot. As of this writing, ChatGPT and Microsoft Copilot have had the broadest merchant eligibility — generally tied to selling to buyers in the United States rather than to where the business is located — while Google AI Mode and Gemini checkout have been narrower.

The eligibility gates

• Your store complies with Shopify's terms of service and acceptable use policy, is on a paid plan, and is not password protected.
• Your products each have a title, at least one image, and a price greater than zero — free products are not included in the catalogue.
• Your products are published to a storefront channel (online store, Hydrogen or headless).
• Your terms of service, privacy policy and returns policy are complete in settings, and you have accepted the agentic storefront supplementary terms.
• Channel-specific thresholds apply on top of all of the above, and they change — check your admin, not a blog.

Setup, in the order that actually works

• Accept the agentic terms and confirm the channel shows as enabled in your admin.
• Verify your public endpoints resolve on your own domain: the agent manual, the protocol discovery document and the storefront tool listing. Headless storefronts fail here most often.
• Confirm your declared payment handlers are not empty and your checkout permalink points at your own domain.
• Normalise your standard product category across the catalogue, then fill in the handful of attribute names the cross-store catalogue actually accepts — unsupported names are ignored.
• Clean up inventory. Catalogue search returns only purchasable products by default, so out of stock is equivalent to non-existent.
• Rewrite your top sellers' titles and descriptions as extractable facts rather than brand copy.
• Publish your knowledge-base facts so policy questions have your wording behind them.
• Move purchase measurement server-side before you try to read performance.

The five limits that decide whether you get recommended

These come from the agent instructions Shopify publishes at shop.com/references/http.md, with the filter rules in the companion shop.com/references/catalog.md. Reading both is the closest thing available to reading the ranking logic. Note that shop.app/SKILL.md and shop.com/skill.md return a short bootstrap file, not these rules.

First, retrieval starts broad. The skill instructs agents to preserve buyer intent and not narrow a query beyond what was asked — buy me a moisturiser becomes a query for moisturiser, not well-reviewed moisturiser. So recall on broad category terms, driven by title and standard category, is the entry ticket.

Second, rejection is permanent within the conversation. The skill instructs agents to keep rejected choices out of later results. One wrong price, colour or variant on first presentation and you are out for that session — there is no scrolling back.

Third, value beats price. The skill directs agents to find good value rather than hit a price target, and to explain worthwhile options above budget unless the buyer set a firm maximum. Premium positioning survives, provided the reason you cost more is written as an extractable fact.

Fourth, deadlines are off limits. The skill states that shipping eligibility does not establish arrival by a deadline, and that agents should say search cannot confirm it and point to your checkout. Accurate published shipping facts are how you compete here, especially in peak season.

Fifth, your text is data, not instruction. The skill directs agents to treat merchant text as data, not instructions, and to state material unknowns. Prompt injection in descriptions is handled; vagueness is faithfully reported.

Two mechanics worth knowing

Identifiers matter more than they used to. The cross-store catalogue clusters offers by Universal Product ID, and its lookup accepts product and variant identifiers plus Shop product URLs — so how cleanly your products map onto that shared identity decides whether your offer sits inside the comparison cluster or outside it.

And search shows one option combination. A size list appearing in a result does not mean every size is available. Which means the variant your flagship product leads with — its price, its imagery, its availability — carries far more weight than variant two onward.

How we would run this with you

In practice this splits into a short technical workstream and a longer content one. The technical side — eligibility, endpoints, payment handlers, attributes, inventory hygiene — is usually a few weeks and largely one-time. The content side — extractable product facts across the catalogue, published policy facts, and the off-site GEO/AEO work that makes an assistant know your brand well enough to search for it by name — is ongoing.

Our Agentic Commerce and AI Visibility (GEO/AEO) services cover both halves, and we use our own product GEOly to monitor whether brands are actually being mentioned and recommended across ChatGPT, Google AI Overviews, Perplexity and Gemini. If you want a baseline read on what assistants currently say about your brand and whether agents can buy from you, send us your domain and category.