Shopify agentic commerce: a merchant checklist for channels, product data and measurement

For a Shopify merchant, preparing for AI shopping starts with three checks: which channels can access your products, whether the product information answers real buying questions, and how you will measure the resulting orders. Buying another AI tool is not the first step.

This checklist was checked against Shopify’s documentation on 2 October 2026. Discovery, direct checkout and a custom shopping agent are different capabilities; a product appearing in an AI answer does not establish that every buyer can purchase inside that conversation.

1. Review access before changing it

Open Sales channels → Agentic. Shopify’s management guide says its default “Allow Shopify to manage for me” setting enables available channels and automatic enrolment in future ones. Turning it off lets you manage channel access and supported direct checkout individually.

Record the current settings before making changes. Decide who owns channel reviews and what should trigger another check: a new market, a changed product range or a new checkout option.

Removing Shopify Catalog access can take up to seven days to stop data sharing. It does not prevent discovery through public web crawling. Treat channel controls and search visibility as separate decisions.

2. Audit a small set of important products

Shopify Catalog uses product titles, descriptions, options, images, prices and availability. Google and Meta also have their own sales-channel feed routes. Check the data source for the channel you actually use.

Our recommendation: choose five products with frequent pre-purchase questions. For each, check dimensions, materials, compatibility, what is included and the differences between variants. Put limitations beside benefits. Replace vague descriptions with details you can substantiate; do not invent attributes to satisfy an imagined ranking formula.

Shopify’s Catalog search preview can help identify listing gaps. Its results are directional: AI channels may re-rank them. Test realistic buying questions, record incorrect or incomplete answers, and fix the underlying product information before drawing conclusions about visibility.

3. Keep developer tools in perspective

Global Catalog MCP supports discovery across merchants; Storefront Catalog MCP searches one merchant’s catalogue. These are tools for building agents, not a requirement to commission a custom integration before reviewing your existing channels.

Some Global Catalog fields are inferred by Shopify and may vary in accuracy. They should not be mistaken for your original product copy. For the browser checkout side, see our Shopify WebMCP explanation.

4. Measure orders and returns separately

Shopify’s agentic reporting includes sales, orders, online-store sessions and conversion. Sales aggregate referred orders and direct checkout; the documentation says these analytics are not yet available for headless stores. Keep that distinction in mind when comparing a sales total with website sessions.

Adobe’s September 2026 release reports a July survey of 5,000 consumers: among respondents who had used AI for online shopping, 77% felt more confident and 69% said they were less likely to return a purchase. That is self-reported sentiment, not a measured reduction in returns or a causal result.

Build your own comparison from completed orders. Allow the normal return window to pass, compare similar products and markets, and record how many orders sit behind each percentage. A handful of AI-referred purchases is useful evidence to investigate, not a reliable forecast for the whole store.

15 October update: test the answers behind the checkout

Bill Hunt’s audit of 122 leading Shopify stores, published by Search Engine Journal on 14 October, adds a practical reason to test policy answers. All 74 live checkout tests successfully created, read and cancelled a checkout. No purchases were completed. Yet none of the 88 policy tools tested gave a usable answer to the combined question about shipping cost and arrival in Boston; only two answered the cancellation question usefully.

This was a selected group of leading stores, probed in August and September, with shipping tested to one Boston address. Different checks covered different subsets. It is evidence of gaps in that sample, not a failure rate for every Shopify store or a measurement of lost sales.

Shopify’s current policy-tool documentation describes a question field and optional context, such as the product being viewed. Its example response includes an answer and source links. Ask your developer or agency to compare those returned answers with the policies your customers actually see.

Our suggested review record is simple: the exact question, product and destination; the expected answer approved by your operations team; the returned answer and sources; and the date checked. Include a normal purchase and an exception, such as a discounted item. Record an unavailable answer as unresolved rather than treating it as confirmation of a policy.

Give each discrepancy an owner. A copy error belongs with the policy editor; an answer that contradicts correct published information needs investigation with the integration provider. Recheck after the correction. Do not promise delivery dates or cancellation rights that your fulfilment team cannot honour.

Background: Shopify’s launch demonstration

This official Shopify video was published on 10 December 2025 for Winter ’26. It explains the original concept; use the current documentation above for availability and settings.

Watch Shopify’s original video on YouTube.

18 October update: interpret the ChatGPT attribution gap cautiously

On 15 October, Adobe’s Cedric Huesler reported that ChatGPT’s link tags were returning after a bug affecting attribution from 8–15 October. He said tagged clicks in his telemetry had recovered to roughly their previous level. This is an analyst’s observation, not a guarantee for every merchant: Glenn Gabe reported in the same discussion that he was not yet seeing the change across his accounts.

Annotate that period in your analytics, then test the ChatGPT links and landing-page parameters you actually receive. Compare tagged sessions, Direct traffic and orders separately. A simultaneous rise in Direct is a clue, not proof that all missing attributed visits moved there or that demand stayed unchanged. Missing UTM parameters and a missing referrer are separate signals; neither identifies an otherwise unattributed visit on its own.