Agent-ready is not ready to buy. We asked our own store.

This year Shopify gave every store an interface for AI agents. At a fixed address, an agent can search products, build a cart, start a checkout and look up the purchase terms. Anyone running a Shopify store is technically agent-ready without lifting a finger.

In August and September, Search Engine Journal checked what that means in practice. The team tested 122 large Shopify stores on several layers: the agents.md file, the merchant profile for the Universal Commerce Protocol, the tools on offer, structured data, accessibility of product pages, live checkouts up to just before payment, and the purchase terms tool.

What works everywhere

Checkout. In 74 of 74 tests the interface created a checkout for an in-stock item and returned typed states and readable error codes. That is the platform's work, and it is done well.

Almost everywhere there is also an agents.md. 89 of 122 stores serve one. It rarely says anything about the store, though. 83 of those 89 files are near-identical copies of a template Shopify generates. Five stores had written their own.

What almost never works

The questions a buyer asks before buying. The purchase terms tool was tested at 88 stores with questions phrased the way a person would ask them:

On top of that, 51 of the 122 stores block the /policies/ path in robots.txt, mostly because of an old template. That is where the legal texts live. And not one store offered tools to cancel, change or return an order after purchase.

Product pages are no better. Of 68 stores checked, 47 percent had unlabeled input fields, 37 percent had empty buttons, and at 19 the add-to-cart button was missing from the delivered HTML. A person hardly notices. An agent that reads the page instead of looking at it finds no button with a name.

The study is a practitioner report from a consulting team, not an independent study. The figures describe large Shopify stores, not an average of all merchants.

The same questions to our pilot store

With naturbohne.de we run a Shopify store ourselves, an organic coffee roastery selling espresso machines and grinders. On 15 October we put the same three questions to the purchase terms tool, once in German and once in English.

In English, the returns question produced a sentence saying customers can start returns in their account. It named no deadline. A question about shipping costs returned the list of countries the store ships to, without prices. In German, the language the store sells in, all three questions returned an empty list.

Yet every answer is in the store. The withdrawal notice states 14 days, the shipping terms state 2 to 3 working days within Germany and 5 to 7 abroad. The tool just does not find them when you ask like a person.

Two things were better than in the test. Our agents.md is our own text and describes range, location and contacts. And robots.txt does block /policies/, but only for crawlers without their own rule. GPTBot, ClaudeBot and PerplexityBot have their own groups with us and may read the legal texts. Googlebot and generic agents may not.

What we are changing

We adjusted three things, none of which requires touching the platform. Since the evening of 15 October all three are live.

First, agents.md now has a section on returns, cancellation and delivery time, in German and English, pointing to the binding legal texts. An agent reading the file finds the deadline, the process and the delivery time without having to query a tool.

Second, /policies/ is now open to everyone in robots.txt. The legal texts are public; there is no reason to hide them from machines.

Third, return and cancellation rules are active where the platform expects them in structured form, not only as running text: 14 days from delivery, cancellation until shipment.

Then we asked the same questions again.

In English the tool now names the deadline, taken straight from the structured rules, along with return shipping at the customer's expense. It does not return the cancellation rule even though it is configured, and it knows no delivery time. In German it is still three empty lists. agents.md and robots.txt help an agent that reads the store. The tool does not draw on them, and it apparently does not understand German questions. For a German store that means: whoever asks in German gets no answer from the store itself, even with everything configured.

What follows

On Shopify, agent readiness is a default. Readiness to buy is not. The platform can build a checkout, but it cannot know how long a merchant accepts returns, whether it cancels and when it delivers. The merchant has to say so, in a way a machine can find.

Three checks are enough to start:

We are currently adding the first check to our checker at check.klariton.com. It asks the three questions and shows which go unanswered. It does not check whether an answer is correct, only whether one comes back at all. For most stores, that is the more important question today.

As of 15 October 2026. Sources: Search Engine Journal, test of 122 Shopify stores. Our own measurement of naturbohne.de on 15.10.2026 via the search_shop_policies_and_faqs tool at /api/mcp.