Preparing a Shopify store for an AI shopping agent

It helps to be unsentimental about the mechanism. An AI shopping agent is not browsing your site the way a person does. It is assembling a shortlist against a constraint the shopper stated — a running shoe under £120, in a wide fit, in stock in a 44, deliverable before Friday — and it will resolve that constraint from whatever source answers it most cheaply and most unambiguously.

That ordering matters. Structured data and feeds are cheap and unambiguous. Rendered HTML is more expensive and requires interpretation. A paragraph of brand copy that says the shoe "runs true to size for most feet" is not a fit signal an agent can filter on; a width attribute is. Whichever store expresses the constraint as data wins the shortlist, and the store with better prose loses it without ever knowing the comparison happened.

This is also why the work is less novel than the framing suggests. The same completeness that makes a product feed perform in Google Shopping, and the same attribute hygiene that makes on-site faceted search usable, is what makes a product legible to an agent. If your filters are bad, your agent readiness is bad, and you can see the first problem today without any new technology.

Product copy still matters, but its job has changed. It is no longer only persuading a reader; it is supplying the specific, liftable sentence that answers a question the structured data cannot express. "Fits a 15-inch laptop with a charger" is that kind of sentence. "Designed for the modern commuter" is not.

The practical test is whether a sentence survives being extracted alone. Name the subject, state the claim, attach the qualifier, and do not start with a pronoun pointing at the paragraph above. This is the same discipline that gets a page quoted by an assistant, and it costs nothing beyond attention.

Two content items are worth treating as product data rather than marketing: your returns policy and your delivery promise. Both are things an agent will be asked about on a shopper's behalf, both are usually written as legal prose on a page nobody links to, and both are answerable in two sentences if somebody bothers to write them that way. Shopify's Storefront MCP server exposes store policies and FAQs to assistants directly, which makes the quality of that writing an operational concern rather than a compliance one.

The place this tends to break down is production. Structured attributes, a clean spec block and an FAQ on every product template is a template change, not a copy change — and if every variation needs a developer, it will not happen at catalogue scale. A section library that lets merchandisers assemble spec blocks and FAQ sections themselves is the unglamorous prerequisite for all of this.

The genuinely new part is agents that complete a purchase rather than hand the shopper back to your site. The Agentic Commerce Protocol is the most developed public standard: an open specification published jointly by Stripe and OpenAI under Apache 2.0, describing programmatic commerce flows between buyers, agents and businesses.

It has three components worth knowing by name. A product feed specification, through which a business declares which products can be sold to agents and how they are presented. A checkout specification, a set of endpoints the business implements so an agent can complete a purchase, covering physical and digital goods, subscriptions and asynchronous purchases. And delegated payment, which passes payment credentials between agent and business through what Stripe calls a shared payment token, so the underlying credential is never exposed to the agent.

Two things in that model deserve emphasis, because they are the opposite of what people assume. The business remains merchant of record and retains control of fulfilment — this is not a marketplace where someone else owns the customer. And merchants control which agents may transact, with listing on any particular surface a separate application rather than an automatic consequence of implementing the spec.

Alongside it, Shopify documents a Storefront MCP server for the assistant-facing side: natural-language product discovery against a store's catalogue, cart creation and management through to checkout, store policies and FAQs, and order tracking and returns. That is the surface through which an assistant can genuinely shop a store rather than merely describe it.

Opinion, plainly: for most mid-market Shopify merchants, building bespoke agent endpoints in 2026 is premature, and the money is better spent on the data layer underneath them.

The reasoning is straightforward. Protocol implementations are a bet on a specific standard and a specific set of surfaces, both of which are moving. Clean variant-level data, honest availability, structured delivery and returns, and stable identifiers are a bet on nothing — they improve your feeds, your on-site search, your merchandising and your traditional search performance whether or not a single agent ever transacts with you. One of those investments has a floor and the other does not.

There is also an uncomfortable commercial question underneath the technical one. An agent that completes a purchase is an agent that owns the buying moment — no upsell, no email capture, no brand impression, and a customer whose relationship is with the assistant. That may still be worth it for a commodity SKU where you compete on price and availability. It is a poor trade for a brand whose margin comes from the experience, and "we should be everywhere" is not an analysis. Decide which of your catalogue you would actually want sold this way before you make it possible.

Where we would spend time now: making variant availability true, filling the recommended structured data properties, defining attributes as metafields, rewriting delivery and returns so they can be quoted in two sentences, and checking what the assistants say about you today. That last one is free and usually instructive.

• Shopify section library
• Getting found by AI assistants
• Structured data for variants and bundles