Agentic commerce readiness: product data, checkout and measurement

What changes when an agent joins the shopping journey?

Three announcements illustrate different ways an agent can participate. On 28 September 2026, Shopify added WebMCP checkout tools to its existing storefront and cart tools. They operate in the buyer’s browser session, can read and update checkout state, and submit checkout after buyer confirmation. When a buyer must complete an authentication step or resolve a blocking interface, control returns to the buyer.

On 1 October, OpenAI described Albertsons’ Safeway experience in ChatGPT: a request such as planning a dinner for four can lead to relevant groceries, offers and a cart, followed by checkout at Safeway. On 15 October, Google announced Universal Cart in the United States, bringing products from different retailers into a shopping experience with price and availability updates and checkout options. These are distinct implementations, not one universal route through which all merchants can sell.

Traditional ecommerce SEO often focuses on attracting a person to a product page and helping that visitor buy. With an agent involved, some comparison and filtering may happen before the person reaches the page. Imagine a buyer asking for black waterproof shoes in a 27-centimetre size, delivered before next week’s trip, for no more than NT$3,000 including shipping. The price is an illustrative Taiwan shopping budget, not a platform fee. The agent needs variant availability, specifications and delivery conditions to check the request.

If those fields are missing, an agent may be unable to establish suitability. If the checkout later changes the price or delivery date, the original recommendation may no longer meet the request. Whether the final purchase happens in an AI interface or on a merchant website depends on the service and supported integration.

Make the product feed describe the actual offer

Product feeds already matter for shopping ads and free product listings. Google Merchant Center lets merchants submit identifiers, titles, images, prices, availability and landing-page URLs. Agentic shopping makes the maintenance of these familiar fields more consequential; it does not make keyword research, technical SEO or useful product pages obsolete.

Google’s January 2026 UCP announcement also described additional Merchant Center attributes intended to support richer shopping information, including answers to product questions, compatible accessories and alternatives. A staged rollout is not the same as general availability for every account. The useful planning question is what information a shopping system needs, rather than whether a new field is guaranteed to improve recommendations.

For example, a laptop feed containing only the model and price cannot establish whether the computer supports two external monitors. A maintained specification page can explain supported displays, ports and compatibility limits. Feed approval establishes eligibility for the relevant listing product; it does not prove that an AI assistant has recommended the laptop.

Keep structured data, the feed and the visible page consistent

Google’s merchant-listing documentation explains how supported Product and Offer markup can describe products and commercial terms and establish eligibility for particular search appearances. That is a narrower claim than AI recommendation or agent checkout support.

A store might show shoes at NT$2,490 while the feed still contains NT$2,990 and the structured data contains a third price. These illustrative values show a consistency problem, not a measured platform response. Installing more schema does not resolve conflicting source data. Check the product page, feed and markup against the same offer, including size and colour variations.

Use the Rich Results Test to inspect supported markup, then check whether Google can access the page in Search Console. If the commerce platform already generates schema, inspect its output before adding another plugin with overlapping responsibilities.

Answer purchase questions with specifications, not adjectives

Search terms such as “waterproof shoes” can help organise a catalogue, but they do not capture every buying condition. A traveller may care about long-distance walking, wet weather, weight and delivery before departure. A product description saying only “high quality” or “comfortable” offers little evidence for those decisions.

Clear material information, waterproofing specifications, weight, limitations, compatibility, warranty, returns and delivery conditions make the offer easier to evaluate. Important conditions should be readable text on the product page, rather than available only in an image, a support conversation or an unrelated website.

This does not require a large FAQ on every product page or long passages written solely for an AI system. Start with real pre-purchase questions and check whether the page can answer them accurately. The overlap with GEO is useful information that an external system can interpret; completeness does not guarantee that the system will choose the brand.

Test the checkout beyond the product listing

Even a suitable product can fail at the transaction stage. Size selection, shipping cost, delivery address, payment authorisation and membership terms can change what is available. A listing at NT$2,800 is not an all-in offer if selecting the required variant raises the total beyond the user’s limit.

From authorization to order confirmation, the AX guide’s task loop provides four checks: discoverability, execution, verification and recovery after failure.

This remains a conversion-rate optimisation problem for human buyers, while adding a question about where an authorised agent can continue and where the buyer must take control. Do not remove payment confirmation, authentication or access controls merely to make automated interaction easier. Shopify’s documented handoff for required buyer input illustrates why those boundaries matter.

Test a complete supported journey: choose a variant, add it to the cart, calculate shipping and review the payment confirmation. A successful test of one integration is evidence for that setup, rather than proof of compatibility with every AI browser or shopping service.

Separate protocol support from merchant and country eligibility

Google developed Universal Commerce Protocol with commerce partners including Shopify, Etsy, Walmart, Target and Wayfair. A common technical specification can reduce the need for a different custom integration for every pairing of agent, commerce platform and payment service. A public specification and access to a consumer shopping entry point remain separate questions.

The Merchant Center UCP guide updated 19 October 2026 describes a gradual US rollout, products sold within the US and checkout for selected merchants. It also points merchants without the integration hub to a direct API route. That route does not itself establish admission to every Google AI shopping experience.

For a Taiwan merchant, Merchant Center support for Taiwan and New Taiwan dollars must not be treated as evidence that UCP checkout is generally available locally. For an English-language reader elsewhere, the same distinction applies to the reader’s own market: check the actual service, account, product and payment eligibility rather than importing US availability.

The source article also discusses the September 2026 Meta Muse example and reporting that Amazon blocked its shopping activity. The lesson is a permission boundary: browser capability does not establish that a platform permits every purchasing action. See the separate Meta Muse analysis for that dated example. Browser operation, a supported commerce API and a formal partner integration are different participation routes.

Four preparation tasks with observable completion criteria

1. Audit a small set of important products. Compare names, identifiers, currency, price, stock and specifications across the backend, feed and page. Include variants. Completion means the same purchasable offer is represented consistently, not merely that a feed upload succeeds.
2. Inspect Product and Offer output. Test the existing markup and access to the page. Correct inaccurate data before adding duplicate integrations. Completion means supported fields are valid and match the visible offer; recommendation remains a separate observation.
3. Fill decision-critical gaps. Add missing compatibility, usage limits, delivery, warranty and returns information. Completion means a buyer can find the conditions that change the purchase decision without having to infer them from promotional language.
4. Verify checkout and integration eligibility. Test the supported flow and ask the commerce provider about its actual agentic integration, regions, payments and permissions. Investigate accidental blocking of authorised services while keeping transaction safeguards. Completion means a documented supported path, including buyer handoffs and failure conditions.

These tasks can improve the accuracy of existing shopping information. They are not a claim that every completed checklist produces AI sales or that a merchant should invest in an unavailable checkout product.

Measure discovery, recommendation and orders separately

A buyer who follows a ChatGPT link and purchases on the merchant website may leave a measurable referral and transaction trail. GA4’s AI Assistant channel can identify eligible assistant-origin visits; it excludes Google AI Overviews and AI Mode. See the GA4 AI traffic guide for channel scope and reporting steps.

If an agent retrieves products through an API, or checkout happens entirely on another platform, the merchant’s website tag may never observe a session. Platform order records, supported order-source fields and documented attribution may then be necessary. Do not assign all orders with an unknown source to AI.

The practical question is whether your information supports a real purchase decision and whether you can observe the resulting business action. Platform adoption and transaction data must answer the larger commercial question. Making product data reliable is a verifiable starting point while those broader outcomes remain uncertain.