Agentic commerce for retailers: what ACP and UCP need from your product data

Quick answer

Agentic commerce lets AI agents help people find, compare and buy products. ACP and UCP connect shopping services with retailers, but they depend on clear product information. OpenAI’s feed guidance is a useful starting point: identify the product, explain its suitability, keep the offer current and check which purchase routes your business can use.

What is agentic commerce?

Agentic commerce is an AI agent carrying out shopping tasks for someone, from researching products to placing an authorised order. Google’s shopping announcement describes buying with the shopper’s permission. How far an agent can go depends on the service: finding a product, building a basket and completing payment are different steps.

For our weekend shopper, the first job is suitability. The jacket needs to suit the weather and come in the right size before a discount matters. Next comes the offer: the selected colour, a current price, stock in that size and delivery before departure. Only then does checkout become useful.

For a retailer, that sequence gives the work an order too. Start by making the product understandable. Make the offer dependable. Then connect the purchase route your platform supports.

ACP vs UCP: how the two agentic commerce protocols compare

ACP and UCP are open standards that let agents and retailers exchange commerce information while keeping the retailer as merchant of record. Their specifications can support different services. For your store, what matters is the particular implementation: how products enter it, which purchases it supports and where the customer pays.

The comparison below separates the standards from their best-known uses. A protocol can support a capability before your chosen platform makes it available to your store.

Sources: ACP governance, OpenAI product data, Google onboarding and Google’s UCP FAQ.

The distinction is easier to see through the shopper’s journey. Shopify’s ChatGPT channel helps someone discover your jacket, then sends them to your store’s checkout in an in-app browser or a new tab. Google’s Universal Cart offers routes involving Google Pay or a transfer to the merchant’s basket.

OpenAI’s documents still need care: its merchant page says it’s moving away from standalone Instant Checkout, while its developer guide continues to describe approved partners. Confirm your available route rather than building around an assumed checkout experience. Our ChatGPT Shopping guide covers the feed formats and application steps.

The retailer announcements reflect that variety. OpenAI’s March 2026 update names Target, Sephora and Wayfair among its discovery integrations, alongside Walmart’s dedicated ChatGPT app. Google’s Universal Cart announcement includes Nike, Walmart and Shopify merchants across its purchase options. A retailer appearing in a shopping experience doesn’t mean every product has the same checkout route.

Whichever route you choose, the next problem is the same: give the service enough reliable information to assess your product.

What do AI shopping agents read from your product data?

For that jacket to reach a useful shortlist, an agent needs to understand both the product and the offer. It may draw on feeds and accessible pages, depending on the service. Google’s Search guidance says special schema isn’t required for its generative AI search features. Clear, consistent information is the work worth doing first.

There is still a minimum to get right. OpenAI’s file specification requires nine discovery fields: item_id, title, description, url, brand, seller_name, image_url, availability and price. Those establish the basic listing. The suitability information in the table explains why someone should choose it.

Notice how ordinary those requirements are. They’re the questions a good shop assistant would work through before recommending a product. The difference is that an AI service needs the answers in information it can retrieve and interpret.

A moisturiser described as “a daily essential” doesn’t tell someone whether it contains fragrance. A sofa’s width doesn’t establish whether its packaged dimensions fit through the doorway. And “available in blue” doesn’t confirm that the shopper’s size is in stock. These are the gaps to close when enriching product data, before adding more copy.

Google’s conversational attributes give that extra detail a more explicit home. The six optional fields are question_and_answer, document_link, related_product, item_group_title, variant_option and popularity_rank. They can carry practical answers, supporting documents and relationships between products and variants, without changing existing product approvals. Choose fields that help resolve an actual buying decision, and avoid repeating information already supplied elsewhere.

The same product knowledge also supports your Google Shopping Graph presence. That makes this catalogue work useful beyond whichever agentic checkout launches next.

How to make your catalogue ready for AI shopping agents

Start with a small group of products and follow them all the way to checkout. OpenAI’s upload guidance recommends testing a sample before the full catalogue. Choose best sellers your team knows well, including items that regularly generate customer questions, so you can tell whether the information is useful.

First, agree which record to trust. Compare the supplier data, feed and product page. If they disagree, decide which value is correct and fix the process that lets the others drift. Give each important attribute an owner who can resolve uncertainty, rather than asking the next export to repeat it.

Then improve the explanation. A title should identify the product, with the brand, type and attributes someone needs to choose. Keep the creative name, but explain what sits behind it. Complete the suitability fields and group variants correctly, so the selected size or finish leads to the right image, price and stock. Use assigned identifiers where available; never invent a missing GTIN.

Take a dress sold in three colours and several sizes. The fabric and care guidance may be shared, while images and stock differ by option. Keep those shared and variant-specific facts separate. OpenAI’s file feed groups variants through group_id, listing_has_variations and variant_dict; Google uses item_group_id. Choosing “navy, size 12” should consistently select the right offer.

Give the offer a reliable update process. OpenAI recommends a full file snapshot at least daily, using a stable filename. Its separate Products API supports partial updates. Work with your platform team to keep changing prices and stock current, and test withdrawals as well as additions. OpenAI can retain an item omitted from snapshots for up to 14 days; setting is_eligible_search=false makes it ineligible at the next processing run.

Finally, test what the service can reach. Check OAI-SearchBot and Googlebot access in robots.txt and your CDN. Follow the chosen product through the feed, page and checkout. Confirm that the purchasable quantity, delivery charge and final price agree. A page that works in your browser hasn’t necessarily passed those checks.

This is where merchandising and development need to work together. One team understands the product and the questions customers ask; another controls how that information travels. Both need to see the same journey before the pilot expands.

Tellis catalogue agents bring together supplier sheets, reviews and search data to fill missing attributes and identifiers, then write changes back to your store. On BuyWholeFoodsOnline’s optimised product pages, conversion rose 17% against a control group and Google AI visibility grew 124% month on month, as the case study sets out.

Judge your own work in the same practical way. Track appearances in AI answers, then follow them through to traffic and completed orders. For participating Google merchants, UCP analytics helps connect buy clicks with purchases and cancellations. Being mentioned is useful; knowing whether the customer could buy is more useful to the next decision.

Is agentic commerce available in the UK and EU?

Where your business is based and where the shopper lives are separate eligibility questions. Shopify’s ChatGPT guide allows stores outside the US if they sell to US customers. Its Google guide lists the US, Canada and Australia as eligible shopper markets, while warning that direct checkout is still rolling out. A UK merchant being eligible doesn’t establish domestic UK shopper access.

For the pilot, choose a market you already serve and confirm the route with your platform. Keep these distinctions in view:

Google’s announcements also describe different parts of the rollout. Its Merchant Center integration-hub guide, updated on 19 October 2026, puts the hub’s Canada and Australia expansion in early 2027. That’s an onboarding timetable, not confirmation of every checkout route. Check the actual feature you intend to use.

There’s no need to put catalogue improvements on hold while those dates settle. Clear product information, consistent offers and accessible pages are useful preparation whichever market you enter first.

What are the risks of agentic commerce for retailers?

Once the product has been chosen, the retailer still has to deliver on the offer. Google’s UCP guidance leaves merchants responsible for live inventory and shipping-address checks. For your pilot, that means testing the awkward cases too: a sold-out item, an excluded postcode, a changed price or a customer who needs help after ordering.

Check eligibility before promising a purchase experience for your whole range. Google’s checkout restrictions exclude products including personalised goods, pre-orders and freight deliveries. An engraved necklace or a sofa requiring freight delivery may therefore need a different purchase route from an ordinary stocked item.

There’s also a reason to keep the shopper in control. Gartner’s 27 May 2026 release reports that 54% of recent AI-shopping users double-checked all information, while a separate survey found willingness to delegate purchase decisions peaked at 11% in lower-stakes categories. The surveys covered 846 US consumers in November and December 2025 and 322 in January 2026, respectively. For your pilot, test whether the assistance makes choosing easier before asking customers to trust more automation.

Agree how payment permission, fraud checks, order updates and returns will work. Applicable product-safety and consumer rules still apply; an AI interface doesn’t take over those responsibilities. Build that operational review into the pilot, alongside the catalogue checks, so a successful recommendation can lead to an order your team can fulfil.

Find out what AI shopping agents see in your catalogue

Bring the work back to a product you know well: a jacket, a moisturiser or a sofa. Check whether someone could judge its suitability, understand the offer and buy it without discovering a contradiction. The free Tellis catalogue scan reads your product pages and helps identify where to begin. Use those findings to choose the first products to improve.

ACP and UCP can carry the information between systems. Your catalogue supplies the answers that make the journey worth taking.

Frequently asked questions