Agentic commerce is buying where an AI assistant researches, compares and in some cases completes a purchase for the shopper. An agent can only choose products it can understand. What decides the outcome is complete, structured, current product data: the same data behind your filters, site search and marketplace listings. It is a new channel on the foundation you already run, not a separate project.
What is agentic commerce?
Agentic commerce covers two different things, and it pays to keep them apart. The first is AI answers that recommend. A shopper asks ChatGPT, Gemini or AI Mode in Google Search for a quiet cordless drill under £150. They get a shortlist with reasons attached. The second is agents that act. They compare options, build a basket and, where the platform supports it, complete checkout once the shopper confirms.
For UK shoppers today, most of this is the first kind. The second is arriving through open protocols that connect AI platforms to merchant systems. Two of them matter most for UK businesses right now.
OpenAI and the Agentic Commerce Protocol. OpenAI launched Instant Checkout and the Agentic Commerce Protocol (ACP) in September 2025. ACP is an open standard OpenAI co-developed with Stripe. It passes an order from ChatGPT to the merchant’s own systems, and the merchant handles payment and fulfilment. In March 2026, OpenAI said the first version of Instant Checkout did not offer the flexibility it wanted. It shifted its focus to product discovery and now lets merchants use their own checkout. ACP has become the layer merchants use to share product feeds and promotions with ChatGPT. OpenAI’s developer guide says feed onboarding is currently open to approved partners, while Shopify merchants are covered through Shopify Catalog.
Google and the Universal Commerce Protocol. Google announced the Universal Commerce Protocol (UCP) in January 2026. It is an open-source standard developed with Shopify, Etsy, Wayfair, Target and Walmart, among others. It lets a shopper buy from a merchant inside AI Mode or the Gemini app, with the merchant staying the merchant of record. Google’s Merchant Center help says UCP-powered checkout currently applies to products eligible in the United States, Canada and Australia, for select merchants. Google’s May 2026 retail update says it will reach the UK later.
Alongside the protocol, Google has added conversational attributes to the Merchant Center product data specification. They are optional fields that help AI systems understand product detail. There are six: question_and_answer, document_link, related_product, item_group_title, variant_option and popularity_rank. Google says they can help products appear on AI-driven surfaces such as AI Mode, and that retailers globally can use them.
The names and the rollout dates will keep moving. The dependency will not. Both protocols keep the merchant as seller of record. Both run on product data the merchant supplies. Neither tidies that data for you.
How does an AI agent decide which product to choose?
An agent matches the shopper’s stated need to product facts. Those facts are specification, compatibility, use case, price, stock, delivery and returns. Take a request like this: “IP65 LED batten, 1500mm, under £40, delivered by Friday.” To put your product on the shortlist, the agent needs to confirm five things. It needs the ingress rating, the length, the price, the stock position and the delivery promise.
If the IP rating only appears in a line of description text saying “suitable for damp areas”, the agent cannot confirm it. If the length is stored as “1.5m” on one variant and “150cm” on another, it may not match either. The product is not penalised. It simply is not considered, and a competitor with cleaner data takes the slot.
OpenAI’s own help page on shopping with ChatGPT search says a product appears when ChatGPT judges it relevant to the shopper’s intent. It says merchants are ranked on factors like availability, price and quality. We cover the selection logic in more depth in how AI assistants choose which products to recommend. The short version: treat an agent as the most literal filter your catalogue will ever face.
What product data do AI agents need?
The fields are familiar. What changes is how complete and consistent they need to be, because no human is in the loop to guess. Our guide to AI-ready product data covers readiness scoring. These are the four areas we check first.
Complete technical attributes for each category
Every category has a set of facts buyers filter on. For a power cable that is conductor size, number of cores, voltage rating, sheath material and length. For a cordless drill it is voltage, chuck size, torque, battery platform and whether a battery is included. Those attributes need to exist for the category and be filled for every product in it.
Fill rate is the number to watch. A category where 90% of products carry torque can be compared. A category where 40% do cannot. OpenAI’s product feed specification includes fields for material, size, weight and dimensions. It specifies product dimensions, not package dimensions. Our piece on product attributes explains how to define the right set per category.
Clean identifiers: GTIN, MPN, brand
Identifiers let an agent work out that your listing, a marketplace listing and a manufacturer page describe the same product. Get them wrong and the agent either merges the wrong things or cannot trust any of them.
Google’s help page on unique product identifiers is blunt. Its advice is not to make up, guess or include values from similar products. Products submitted without identifiers they do have may get limited visibility. OpenAI’s spec makes brand a required field and asks for a GTIN only when it has a valid check digit. The faults we find most often are internal SKUs pasted into the GTIN field and supplier part numbers stored as the MPN.
Use case and compatibility stated plainly
“Fits all Bosch 18V Professional batteries” is a fact an agent can match. “Compatible with leading platforms” is not. Compatibility, application and suitability are often the deciding facts in trade and technical buying. They are also the facts most often buried halfway down a description.
OpenAI’s feed spec has no dedicated compatibility field, so these statements need to be explicit in your description and on the product page. Google’s question_and_answer and related_product attributes give you a structured place for them. Either way, hold compatibility as structured data in your PIM first, then write it out for each channel.
Current price, stock and delivery data in your feeds
An agent shopping for “delivered by Friday” needs your stock and delivery data to be right today. OpenAI’s spec tells merchants to update price and availability when they change. It also says an omitted shipping price means unknown, not free. A product with no delivery data loses to one that states it.
This is the same discipline as a well-run Google Shopping feed. Feed refresh frequency, price parity with the website and accurate stock status all carry straight across to agents.
Where complex catalogues fall short today
Across the more than 50 PIM projects our centre of excellence in Hyderabad has delivered, the same five gaps come up again and again.
- Supplier data loaded as received. Take a distributor with 40,000 SKUs from 300 suppliers. Each supplier names attributes differently, so “Colour”, “Finish” and “RAL” all describe the same thing. Nobody mapped them, so none of them filter.
- Specs trapped in PDFs. The datasheet holds the IP rating, the operating temperature and the certifications. The product record holds a link to the datasheet. Google’s document_link attribute can point an agent at the PDF, but the values themselves still need to be in the record.
- Attributes only in descriptions. A builders’ merchant writes good copy: “treated softwood, 47 x 100mm, 4.8m length, C24 grade”. All four facts are in a sentence, and none of them are in fields.
- Mixed units. Length as 1500, 1500mm, 150cm and 1.5m across one range. A person can read past it. An agent comparing lengths cannot.
- Half-empty categories. The attribute set exists, the fields are on the screen, and half of them are blank across a whole category.
None of these are new problems. They already hurt site search, filters and marketplace listings. Agents simply make the cost harder to ignore. Fixing them at volume is what our product data enrichment services are built to do.
Do you need a separate data strategy for agentic commerce?
No. Agentic commerce is a new channel on the same foundation. That foundation is one governed source, usually your PIM, feeding the website, Google, marketplaces and now agents.
The protocols add feed fields and connection points. They do not add a new kind of product data. Google’s conversational attributes are optional extras on the existing Merchant Center specification. OpenAI’s feed asks for titles, descriptions, identifiers, price, availability and attributes you should already hold. If those are right in your PIM, each new channel becomes a mapping and syndication job, not a rebuild.
The businesses that struggle are the ones running a separate spreadsheet per channel. Every new channel means another copy to keep in step, and the copies drift. If you do not yet have a governed source, start there. Our PIM implementation work is about making that source the one every channel reads from. SKULaunch, our own platform and the technology behind our delivery, works upstream of or alongside your PIM. It onboards and enriches supplier data before it reaches the PIM.
What to fix first: a 90 day plan
You do not need to fix the whole catalogue before agents matter. You need to fix the products that earn the money, then keep them right.
- Days 1 to 30: measure where you stand. Score attribute completeness, identifier accuracy and feed accuracy on a sample of your top sellers. Run AI visibility checks on the prompts your buyers actually use. Our guide on how to measure your AI visibility sets out the method. A free Snapshot gives you a fast baseline.
- Days 31 to 60: fix attributes in your top revenue categories. Pick the three to five categories that earn the most. Agree the attribute set buyers filter on, normalise units, pull values out of PDFs and descriptions, and fill the gaps.
- Days 61 to 90: correct feed accuracy. Clean identifiers, align price and stock with the website, and add delivery and returns data. Add Google’s conversational attributes through a supplemental data source, which is the route Google recommends.
- From day 90: set the routines. Name an owner per category. Set onboarding rules so new supplier data arrives mapped. Check freshness and completeness every month, then re-run your visibility checks and trace each gap back to the data.
The goal is products found, filtered and chosen on every channel, with less manual effort. Nobody can promise an agent will pick your product. You can make sure it is eligible, complete, accurate and current, and measure the difference.
Frequently asked questions
Buying where an AI agent researches options, compares them and can complete the purchase for a shopper. The shopper sets the limits and confirms the order.
Both, depending on the platform. Assistants read pages they can crawl and the structured feeds merchants submit, such as Google Merchant Center. Accurate feeds and clear product pages both matter.
No. What you can do is make products eligible, with complete attributes, correct identifiers and current price and stock. Then measure how often you appear and trace the gaps back to the data.
Usually not. Most of the work is in product data you already hold, governed in a PIM and sent to every channel. Agents add feed fields, not a new data strategy.
With the categories that earn the most. Check attribute completeness and feed accuracy there first, fix the gaps, then move category by category.
Get a free Product Data Performance Snapshot. We score 50 of your SKUs, show you where the gaps are, and send the results back within five working days.