AI agents increasingly shop on their customers' behalf, and ACP and UCP define how that works. But which product the agent recommends is decided by product data.
The customer is king. Retail has built an entire century on that sentence: the shop window, the product photos, the landing page, the line in the newsletter that sparks desire at exactly the right moment. Everything was designed to win over a human who looks, reads and then decides.
The king is still king. But he no longer goes shopping himself.
More and more often, he sends an envoy ahead: an AI agent in ChatGPT, in Gemini or in Google AI Mode. The agent searches, compares, checks and finally presents the king with a recommendation. Or it simply buys. What this envoy is allowed to do and how it talks to a shop is now governed by two protocols: ACP and UCP.
The deciding question is no longer whether the king likes your product. It's whether his agent understands it.
A human buyer forgives a lot. If the waterproofing spec is missing, they look at the photo. If the shoe width only appears in a graphic, they zoom in. If the description sounds like advertising, they read between the lines.
An agent forgives none of that. It doesn't read images it can't interpret, and it doesn't trust promises it can't verify. It works with whatever your shop hands over as data: attributes, prices, stock levels, delivery times. If something is missing there, it doesn't exist.
That sounds like a detail for the IT department. In fact, it's a shift at the core of commerce. Until now, you wrote for the king. From now on, you also write for his agent.
For agents to shop at all, they need rules: How does an agent query a catalog? How does it build a cart? How does it pay without ever seeing the customer's card details? That's exactly what the two protocols describe. Think of them as court rules: the terms under which the envoy negotiates with retailers.
ACP – Agentic Commerce Protocol. OpenAI and Stripe introduced ACP as an open standard in September 2025, together with buying directly in ChatGPT. The protocol covers four things: the product feed, checkout, order updates and payment via a single-use token. The feed regularly supplies ChatGPT with identifiers, descriptions, prices, availability, images and shipping information.
UCP – Universal Commerce Protocol. Google introduced UCP in January 2026, developed together with Shopify, Etsy, Wayfair, Target and Walmart and backed from day one by more than 20 other companies from retail and payments. UCP covers the entire purchase, from discovery through checkout to post-purchase service. It works alongside other agent standards such as MCP and Google's payment protocol AP2, and it enables buying directly in AI Mode and the Gemini app.
Two protocols, two camps. For retailers, though, “ACP or UCP?” is a secondary question. Both need the same thing to start with: a catalog a machine can read.
OpenAI itself showed how quickly the court rules can change. In March 2026, the company scaled back buying directly in the chat: merchants are to use their own checkout again, while ChatGPT focuses on product discovery. Walmart, one of the first major partners, reported that purchases completed inside the chat converted at only a third of the rate of purchases after a click through to its own store.
You could read that as a setback. A more useful reading: checkout is the part platforms will keep rebuilding – sometimes in the chat, sometimes in the shop, sometimes in an app. The feed, on the other hand, has survived every version. Without product data, there's nothing to find, nothing to compare and nothing to buy.
The protocols change how things are bought. Your data decides what gets recommended.
Take a perfectly ordinary request. The customer tells their agent: “I need waterproof hiking boots, size 44, wide fit, under €160, delivered by Saturday.”
The agent finds three boots that could fit.
Shop A describes its boot as “braves any weather”. A nice line. There is no “waterproof” attribute. The agent can't verify the requirement and drops the boot.
Shop B has all the details, but the width only appears in a size chart image. The agent can't confirm “wide fit”. Dropped.
Shop C provides everything in structured form: waterproof membrane yes, width F, price €149, 14 pairs in stock, delivery in one to two days. The agent presents exactly this boot to the king.
Shop C wasn't the cheapest. It didn't have the best photo either. It was the only one whose data fully answered the question.
For agents, one quality matters that hardly anyone has measured so far: the agent readability of your product data. It describes how completely, unambiguously and currently a product is described in machine-readable form, so that an agent can compare and buy it without asking back.
In practice, that means:
None of this is new. Good product data quality has always mattered for marketplaces and search engines, as we also described in “If AI search doesn't understand your product, it won't recommend it”. What's new is that it now decides the purchase directly – and that nobody fills the gaps with their own eyes anymore.
Today, the agent still asks back and presents the king with a shortlist. In a few years, it will act on its own for many purchases: reordering supplies when they run low, comparing delivery time and total cost on every purchase, and avoiding retailers whose information has proven wrong in the past.
On the other side of the table, there will be agents too: those of retailers and brands, distributing catalogs, adjusting prices and answering questions in real time. Commerce becomes, in part, a conversation between machines.
In that conversation, trust is no longer a feeling but a data trail. Whoever delivers consistently correct data gets asked more often. Whoever delivers contradictory data drops out of the selection before a human ever looks.
The protocols make every retailer equally reachable. What will set them apart is the quality of their data.
This is exactly where AINAVIO comes in. As an Agentic PIM, AINAVIO makes sure product data doesn't just look good to people but is understandable to agents as well.
It starts with data onboarding: supplier data from Excel lists, PDFs or ERP exports is imported automatically and brought into a consistent structure. AI agents enrich the data, fill in missing attributes, harmonise values and flag where information is missing or contradictory. Finally, the data is published channel-ready – for marketplaces, shops and the feeds the new agent platforms draw on.
That's how a raw data record becomes a product an agent can understand, compare and recommend without asking back.
The customer is king – ACP and UCP don't change that. They only change who presents the products to him. The envoy is thorough, fast and incorruptible. It isn't impressed by beautiful words, only by complete data.
If you want to reach the king, you first have to convince his agent.
And the agent reads the fine print.
ACP from OpenAI and Stripe and UCP from Google define how AI agents shop. They don't define who gets bought from. Because the protocols give every retailer the same access, product data quality becomes the deciding difference: what an agent can't read and verify, it won't recommend.
Björn Thomsen is Head of Marketing at ainavio, the agentic PIM/DAM system. His focus is marketing operations & engineering: data, automation and the systems behind marketing, increasingly run by AI agents. He brings deep experience from B2B SaaS, cloud and data center companies as well as Tokyo's startup scene. At ainavio, he works on how e-commerce product data serves both people and machines, from the online shop to AI search.
https://www.linkedin.com/in/bjoern-thomsen/
contact@ainavio.com
+49 (0) 2842 - 929987-3
