Agentic commerce is a purchase made by an AI agent on a person's behalf. To be chosen, a store needs complete Schema.org product data, real-time stock and price through an API, machine-readable delivery and returns policies, and a checkout agents can use through protocols such as UCP (Google and Shopify) or ACP (OpenAI and Stripe).
Key takeaways
Agentic commerce rewards stores whose catalogue a machine can read, compare and buy from without help.
- Agentic commerce is a purchase made by an AI agent on a person's behalf: it takes a goal in plain language, queries merchants, weighs options against the user's constraints, authorises payment within set limits and triggers fulfilment.
- Two open standards are competing: UCP (Universal Commerce Protocol, launched by Google in January 2026 and co-developed with Shopify, Etsy, Wayfair, Target and Walmart) and ACP (Agentic Commerce Protocol, from OpenAI and Stripe, September 2025). Both accept MCP as a way to connect agents to the merchant.
- Checkout inside the assistant is still settling: in March 2026 OpenAI let merchants use their own checkout from ChatGPT. The operational question is not whether agents arrive, but whether your catalogue will be machine-readable when they do.
- What to do today is not exotic: structured product data, real-time stock and price via API, explicit policies and machine-accessible checkout. Almost all of it is data hygiene that also improves classic SEO.
What is actually changing
Agentic commerce changes who reads your product page first: an AI agent filters before the human sees anything. For twenty years, e-commerce has been optimised so that a person finds, compares and buys. Convincing photos, persuasive copy, urgency, social proof. Agentic commerce introduces a new intermediary: an agent that reads, compares and decides before the human sees anything.
The user no longer says "I'm looking for trainers". They tell their assistant: "I need trail running shoes in size 43, waterproof, arriving before Friday and under this budget". The agent queries catalogues, verifies real stock and delivery windows, discards what doesn't qualify and presents two or three options. Or buys directly, if authorised.
What this means for a merchant: your photos and copy still matter for the human who confirms. But the first filter is run by a machine that needs structured, verifiable, real-time data. If your sizing isn't normalised, your stock isn't queryable or your delivery window isn't explicit, the agent discards you before anyone sees your product.
What agentic commerce is NOT: it is not a chatbot on your site helping people buy (that is conversational assistance, covered in AI agents for e-commerce). Agentic commerce involves an external agent, belonging to the buyer, operating on your catalogue from outside.
The protocols: UCP, ACP and MCP
UCP and ACP are competing commerce protocols for agents, and MCP is a connection standard both accept. There is legitimate confusion here because the terms appear together and do different things.
UCP (Universal Commerce Protocol). An open-source standard that Google announced on 11 January 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart (Google for Developers). It lets agents interact with merchant catalogues, retrieve real-time pricing and inventory, manage carts and complete purchases through a single normalised interface. Its specification supports REST, MCP and A2A, and Stripe joined Google and Shopify on its governing council in April 2026.
ACP (Agentic Commerce Protocol). Co-developed by OpenAI and Stripe and open-sourced on 29 September 2025 under the Apache 2.0 licence (OpenAI, Stripe). It launched with Instant Checkout in ChatGPT for US Etsy sellers, with Shopify merchants announced next. It covers the same functional territory from the OpenAI ecosystem, with a focus on the payment flow, and merchants can publish their checkout through an API or through MCP (agenticcommerce.dev).
MCP (Model Context Protocol). Not a commerce protocol: it is the standard that connects an agent to tools and data such as inventory, pricing and product details, and both UCP and ACP accept it as a way to connect. We covered it in detail in MCP for business.
AP2 (Agent Payments Protocol). Google's open protocol for authorising payments made by agents, announced on 16 September 2025 with more than 60 partners, which works as an extension of A2A and MCP (Google Cloud).
| Protocol | What it solves | Backers |
|---|---|---|
| MCP | Connecting agents to tools and data; accepted by UCP and ACP | Anthropic, now the Agentic AI Foundation (Linux Foundation) |
| UCP | Catalogue, cart and checkout for agents | Google and Shopify, co-developed with Etsy, Wayfair, Target and Walmart |
| ACP | Agentic checkout and payment | OpenAI, Stripe |
| AP2 | Authorising agent-led payments | Google, with more than 60 partners |
Our recommendation on the standards war: don't bet on one. The expensive part of the work (normalising the catalogue, exposing stock and price via API, making policies explicit) is common to all of them. Do that part and adopting any protocol becomes a thin layer on top. Picking a side now takes on risk you don't need to take.
Key market data
Agentic shopping already works at scale, but checkout inside the assistant is still settling.
- In March 2026 OpenAI changed course on Instant Checkout: "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide", so merchants can now use their own checkout while OpenAI focuses on product discovery. Shopify says "millions of Shopify merchants are open for business in ChatGPT", with purchases completed on the merchant's store through an in-app browser (OpenAI, 24 March 2026).
- UCP "already enables direct checkout for hundreds of thousands of brands and retailers across Google" (Google, 16 September 2026).
- Traffic to US retail sites from generative AI tools grew 693.4% year on year in the 2025 holiday season, from 1 November to 31 December (Adobe, January 2026).
- "Half of all consumers now use AI when searching the internet", according to McKinsey (October 2025).
Our reading: European markets lag the US on agentic shopping adoption, and that is a temporary advantage. It gives room to prepare the catalogue without the pressure of losing sales today. That room closes fast.
What to do today: five concrete fronts
Preparing a store for AI agents comes down to five fronts, and almost all of them also improve classic SEO. It is not a speculative investment: the same work improves human conversion.
1. Structured, complete product data. Schema.org Product with offers, availability, price, sku, gtin, brand and normalised technical attributes. An agent does not interpret "standard size" or "assorted colours": it needs discrete, comparable values. It is the same structured data discipline Google has asked for for years, now with more immediate consequences.
2. Real-time, queryable stock and price. If your availability updates nightly, the agent works with stale data and cancels orders. Availability must be queryable at the moment, not a static field.
3. Explicit delivery windows, by postcode. "Delivery in 24-48h" is not usable data if it depends on destination and cut-off time. The agent filters by real arrival date: one of the most frequent user constraints and one of the fastest disqualifiers.
4. Machine-readable policies. Returns, warranty, shipping costs, conditions. An agent that cannot verify the returns policy before purchase will lean towards the competitor that exposes it clearly.
5. Accessible checkout. This is where UCP and ACP come in. While you decide, at minimum make sure your purchase flow does not depend on steps only a human can clear without friction.
Common use cases
The three most common agentic commerce projects are catalogue normalisation, real-time availability and structured policies.
Case 1: Catalogue normalisation for agent readability.
- Problem: inconsistent product attributes across categories, sizes written three different ways, materials in free text.
- Solution: a normalised attribute taxonomy plus AI-assisted extraction to fill gaps in the historical catalogue, with sampled human validation.
- Stack: extraction model + attribute schema + batch validation + Schema.org output.
- Result: the catalogue becomes machine-comparable, and faceted navigation improves for humans too.
Case 2: Exposing real-time stock and delivery windows.
- Problem: published availability updates in batches and causes post-purchase cancellations.
- Solution: an availability and delivery-window endpoint by SKU and postcode, fed from the ERP and the carrier.
- Stack: availability API + ERP integration + zone-based window rules.
- Result: what the agent sees matches reality, and out-of-stock cancellations fall.
Case 3: Structured, verifiable policies.
- Problem: returns and warranty conditions live as free text spread across several pages.
- Solution: modelling policies as structured data and as queryable functions, with one source feeding both the site and the agent.
- Stack: policy schema + Schema.org + query endpoint.
- Result: a single source of truth for humans, agents and the support team.
How to prepare, step by step
Preparing for agentic commerce takes seven steps, and the catalogue comes before any protocol.
- Audit your catalogue from a machine's perspective. Take ten products and ask whether an agent could filter by size, material, compatibility and delivery date using only what you publish. It is usually a revealing exercise.
- Normalise the attribute taxonomy before touching protocols. It is the longest piece of work and it conditions everything else. And it serves UCP, ACP or neither equally.
- Complete your product Schema.org.
availability,price,priceValidUntil,sku,gtin,shippingDetailsandhasMerchantReturnPolicyare the fields that separate being readable from being ignored. Google now also accepts shipping and returns policies declared once for the whole business, underOrganization. - Expose availability and delivery windows via real-time API. If you do only one thing on this list, make it this one. It prevents the most disqualifications.
- Structure your returns, warranty and shipping policies as data, not just as text pages.
- Evaluate UCP and ACP once your base is ready, not before. On a normalised catalogue, adopting a protocol is a matter of weeks; on a chaotic one, it is an entire project.
- Measure agent traffic separately. Start identifying automated assistant access in your logs. Without that segmentation you will not know when the channel starts to matter.
Common mistakes (and how to avoid them)
The most common agentic commerce mistake is waiting for the standards to settle before preparing the catalogue.
Mistake: waiting for "the standards to settle". → Reality: 90% of the work (normalising the catalogue, exposing stock, structuring policies) is common to every protocol. Waiting only delays the slow part.
Mistake: betting everything on one protocol. → Reality: UCP and ACP compete and both have serious backing. Build the agnostic base and treat each protocol integration as a thin, replaceable layer.
Mistake: blocking agents in robots.txt without analysis. → Reality: there is a difference between a training crawler and an agent buying on behalf of a customer. Blanket blocking can close a sales channel. Distinguish by user agent and decide deliberately.
Mistake: publishing stale stock. → Reality: an agent buying on outdated data generates cancellations, and platforms penalise unreliable merchants. Real-time data is not a luxury, it is the entry requirement.
Mistake: assuming persuasive copy stops mattering. → Reality: the agent filters on data, but in many flows the human confirms. Copy and images still decide between the two or three options the agent presents. What changes is a prior filter that does not read emotional text.
Mistake: treating it as a marketing project. → Reality: it is a product data and integration project. The team that must lead it is the one managing the catalogue and the ERP, not the content team.
Realistic timelines and effort
Catalogue normalisation is the long block, 4 to 10 weeks; a protocol integration then takes weeks.
- Agent-readability audit of the catalogue: 1-2 weeks. It is cheap and determines the real scope of everything else.
- Attribute taxonomy normalisation: 4-10 weeks depending on catalogue size and disorder. This is the long block, and it can be parallelised by category.
- Complete product Schema.org: 2-4 weeks once attributes are normalised.
- Real-time availability and delivery API: 3-6 weeks, depending on how accessible the ERP is.
- UCP or ACP integration: weeks, not months, if the base is done. Without the base, it is an entire project.
- Where the hidden effort sits: in the legacy catalogue. New products get created under the right taxonomy; the ten thousand old ones need cleaning, and that is where AI-assisted extraction saves weeks of manual work.
Metrics to track from day 1: share of products with complete, normalised attributes, Schema.org coverage, latency and accuracy of the availability endpoint, out-of-stock cancellation rate and (once you can segment it) traffic and conversion coming from agents. That last metric does not exist in most analytics setups today: building it now is what will tell you when the channel takes off.
Frequently asked questions
What is agentic commerce?
It is a purchase made by an AI agent on a person's behalf: the agent takes a goal in plain language, queries merchant catalogues, compares options against the user's constraints, authorises payment within set limits and triggers delivery. It runs on open protocols such as UCP and ACP.
What is the difference between UCP, ACP and MCP?
UCP (launched by Google with Shopify and other retailers) and ACP (OpenAI and Stripe) are commerce protocols letting agents access catalogues, manage carts and complete purchases. MCP is the standard that connects agents to tools and data such as inventory, pricing and product details, and both protocols accept it as a way to connect.
What do I need to do in my store so agents can buy?
Structured, normalised product data, real-time queryable stock and price, explicit delivery windows by destination, machine-readable returns and warranty policies, and an accessible checkout. Most of that work also improves your SEO and human conversion.
Do I have to choose between UCP and ACP?
Not today. The bulk of the work (normalising the catalogue, exposing availability, structuring policies) is common to both. Build the base first and treat each protocol integration as a thin, replaceable layer.
Should I block AI agents in robots.txt?
Distinguish carefully. A crawler collecting data to train models and an agent buying on behalf of a customer are different things, even though both are automated. Blanket blocking can close a sales channel. Decide by user agent and on commercial grounds.
Do product copy and photos stop mattering?
No. The agent filters on structured data, but in most flows the human confirms among the presented options, and there copy and images still decide. What changes is a prior filter that reads only data, and you don't reach that second stage without passing it.
When will this matter outside the US?
Adoption is furthest along in the US, where agentic checkout started: Instant Checkout in ChatGPT launched with US Etsy sellers in September 2025. That lag is a temporary advantage for preparing your catalogue without urgency, but it closes fast.
Does this replace traditional SEO?
It does not replace it, it extends it. Structured data, accurate product information and explicit policies were already SEO signals. The difference is that now they also determine whether an agent can consider you at all.
Could an AI agent buy from your store today?
NAXIA prepares catalogues for agentic commerce, starting with an audit of what an agent finds today. The work covers the machine-readability audit, attribute normalisation with AI-assisted extraction over the legacy catalogue, complete Schema.org, real-time availability and delivery endpoints, and structured policies. Agnostic base first, protocol second.
If you want to know what an agent would find in your catalogue today (and what's missing), talk to us. No commitment, no 40-slide decks.
Or explore our implementation process first.
NAXIA is a Spanish artificial intelligence and automation agency for companies of all sizes: it designs, deploys and maintains AI agents, voice agents and automations integrated with each client's systems. Founded in 2024 and based in Pamplona (Navarre), it works with companies across Spain and has delivered more than 20 projects.