EN
Language · same page NLNederlands/agentic-commerce-google-ai-koopt-namens-je-klant/ ENEnglish (UK)/en/agentic-commerce-google-ai-buys-on-behalf-of-your-customer/ ESEspañol/es/comercio-agentico-google-ia-compra-en-nombre-de-tu-cliente/ We do not remember your choice and never redirect you automatically.
DOC.K · KNOWLEDGE BASE

Agentic commerce: Google AI buys on behalf of your customer. Can it read your site?

Agentic commerce means that an AI agent takes over the buying process: searching, comparing, choosing and, increasingly, paying on behalf of your customer. In January 2026, Google announced the Universal Commerce Protocol for this together with Shopify, the Universal Cart followed at Marketing Live, and the Agent Payments Protocol has provided a payment standard with more than sixty partners since September 2025.

At the same time, the market learned a hard lesson: in March 2026, OpenAI pulled the plug on checkout inside ChatGPT. This article sets out what has actually been announced, what the figures from Adobe and Salesforce really say, and what you can sort out today so an agent can read your offer. Because that is the thread running through it: the agent does not skip the best shop, but it does skip the unreadable one.

Download the agentic commerce check as a skill ↓3.7 KB · zip with SKILL.md · MIT licence · no account needed · runs through the four layers of agent readiness, from catalogue to transaction, and delivers a prioritised action list with no promised results
CLAUDE CODE
  1. Unzip it into ~/.claude/skills/ (or .claude/skills/ in your project).
  2. Claude then recognises the skill by itself as soon as you ask whether your site is ready for AI agents.
  3. You can also call it directly, with /agentic-commerce-check.
CLAUDE.AI
  1. Go to Customize and then Skills.
  2. Upload the zip there as a skill.
  3. Or paste the contents of SKILL.md into the project instructions of a Project.
CODEX
  1. Open AGENTS.md in your repo.
  2. Paste the contents of SKILL.md into it, or place SKILL.md alongside it as a separate file and refer to it from AGENTS.md.
  3. Codex reads that along at every session.
Gianluca, who wrote this explanation
image · whoever does the work also writes the explanation
Section 01 · The buying route

From instruction to order, with no website visit

reg. A.001

This is what a purchase looks like when an agent carries it out. Notice what gets read at every step: not homepages and banners, but feeds, structured data and terms. The shop no longer decides for itself whether it takes part in the comparison; its data decides that.

FIG.01: The agentic buying route in four stepssheet 1/4
agent-log · from instruction to orderwhat the agent reads at each step
$ customer: "sort running shoes, size 9, budget £100, delivered by the weekend"
STEP 01 · SEARCHThe agent searches catalogues through AI Mode, Gemini or another assistant. It reads product feeds and structured data, not homepages. Whoever is not in there does not exist for this search.
STEP 02 · COMPAREPrice, stock, delivery time, shipping costs, returns policy and reviews are placed side by side. Machine-readable counts, unreadable is out, however good the offer is.
STEP 03 · CHOOSEAt Google, the choice lands in the Universal Cart: a basket that works across shops, in Search and Gemini. The customer sees the selection and decides.
STEP 04 · CHECKOUTThrough the Universal Commerce Protocol and Google Pay, within limits the customer signs in advance (Agent Payments Protocol), or the customer checks out on the shop's own site.
The selection happens in step 01 and 02. Everything that comes after that is execution. That is why machine-readability is not a technical detail but the ticket into the comparison.
3A.01 · open protocols carry agentic commerce: UCP, AP2 and ACP
60A.02 · launch partners Google announced AP2 with in September 2025
693A.03 · per cent growth in AI traffic to US retail sites in the 2025 holiday season, year on year
20A.04 · per cent of global online holiday sales in 2025 where AI helped decide
A.01 and A.02 are properties of the protocols themselves. A.03 comes from publicly published data from Adobe Analytics, with the link in the text below. A.04 comes from Salesforce's holiday report; we have not checked that figure against the source ourselves, so it stands on Salesforce's authority and not on ours. None of the four is a client result of ours.
DOC.01

What is agentic commerce, and why is it tipping now

Imagine your customer no longer scrolls through Google themselves. No ten tabs to compare prices, no hesitating at the checkout. Instead, that customer tells an AI assistant: find a good plumber in Bristol who can come this week, or sort running shoes within a budget. The assistant does the rest: searching, comparing, choosing and, where that is already possible, paying. That model is called agentic commerce, and it is the logical next step after AI Overviews and zero-click search. First AI took over the search results page, now it is moving into the whole journey from question to order.

The difference with earlier waves of hype: this time there are working standards, real partners and measurable flows

Google announced the Universal Commerce Protocol in January 2026 and built visible products around it at Marketing Live 2026, from a basket that works across shops to an agent checkout with Google Pay. OpenAI had already tried checkout inside ChatGPT earlier and partly backed away from it in March 2026, a lesson that is at least as instructive as the announcements. Meanwhile, the flow of visitors landing on websites through AI assistants is growing fast, with traffic that, according to Adobe Analytics, converts better than average.

For you as a business owner, the core question is therefore not whether this continues, but which part already applies to you today

The honest answer comes back throughout this article: checkout by agents is still young and changes every quarter, but being read by agents is already decisive today. An assistant that cannot read your offer cannot recommend you either. Whoever has their technical foundation NL and structured data in order takes part in every next step of this shift; whoever does not have that misses all of them at once.

DOC.02

What Google has actually announced

On 11 January 2026, Google presented the Universal Commerce Protocol, UCP for short, at the NRF retail trade show, developed together with Shopify. The idea: a shared, open language that lets AI agents and shop systems understand each other across the whole buying journey, from browsing the catalogue to basket and checkout, so a shop does not have to build a separate connection for every agent.

Alongside Shopify, the launch featured Etsy, Wayfair, Target and Walmart on stage, with backing from payment players such as Stripe, Adyen, Visa, Mastercard and American Express. In April 2026, the standard's governance grew further: Amazon, Microsoft, Meta and Salesforce also joined the technical council. The fact that a competitor like Amazon joins a standard that Google started says a great deal about how seriously the market takes this.

At Marketing Live 2026, Google turned it into visible products

The Universal Cart is a basket that works across shops and travels with the customer through Search and Gemini: the customer collects products from different online shops in one basket and pays with Google Pay, or moves the basket to the shop's own site to complete it there. In the US version of AI Mode, adverts and offers can get a direct checkout, without the user leaving the AI environment.

And with Business Agent, retailers can activate their own brand assistant in Merchant Center; retailers such as Lowe's, Reebok and Poshmark are already using it

Google also opened dozens of new data fields in Merchant Center, meant for conversational questions: which accessories go with this, what is the alternative if this is sold out.

The third building block is the Agent Payments Protocol, AP2, which Google already launched in September 2025 with more than sixty partners, including Mastercard, PayPal and American Express. AP2 handles the most sensitive part: paying. The customer signs digital mandates for intent, basket and payment in advance, so it can be proven afterwards what the agent was allowed to do and what the customer approved. Google has since placed the protocol with the FIDO Alliance, the organisation behind passkeys, so that it does not remain one vendor's house standard.

For the Netherlands, a sober footnote applies to all of this: the rollout of these features starts in the United States, and what lands here, and when, has not been announced. The underlying requirement does not change because of that, because being read by assistants is already happening here today, as we described earlier in the shifts in ChatGPT NL.

Section 02 · The standards

Three protocols, three roles

reg. A.002

The names look alike and get mixed up in every article. This is the split to remember: UCP handles the buying journey, AP2 handles paying, ACP handles discoverability in ChatGPT. All three are open standards, and none of the three asks you, as a shop, to implement a protocol yourself: your platform and your payment provider do that. Your work sits one layer deeper, in the data that these protocols pass on.

FIG.02: The three protocols side by sidesheet 2/4
protocols.logwho, what and status
UCPUniversal Commerce ProtocolGoogle and Shopify, announced 11 January 2026 at NRF. Catalogue, basket and checkout in one language. Backed by Target, Walmart, Etsy, Stripe, Visa and Mastercard among others; since April 2026 also Amazon, Microsoft, Meta and Salesforce on the technical council.rolling out, US first
AP2Agent Payments ProtocolGoogle with more than 60 partners, September 2025, since placed with the FIDO Alliance. Signed mandates for intent, basket and payment, so it is fixed what the agent was allowed to do.open standard
ACPAgentic Commerce ProtocolOpenAI and Stripe, September 2025. Initially also handled checkout in ChatGPT; since March 2026 the focus has been on product discovery, after which the customer checks out on the shop's site.course changed
What this means for you: you do not have to build a protocol. If you run on Shopify or a large platform, UCP will come to you. Your homework is the layer underneath: the feed, the structured data and the terms that all these standards draw on.
DOC.03

The lesson from OpenAI: why checkout in the chat stalled

Whoever only reads the announcements misses half the story. In September 2025, OpenAI launched Instant Checkout together with Stripe: buying without leaving ChatGPT, carried by the Agentic Commerce Protocol. Walmart put around two hundred thousand products into the assistant, Etsy and Shopify joined in. A quarter and a half later, in March 2026, OpenAI pulled the plug on that first setup. Of the millions of Shopify shops, only a few dozen had actually connected checkout, and the figures Walmart shared through CNBC explain why: checkout inside the chat converted roughly three times worse than clicking through to the shop's own site. Missing basic functions, such as handling US sales tax, did not help either.

More interesting than the stopping is what remained

That same Walmart found that ChatGPT delivered roughly twice as many new customers as classic search traffic. So the demand side does work: people let an assistant advise them and discover shops they would otherwise never have found. It was only the checkout itself that they preferred to do somewhere they trust, the shop's own site. OpenAI then shifted course to what the sector now calls "discover in AI, buy on the site": the Agentic Commerce Protocol now mainly serves discoverability, and retailers such as Target, Sephora, Lowe's and Wayfair use it to make their catalogue readable for ChatGPT, after which the customer checks out on their own site.

For you, this is where the most important strategic lesson of this whole topic lies

The transaction layer is still in motion: Google is now trying it with a broader coalition and a basket that can also flow through to the shop's own site, and whether that catches on remains to be seen.

But the discoverability layer is the same in every variant: a catalogue or service offering that machines can read

Invest in that, and it works for Google, for ChatGPT and for every assistant still to come. Invest only in a checkout connection, and you are betting on a horse whose race has not yet been run. How you build pages that convince both people and machines is something we also describe in the structure of a product page that actually sells.

DOC.04

What the figures really say, with the source included

Plenty of figures about agentic commerce circulate without a source, so only figures with a name attached are given here. Adobe Analytics, which measures traffic to US retail sites, reported growth of 693 per cent in traffic that came from AI assistants to retail sites over the November and December 2025 holiday season, compared with a year earlier (Adobe on AI traffic by sector). The growth figure of 393 per cent for the first quarter of 2026 that is cited elsewhere does not appear in any Adobe publication we could find, so we do not mention it.

That figure of 693 per cent sounds huge, and it is, but do not forget the context: the flow starts small, so tripling or more is easier now than it will be later.

More interesting is the quality of that traffic: in March 2026, visitors from AI sources converted 42 per cent better than other visitors according to Adobe, stayed 48 per cent longer on the site and viewed 13 per cent more pages (Adobe AI Traffic Report, second quarter 2026). Someone arriving through an assistant is therefore not a casual passer-by but someone who has already been advised and pre-sorted.

Salesforce looked at the same holiday period from the sales side and estimated that AI and agents helped decide more than 20 per cent of global online retail sales, worth around 262 billion dollars in spending. A striking detail from the same report: retailers running their own brand assistant grew by 6.2 per cent over the holidays, against 3.9 per cent for retailers without one. That is a correlation, not proof of cause, because whoever leads on agents usually also leads elsewhere. But the pattern fits what Adobe measures: the buying journey is shifting to places where an assistant has a say.

What these figures do not say is part of the picture too

They do not prove that checkout by agents is already big; it is not, as the OpenAI episode above shows. Nor do they prove that every business will get customers from AI Mode tomorrow; the rollout differs by country and by sector. What they do show: the flow of visitors pre-sorted by an assistant is growing fast and is more valuable than average traffic. Every month, we use the National AI Index NL to measure which Dutch businesses are named by AI assistants for fixed customer questions per sector, and the differences in that are exactly what you would expect: whoever is readable gets named, whoever is not does not exist in those answers.

Section 03 · The reading layer

Four layers the agent reads from your site

reg. A.003

All protocols and baskets ultimately draw on the same four layers of your site. This is the order in which an agent needs them, and therefore also the order in which you put them in order. The first two apply to everyone, even without an online shop.

FIG.03: The four reading layers of agent readinesssheet 3/4
LAYER 01 · CATALOGUE Can the agent find your offer? Online shop: an up-to-date product feed in Google Merchant Center, with accurate price and stock. Service provider: every service its own named page with service area and offer.
LAYER 02 · STRUCTURED DATA Does the agent understand what you sell? JSON-LD that matches what is on the page: Product and Offer for online shops, LocalBusiness and Service for service providers, Organization for everyone.
LAYER 03 · TERMS Does the agent know your rules? Shipping costs, delivery time, returns policy and warranty as readable text on the site, not only in a PDF. The agent that cannot read this chooses the shop where it can.
LAYER 04 · TRANSACTION How close is checkout? A fast checkout with no mandatory account, and a platform that supports or is going to support UCP. For service providers: a route that can be booked or requested directly.
The highlighted layers 01 and 02 already work today, for every assistant and every search engine. Layer 03 strengthens your position in every comparison, layer 04 is preparatory work for the rollout of agent checkout.
DOC.05

Machine-readable or invisible: structured data as the ticket in

For years, structured data was something for enthusiasts: nice for star ratings in the search results, but you could rank perfectly well without it. That optionality disappears in an agentic world, because an agent chooses based on data it can compare: price, availability, reviews, location, opening hours, payment options. If that information is not on your site in a machine-readable form, the agent cannot weigh you in, however good you are at your trade. It is the same mechanism we have seen for years at SEO for online shops NL, but without the human visitor as a safety net who still comes to look for themselves.

The tool for this already exists and is called JSON-LD: JavaScript Object Notation for Linked Data

In practice, it is a block of code on your page, invisible to visitors, that tells machines what is on that page: your company name, address and opening hours, your products with price and stock status, your services with service area. Google has documented JSON-LD for years as the preferred format for structured data, and it is deliberately easier to implement than older alternatives such as Microdata: you do not have to rebuild your existing HTML, you add a block. Whoever goes from nothing to complete markup makes their pages machine-readable for the first time; how much visibility that delivers differs per site, and we do not promise a figure.

Where it goes wrong in practice: many WordPress sites rely on an SEO plugin to sort this out, and that usually only lays the foundation

A standard installation puts down Organization and WebPage markup, but no Service markup with a service area, no Product markup with current stock, no Review markup attached to the right entity.

And worse than missing markup is correctly formed markup that lies: a price in the JSON-LD that differs from the price on the page, or stock that has said sold out for weeks. An agent that reads contradictory data has no reason to give you the benefit of the doubt. So check not only whether markup is there, but whether it is accurate; how you do that is in the step-by-step guide further on, and our free SEO scan already looks at 24 points for this.

DOC.06

For online shops: feed, data and platform

If you have an online shop, your most important asset in this story is the product feed. The Universal Cart, AI Mode and also ChatGPT's shopping features draw on catalogue data, and for Google that route runs through Merchant Center. A good feed is current, complete and consistent with your site: the same price, the same stock status, the same variants. At Marketing Live 2026, Google also opened dozens of new data fields in Merchant Center built for conversational questions: which accessories go with this, what is a comparable alternative, what are frequently asked questions about this product. Every field you fill in is a question the assistant can answer with your answer instead of a competitor's.

The second pillar is the structured data on your product pages themselves: Product, Offer, AggregateRating and BreadcrumbList

If you run on Shopify, you get part of this, and soon the UCP connection too, largely as a gift from your platform. If you run on WooCommerce or a custom build, this work falls to your developer, and it pays to combine it with the speed and technology of the site itself, because a slow or technically messy site remains a poor landing place for agent traffic too. How Google judges site speed is covered in our article on Core Web Vitals NL, and what belongs to a healthy technical foundation in technical SEO NL.

And do not forget the text around the data

An agent also reads your returns policy, your shipping information and your product descriptions as plain text, and assistants prefer to cite pages that give concrete answers: what does shipping cost, when does it arrive, what happens with a return. Anyone who also wants to open their site up to language models can consider an llms.txt as well; our llms.txt generator makes one for you for free. These are all variants of the same shift: from a site that only looks good to people, to a site that also answers a machine's questions.

DOC.07
Section 04 · The route

Ready for the agent, in four steps

reg. A.004

The readiness route, in the order that delivers the most. You can take step 01 and 02 this month, step 03 is ongoing maintenance, step 04 is watching and being ready. The skill at the top of this page walks through the same route with you, with your site as the subject.

FIG.04: The readiness route from audit to agent checkoutsheet 4/4
STEP 01 Measure what is there now Check your structured data with the Rich Results Test or the Schema Markup Validator, and place the result next to what is visible on the page. Do price, stock and opening hours match up?
STEP 02 Add the right types Online shop: Product, Offer, AggregateRating, BreadcrumbList. Service provider: LocalBusiness, Service, OpeningHoursSpecification. Everyone: Organization with name, logo and contact details.
STEP 03 Keep feed and text current Merchant Center feed in sync with the site, terms as readable text, and fill in new data fields as soon as your platform offers them. Outdated data is worse than no data.
STEP 04 Watch the checkout layer If you run on Shopify or a large platform, UCP support arrives through your platform. Do not build anything yourself before your platform or payment provider offers a route; the standards move every quarter.
The order is the advice. Whoever does step 04 first is building a checkout for a flow that does not exist yet; whoever does steps 01 to 03 is readable for every assistant already watching today.
DOC.08

What you can do today, and what it costs

The good news in this whole story: preparing for agentic commerce is not about gambling on a protocol, but about work that makes your site better even without agents. Start with measuring. With Google's Rich Results Test, you can see in a few minutes what structured data is currently on your most important pages, and our free SEO scan checks your site on 24 points, including the technical foundation. If you want to learn to understand it yourself rather than just have it done, the background is in our knowledge base, and anyone working with AI assistants such as Claude can use the skills NL from our free library NL, including the agentic commerce check at the top of this article.

Implementation follows after that, and an honest word about cost belongs there

What structured data costs depends on your platform and the state of your site: a WordPress site where only the right types need switching on and filling in is a few hours' work, a custom shop with hundreds of products is a project.

We do not charge separate rates for this: standalone work goes against our hourly rate of 65 euros, and anyone who wants the full picture first can have a complete audit done for 750 euros, in which structured data, technology and content are reviewed together. We do not make promises about what that delivers, because that depends on your market; what we do promise is that after the audit you know exactly what is missing and in what order it pays to tackle it.

And stick to the order that kept coming back in this article: readable first, transactional only after that

The checkout layer changes every quarter and, for most shops, arrives through the platform on its own; the reading layer is already decisive today and changes far more slowly. Whoever invests now in feed, structured data and concrete text is ready for the Universal Cart when it lands here, for ChatGPT that sends people through to their own site, and for every assistant still to follow. Not sure where your site stands? Book a call, and we will go through the four layers together, or start at AI and automation to see how we approach this work.

DOC.09 · FAQ

Frequently asked questions about agentic commerce

What exactly is agentic commerce?

Agentic commerce is the model in which an AI agent handles the buying process on behalf of a customer: searching, comparing, choosing and, increasingly, paying too. The customer gives an instruction to an assistant such as Google AI Mode, Gemini or ChatGPT, and that assistant searches catalogues, weighs up price, delivery time and reviews, and either completes the order or sends the customer through to the shop. The difference from ordinary search: it is not the human who visits ten websites, it is the agent that reads the machine-readable data of those websites.

What is the Universal Commerce Protocol (UCP)?

UCP is an open standard that Google announced on 11 January 2026 and developed together with Shopify. The protocol gives AI agents and shop systems a shared language for the whole buying journey: browsing the catalogue, managing the basket and checkout. Among the partners and backers are Target, Walmart, Etsy, Wayfair, Stripe, Visa and Mastercard, and since April 2026 Amazon, Microsoft, Meta and Salesforce have also joined the technical council. For shops, UCP means they do not have to build a separate connection for every agent.

Does agentic commerce already work in the Netherlands?

Partly. The rollout of features such as Universal Cart and agent checkout starts in the United States, and part of it is not yet available in the Netherlands. But the underlying layer already works here: AI assistants also send Dutch visitors to websites, and those assistants read structured data, feeds and terms to decide who they recommend. So becoming machine-readable already has an effect today, even without checkout running through an agent.

Is this useful to me if I do not have an online shop?

Yes. AI agents look for services just as much as products: a painting firm that can start next week, a physiotherapist nearby, an accountant for a sole trader. For service providers, the same principle applies as for online shops: the agent can only recommend what it can read. LocalBusiness and Service markup, a current Google Business Profile and clear service pages with service area and offer are, for a service provider, what the product feed is for an online shop.

What is the difference between UCP, AP2 and ACP?

They are three open standards, each with its own role. UCP, the Universal Commerce Protocol from Google and Shopify, handles the buying journey: catalogue, basket and checkout. AP2, the Agent Payments Protocol that Google launched in September 2025 with more than sixty partners and later placed with the FIDO Alliance, handles paying: the customer signs digital mandates in advance that allow an agent to check out within limits. ACP, the Agentic Commerce Protocol from OpenAI and Stripe, handles how products are discoverable in ChatGPT; since March 2026, the focus there has been on discoverability, after which the customer checks out on the shop's own site.

What can I do today to be ready?

Four things, in this order. First check what structured data is currently on your site, for example with Google's Rich Results Test or with our free SEO scan. Then implement the JSON-LD types that fit your business: Product and Offer for online shops, LocalBusiness and Service for service providers. As an online shop, make sure you have a current product feed in Google Merchant Center. And make your terms readable: shipping costs, delivery time and returns policy as plain text on the site, not only in a PDF. Everything after that, such as an agent checkout through UCP, builds on this foundation.

DOC.K.09 · Next step

Remember one thing: the agent reads, it does not look

All the announcements, protocols and baskets in this article boil down to a shift you can sum up in one sentence: the next visitor to your website is increasingly a machine that comes to read on behalf of a human. That machine cannot be won over with a nice banner and has no patience for a PDF; it reads your feed, your structured data and your terms, and decides on that basis whether you make it into the comparison. That is no reason for panic, nor a promise of gold, it is a change in who is reading.

So start not with checkout but with readability: measure what is there, fill in what is missing, and keep it current. The free SEO scan is the shortest first step for that, the knowledge base at self learning the longest haul, and anyone who would rather do this together finds through contact the shortest route to an honest conversation about where your site stands today.

Section · Next stepReachable 24/7 · sales@theseo.nl
Book a call
Written by GianlucaFounder. Building visibility for Dutch businesses since 2017, in Google and in AI answers. More about the institute.