Almost everything written about AI shopping agents right now was written for a merchant in the United States. If you run a web store in Slovenia and follow that advice literally, you will spend money on the wrong half of the problem.
Here is the short version. The buying layer is not available to you yet. The deciding layer already is. Agents are recommending, comparing and shortlisting products for Slovenian shoppers today, and the work that determines whether you appear in those answers is unglamorous, cheap, and entirely within your control.
This post is about which work that is.
What actually happened
Three things changed the landscape inside twelve months.
Google unveiled the Universal Commerce Protocol (UCP) at NRF in January 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart. It is an open standard covering the full commerce journey, discovery, cart, checkout and post-purchase, so agents donāt need a bespoke integration with every merchant.
OpenAI and Stripeās Agentic Commerce Protocol (ACP) has been live in ChatGPT since late 2025. It is checkout-centric, and by early 2026 was processing live transactions for Etsy and expanding across Shopifyās merchant base.
Microsoft Copilot Checkout launched in the US in January 2026, with Shopify, Stripe, PayPal and Etsy integrations. Shopify merchants were enrolled automatically, with an opt-out window.
Underneath all of it sits the Model Context Protocol (MCP), originally from Anthropic and now governed by the Linux Foundation, which lets AI systems query live catalogue, inventory and pricing data rather than scraping rendered pages.
That is the hype cycle. Now the part the hype cycle leaves out.
The reality check
Agentic checkout is not a European product yet. Googleās agentic shopping features, agentic checkout, virtual try-on and Business Agent, launched US-first and are rolling out progressively, with UCP-powered checkout extending to Canada and Australia, and the UK later. As the German consultancy diconium summarised in May 2026: autonomous checkout is not currently available in Europe, held up by data protection, payments and regulatory requirements.
And where it has launched, it has underperformed. Conversion rates for purchases completed inside ChatGPT ran roughly three times lower than for purchases redirected to merchant websites. OpenAI subsequently removed in-app checkout and now redirects buyers to merchant sites and apps. Amazonās āBuy for Meā agent frustrated both buyers and sellers over the holiday season because listings didnāt match merchantsā real inventory.
Notice what broke in both cases. Not the protocol. Not the model. The data. Reporting on OpenAIās rollout points at exactly this: maintaining real-time accuracy on inventory and pricing, information usually scattered across teams and systems, is the hurdle.
Which is convenient, because data is the one part you can fix without waiting for Google, Brussels, or anyone else.
The problem nobody is writing about: Slovene is hard for agents
If your store sells in Slovene, you are operating with a handicap that no US-focused guide will mention.
Slovene is a low-resource, morphologically rich language. Benchmarks for it are limited, most standard LLM evaluations canāt simply be machine-translated into Slovene because morphology and context break the translation, and Slovene-language model performance still trails English by a meaningful margin.
For a product catalogue, that has three concrete consequences:
Declension eats attributes. A model extracting colour from bela srajca, belo majico and v beli barvi is doing three different jobs where an English catalogue does one. Every inflection is a chance to lose the attribute.
Machine-translated copy compounds. A lot of Slovenian stores translate supplier descriptions automatically. An agent then re-parses that already-degraded text. Errors stack.
Prose is the worst possible container for a fact. Anything an agent needs in order to match a query, size, material, compatibility, delivery window, availability, should never live only in a description paragraph.
The fix is language-neutral, and that is the point. Structured fields donāt decline. A color attribute with a controlled value, a size attribute in a declared unit, a gtin, these carry the same meaning in Slovene, German and English. Move every purchase-relevant fact out of prose and into structured data, and your Slovene catalogue stops being a disadvantage.
If you sell cross-border, publish structured data in English too. Not translated marketing copy, structured attributes.
Why this is a threat, not an opportunity
The tempting frame is ānew channel, new customersā. For most Slovenian merchants it is the opposite.
Slovenian shoppers already split their spending. In 2024 they averaged roughly 95 EUR per month in Slovenian online stores against 78 EUR cross-border, and those two numbers have been converging. Cross-border is no longer the exception.
An AI agent handed a query like ānajboljÅ”e zimske pnevmatike 205/55 R16 do 400 EURā does not care where the merchant is registered. It compares whatever it can parse. If your listing is missing three attributes the query needs, and a German competitorās listing isnāt, you are not in the comparison at all, regardless of price, stock or delivery time.
IAB Slovenijaās Ecom 25/26 research points the same direction: competition is no longer only local, foreign stores and price pressure are increasing, and measurement, attribution and first-party data quality are the sectorās central weakness, with the explicit finding that poor data quality is what keeps AI use stuck at basic tasks. The same research found generative AI is already widespread among Slovenian companies, but mostly for copywriting, translation and visuals, rarely integrated into processes, analytics or customer experience.
That is the gap in one sentence: the market is using AI to produce more content, while remaining invisible to the AI that its customers are actually asking.
What to do now, by platform
The Slovenian market is not Shopify-dominated the way US coverage assumes. WooCommerce leads on install count, Magento and Adobe Commerce run the larger and more complex stores, and Shopify is growing but far from default. The work is different for each. If you are still choosing, our Shopify vs WooCommerce comparison covers the trade-offs.
If youāre on Shopify
Mostly a switch. Shopify co-developed UCP and ships first-class tooling. The Universal Commerce Agent app takes minutes to install, after which you configure an agent policy defining what permissions agents have. Also check what has already been switched on for you, Copilot Checkout enrolled Shopify merchants automatically.
Your remaining work is catalogue quality, which is the same work as everyone elseās.
If youāre on WooCommerce, Magento, or something custom
You own the whole job, and for now, you should deliberately do only half of it.
The discovery half is worth building today:
- Feed completeness. Every attribute a buyer would filter on. Not the ones your PIM happens to have.
- GTIN / EAN on everything. The one identifier that survives language entirely.
- Real-time price and stock. If your feed and your product page disagree right now, nothing else on this list matters.
- Structured shipping and returns terms, expressed as data. Agents weigh delivery windows. ā3-5 delovnih dniā buried in a footer is invisible.
- Schema.org markup that renders server-side. Client-side-only structured data is a common and expensive mistake.
robots.txtsanity check. Confirm you arenāt blocking the AI crawlers you want, while blocking the ones you donāt.- Response times and rate limiting. Agents produce different traffic patterns than browsers, bursts of parallel structured requests rather than a humanās sequential browsing. Guidance circulating in the UCP community targets sub-200ms for product discovery. Rate limiting should distinguish legitimate agent traffic from abuse rather than blanket-throttling both.
The checkout half, the three core REST endpoints for session creation, update and completion, OAuth for account linking, payment handler configuration, public keys for signature verification, can wait. It requires Google approval to go live, and the surface it goes live on isnāt available to your customers yet.
Reuse what you already have
Most Slovenian merchants already maintain a product feed for Ceneje.si and Google Merchant Center. That feed is your agent-readable catalogue. You do not need a new system. You need an audit of the one you have, and it is almost certainly incomplete in ways that never mattered when only a price comparison site was reading it.
Five minutes, five checks
Run these yourself before you talk to anyone about a project:
curl https://yourstore.si/.well-known/ucpand expect a 404 today. It tells you where the industry is heading and what the file will be.- Ask ChatGPT, Gemini and Perplexity, in Slovene, for a product in your category with a real constraint (budget, size, compatibility). Write down who gets named. Then ask the same question in English and compare.
- Open your product feed. Count the products missing GTIN. Count the ones with fewer than five structured attributes.
- Pick five products. Does the price in your feed match the price on the page, right now?
- Compare your attribute completeness against your best-known foreign competitorās listing for the same product.
If checks 2 and 5 disagree with your self-image, that gap is the whole project. The same logic drives generative engine optimization for the rest of your site.
The regulatory footnote
Agent-initiated purchases raise unresolved questions under EU consumer law, pre-contractual information duties, right of withdrawal, who the contracting party actually is, alongside GDPR and payment regulation. This is a meaningful part of why European rollout lags the US.
Read that as a reason to sequence, not to stall. Build the data layer now, because it pays off immediately in AI-assisted discovery whether or not autonomous checkout ever arrives. Defer the checkout layer until the surface exists in your market and the legal position is settled.
The uncomfortable summary
The stores that win the agent channel will not be the ones that adopted a protocol first. They will be the ones whose product data was accurate, complete and machine-readable before anyone was checking.
That work has been worth doing for a decade. It is now the difference between being in the answer and not existing.
If you want to know how your catalogue reads to an agent and what to fix first, get in touch. We will go through your feed, your structured data and your visibility in AI answers, and tell you which half of the work is actually yours to do.