What changes for your online store when the buyer is an AI agent

For your online store, this changes less than the headlines suggest, but more than your orders show today. Right now, in Italy too, you can ask ChatGPT, Gemini or AI Mode in Google Search to compare products and tell you where to buy them. The next step is an assistant that completes the purchase inside the chat on the customer’s behalf. It already exists, but for now almost only in the United States and with a lot still being adjusted.

Either way, the assistant decides based on the data it finds about your store. If that data is incomplete, outdated or contradictory, it picks someone else.

What agentic commerce means, in plain terms

An AI agent is an assistant that doesn’t just answer but carries out a task for you. In shopping, that means it searches for products, compares them against your criteria, adds them to the cart and, if you’ve allowed it, pays. This is what people call agentic commerce.

For it to work, agents and stores need shared rules, which set out how to read a catalogue, how to create an order and how to prove the customer really authorised that payment. These rules are called protocols, and several have appeared since autumn 2025, pushed by big tech companies and payment networks.

Where things actually stand in October 2026

OpenAI and ChatGPT

In September 2025 OpenAI and Stripe launched Instant Checkout, buying directly inside ChatGPT, together with the Agentic Commerce Protocol (ACP), an open standard for agents and stores to talk to each other. At launch it was limited to US users and US Etsy sellers.

In March 2026 OpenAI changed course. It said the first version of Instant Checkout didn’t offer the flexibility it wanted. It therefore let stores use their own checkout and focused on product discovery, with visual product cards, comparisons and search by image. ACP is still there, but today it’s mainly used to send catalogues and promotions to ChatGPT, and direct catalogue onboarding is limited to approved partners for now.

Google

In September 2025 Google introduced the Agent Payments Protocol (AP2), designed for payments started by an agent. What the customer authorises, such as a product and a spending limit, is recorded as digitally signed proof. In January 2026 came the Universal Commerce Protocol (UCP), developed with Shopify, Etsy, Wayfair, Target and Walmart, which lets shoppers buy from eligible US retailers without leaving AI Mode or the Gemini app. The store remains the seller.

In May 2026 Google announced a single cart across multiple stores for the US and said checkout would come to Canada and Australia, then the UK. Google talks about expanding globally, but for Italy and the rest of the European Union it hasn’t given any dates yet. AI Mode itself has been available in Italy, in Italian, since October 2025. People can already search for products conversationally there, but they buy on the store’s own site.

Perplexity

In November 2025 Perplexity and PayPal launched instant buying inside Perplexity’s answers, again for US users.

So for an Italian store, the real effect today is on discovery and not on payment. That’s where the game is played, and it’s also the part you can work on right away.

What an agent looks for in a store

An AI assistant isn’t swayed by a beautiful photo or a slogan, because it reads the data, compares it and discards whatever doesn’t add up. Here is what it needs.

• Product data has to be complete, with a clear name, a description that answers real questions (materials, dimensions, compatibility, use, care), product codes and photos. A description written only to be evocative is almost empty to an agent.
• Every size, colour or finish needs its own availability and price. If variants are muddled, the agent doesn’t know what it’s recommending.
• Prices and stock should be up to date and consistent, with the same price on your site, in your feed and on marketplaces. An agent that finds two prices for the same product has a good reason to go elsewhere.
• Shipping and returns need to be spelled out, with delivery times, costs, countries served and return conditions. Since November 2025 Google lets stores state these policies for the whole site even without a Merchant Center account.
• The product feed is the automatically updated list of all your products and their details, and it needs looking after because Google Merchant Center and marketplaces read it. Google has said retailers worldwide can now use new attributes designed to adapt product descriptions to the questions people ask in chat.
• Your pages should carry structured data, meaning information in a format machines can read without guessing. It covers what the product is, what it costs, whether it’s in stock and how it’s rated.
• Checkout has to be fast and reliable, with few steps, no surprise costs and common payment methods. Many large payment providers are already involved in these protocols, so when they reach Europe, stores using one of them will probably have less work to do.
• Reviews and reputation count too. Assistants show them and take them into account, so you need genuine reviews, collected steadily and consistent between your site and other platforms.

An agent doesn’t buy from the prettiest store but from the one it trusts, and it trusts data.

The risks and the open questions

The first risk is margins. If an assistant compares ten stores’ prices in a moment, price pressure grows. Stores competing on price alone are most exposed, while those with a distinctive product, better service or reliable delivery have more to stand on.

Then there’s your relationship with the customer. If the purchase happens inside an assistant, the customer sees less of your brand, your site and your advice. Platforms say the store remains the seller, but someone else designs the experience.

Fraud and liability are still open questions too. Who is responsible if an agent buys the wrong thing, or if someone tricks it? Protocols like AP2 exist precisely to prove what the customer authorised, but they’re young.

In Europe, online payments also generally require strong customer authentication, meaning at least two proofs of identity. How that fits with an agent paying on its own is one of the questions still to be settled.

Finally, plans change fast. OpenAI launched and then replaced its checkout within six months, and chasing every announcement gets expensive.

What to do now without overspending

The good news is that what agents need is the same thing that already helps today on Google, on marketplaces and on your own site. You’re not betting on an uncertain future. Here is where to start.

• Get your product data in order, with complete descriptions, correct variants, codes, dimensions and materials. It’s the most tedious work and the one that pays back most.
• Keep a single source for prices and stock, so your site, feed and marketplaces never contradict each other.
• Look after your Google Merchant Center feed and fix the errors it flags.
• State your shipping and returns policies so that Google and other systems can read them without ambiguity.
• Simplify checkout, because every step you remove helps whoever is buying today, person or agent.
• Collect genuine reviews after every order.

What I wouldn’t do is rebuild your store from scratch for a protocol that isn’t available in Italy yet, or pay for services that promise guaranteed sales from AI assistants. If you want to understand how businesses get cited in AI answers, I cover it in a separate article.

How I can help

I rebuilt LaCurvaDesign from scratch, an online store for bent-wood design furniture in five languages, from the storefront and checkout flow to the management of orders, products, customers and payments. I can go through your store’s product data, feed, structured data and checkout with you, and tell you what to fix now and what can wait. If that sounds useful, get in touch.