Agentic commerce: selling through ChatGPT from a Swiss online store

Customers are already asking ChatGPT “which insulated bottle for hiking, delivered in Switzerland, under CHF 50?”. The assistant compares, summarises, then suggests links. If your store does not appear in that answer, the customer will not see it. Here is what has changed over the past year, and what an online store in French-speaking Switzerland can do right now.

Agentic commerce: what has happened since 2025

We speak of agentic commerce when an AI assistant does more than answer and acts on the buyer’s behalf: it searches, compares and prepares the purchase. Three players have laid the groundwork.

• OpenAI and Stripe, September 2025. Launch of Instant Checkout, buying without leaving ChatGPT, and of the Agentic Commerce Protocol, an open standard connecting the assistant to merchants. At first, only US users could buy, from Etsy sellers in the United States.
• OpenAI, 2026. According to OpenAI, merchants can now use their own checkout, and ChatGPT is focusing its efforts on product discovery. Merchants share their product feeds so that their catalogue is properly represented.
• Google, 2026. Google presented the Universal Commerce Protocol, an open protocol for discovery and purchasing in AI Mode and Gemini. The merchant remains the seller. The rollout starts with the United States.
• Shopify. Shopify offers eligible merchants a way to distribute their store’s catalogue to AI assistants from the admin. Availability for a Swiss store should be checked in your account.

Discover in AI, buy on the website

OpenAI’s refocus points to a simple trend. The assistant becomes a shop window that sorts and recommends. The purchase itself happens on the merchant’s website, with its basket, its payment methods and its customer relationship. For a Swiss store, this is rather good news: there is no need to wait for payment inside ChatGPT to reach Switzerland. You need to be recommended, then convert the visitor who arrives.

The logic is the same as ranking in answer engines, covered in getting recommended by ChatGPT. The difference: for a product, the assistant needs precise, up-to-date data, not just a good reputation.

How an assistant chooses the products it recommends

No platform publishes its formula. But the merchant documentation from OpenAI, Google and Shopify all stress the same elements. An assistant recommends what it understands and what it can verify.

• Complete product data. A clear title, brand, material, dimensions, variants, identifiers (GTIN, SKU). A poorly described product matches no precise question.
• Price and availability. An up-to-date price, in the right currency, and real stock. An assistant avoids recommending a product that is out of stock or priced differently on the website.
• Reviews. Visible, credible customer reviews help the assistant choose between two similar products.
• Terms. Delivery costs and times, countries served, returns policy. These are often the criteria in the question itself.
• A clear page. A product page that answers questions in readable text, not only in an image or a hidden tab.

What a Swiss store needs to prepare now

None of the following is specific to one assistant. These are the foundations of a good catalogue, which also serve Google Shopping and your Meta campaigns. Agentic commerce simply makes them mandatory.

• The product feed first. It is the source Google reads and, increasingly, assistants too. Descriptive titles, accurate categories, clean images, and automatic synchronisation with stock.
• Product pages that answer. Take the questions your customers ask by email or phone, and answer them on the product page, in text.
• Swiss prices without surprises. Prices in CHF, VAT included, shipping costs and delivery times from Switzerland visible on the product page. If you deliver to the European Union, state who pays customs duties.
• Reviews people can believe. Collect them after every order, publish the less positive ones too, and reply. Invent nothing: a made-up review deceives the customer.
• French and German. A question asked in German more easily finds a page in German. A store targeting the whole of Switzerland should translate its product pages, not just its menus.

Measuring traffic from assistants

An assistant that recommends your product sends a visitor. You still need to see it in your analytics, otherwise you will never know whether your efforts are paying off.

• Referrers. In Google Analytics 4, filter traffic sources: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com.
• A dedicated channel group. Create an “AI assistants” channel grouping these sources, to track their visits, sales and average order value over time.
• UTM parameters. Some assistants add a utm_source parameter to links themselves. In your product feeds, add your own UTMs where the platform allows it.
• Sales, not just visits. Match each order to its source in Shopify or your CRM. Some traffic arrives without a referrer: ask “how did you hear about us?” at checkout.

Keeping a checkout that converts

Since the purchase comes back to your website, every visitor sent by an assistant must find a frictionless journey. They often arrive well informed: they have already compared and want to buy. Shipping costs revealed at the last moment, mandatory account creation or no TWINT are enough to lose them.

Check your leaks with our conversion rate method: a store that converts poorly wastes assistant traffic just as it wastes ad traffic. On Shopify, the native checkout, Swiss payment methods and the product feed are configured together; that is what we set up in our Shopify stores.

What remains uncertain

• The Swiss timetable. Neither OpenAI nor Google has announced a date for Switzerland in the sources consulted.
• The exact criteria. The platforms do not publish how they rank recommended products. They can change these rules without notice.
• Advertising. The place of paid ads in assistant answers is still evolving; nothing allows you to plan for it today.
• Protocols. Several standards coexist. Clean data, reusable everywhere, is a better bet than a single protocol.