ChatGPT, Gemini and Copilot recommend products now, and the sale happens on your store. This is what they read in a feed, which of the new programs reach a Nordic store yet, and how to see their visits in GA4 without fooling yourself.
Ask an assistant for waterproof trail shoes and it answers with products, prices and links. Some of those links lead to stores. In 2026 the plumbing for that arrived, and the merchant stayed the seller. The assistant recommends, you sell.
That leaves two questions. Is your product data good enough to be recommended? And can you see the visitors when they arrive? One caution first: most of the shopping programs open in the United States, so for a store in Sweden the useful parts are the ones that do not wait for a program. What follows is a design to test, built from the vendors’ own documents. It is not a report of results.
Where does a store in Sweden stand? OpenAI’s product feed currently targets the US, with Canada and Mexico needing market setup. Google’s UCP checkout covers the United States, Canada and Australia for select merchants. Sweden is on none of these lists today, and Google says it will work with retailers to expand globally. Check your own markets before you plan around any of it.
Whatever the program, the assistants read one kind of thing: a row per purchasable item. OpenAI’s spec has nine required fields (ID, title, description, URL, brand, seller name, image, availability and price), and Google’s product data specification asks for the same facts and more, including a GTIN where one exists. This is the checklist I would work through:
| Field | Why an assistant needs it | What I check |
|---|---|---|
| Title | It is what the shopper’s question is matched to | The product and the chosen variant as plain text under 150 characters, with no promotion and no capital letters |
| GTIN and brand | They tell one product from lookalikes | A real GTIN (8, 12, 13 or 14 digits, valid check digit) where one exists, empty where none does, and the real brand |
| Price | It is compared across stores | The same on the product page, at checkout and in the feed, in the right currency, with VAT handled the same way everywhere |
| Availability | A wrong “in stock” costs trust | Explicit, current and the same on the page and at checkout; OpenAI rejects a row with an empty or unrecognized value |
| Variants | Size and color are what the shopper picks | One row per variant, with its own URL and an ID that never changes |
| Shipping and returns | They separate two similar offers | Cost, time and return window, matching the policy pages |
| Images and description | They are the evidence | A direct image of the right variant and a plain-text description under 5,000 characters |
A title that works names the product and the variant, like “Northline trail running shoes, black, size 10” (an invented example). One that does not carries a promotion, like “SUMMER SALE trail shoe NEW”. The shopper’s question will not contain the second.
Two things sit outside the table. First, one source: build the feed and the markup on the product page from the same catalog data, because the usual failure is a price or a stock status that differs between them. Google’s Product and Offer markup carries the same facts on the page (name, image, offer, GTIN, and shipping and return information). Second, let the crawlers in. OpenAI says to make sure OAI-SearchBot is not blocked, and a firewall or CDN setting can block a bot that robots.txt allows.
Google also added conversational attributes, such as answers to common product questions and compatible accessories. In its test with lululemon, they showed up in half of the relevant AI Mode recommendations. If your market has them, fill them in.
Pick ten products, five that sell well and five from the long tail, and three assistants your shoppers might use. Ask each the way a shopper would, with a need and a budget instead of a product name. For every answer, write down whether the store and the product appear, and whether the price, the stock status and the link are right. Keep screenshots.
Answers change from day to day, so look for a trend, not a single result, and repeat the test every month. When something is wrong, find out where the wrong fact comes from (the feed, the product page or somewhere else) and fix that, not the assistant.
Since May, GA4 has a default AI Assistant channel. It covers arrivals from sources like ChatGPT, Gemini, Deepseek, Copilot and Grok, and it leaves out Google’s AI Overviews and AI Mode, which stay in Organic Search. It works from the referrer: when the referrer matches Google’s list, the medium becomes ai-assistant and the campaign (ai-assistant). ChatGPT also adds utm_source=chatgpt.com to the links in its search results, according to OpenAI’s publisher FAQ, which gives you a second way to find them.
Three limits. A visit with no referrer and no tag is filed as Direct, so the channel is a floor, not a total. The list is Google’s to maintain, so anything it misses falls back to Referral. And clicks from AI Overviews and AI Mode stay inside Organic Search, so this channel does not show them.
For assistants the default misses, and for older data, make a custom channel group. Google’s help page has an example with a regex for ChatGPT, Gemini, Copilot, Claude and Perplexity, tells you to place it above Referral, and says custom channel groups apply to past data too.
Then build one table in Looker Studio or a GA4 exploration: sessions, purchases and revenue by channel group, with the landing page as a second dimension so you can see which pages the assistants send people to. Compare the purchases with the store’s orders the way the reconciliation article does. On Shopify, orders already carry ChatGPT referral attribution in the admin, which is a second count to compare.
Finally, test the click itself. Open ten links from each assistant on a phone and on a laptop and watch GA4’s realtime report. Shopify says the store opens in an in-app browser on mobile and in a new tab on desktop, and whether the referrer survives each is something to see, not assume.
Read it as a floor. For most stores it will be small and noisy, so comparing its conversion rate with other channels needs patience. I would put the AI Assistant row at the bottom of the weekly report until it is big enough to matter, look at the landing pages first, and not move a budget because of it. MER stays the anchor: whatever the assistants send is already inside total revenue.
The assistants also influence visits they do not send. Someone who asks an assistant and then searches for the brand shows up as Organic Search, and someone who types the address shows up as Direct. So I would watch branded search and Direct next to the AI Assistant row. If all three rise together, that is a hypothesis for the ten-product test to check, not a measurement. It is the same lesson as Meta’s view-through: a channel report only credits what it can see.
The shopping programs will keep changing, and most are not open in Sweden yet. The feed hygiene and the measurement are not waiting for them. Both pay off in Google today and in whichever assistant your shoppers use next. The assistant recommends. You are still the seller.
Sources: OpenAI, Products – Agentic Commerce and Publishers and Developers FAQ. Google, New tech and tools for retailers to succeed in an agentic shopping era (January 11, 2026), Boost your holiday sales with these agentic commerce updates (September 16, 2026), About the Universal Commerce Protocol (UCP) and UCP-powered checkout feature on Google and Product data specification and Merchant listing structured data. Google Analytics Help, Default channel group and Custom channel groups. Shopify, Millions of merchants can sell in AI chats (March 24, 2026). Search Engine Land, Walmart: ChatGPT checkout converted 3x worse than website (March 19, 2026) and Google Analytics adds AI Assistant channel to measure AI traffic (May 14, 2026). These programs change often, so some details here may have moved on.
Both tools lose data when a visitor declines cookies, and they lose different things. This is what GA4's consent mode and Piwik PRO's anonymous mode each record, what is observed and what is estimated, where the EU's cookie reform stands, and a test to run on one site.
Google now ships official MCP servers for Analytics, Ads and BigQuery, so an agent can answer questions about your numbers. This is how I would give it a seat instead of the keys: read-only identities, ten questions with known answers, and the ways it still goes wrong.
Meta, GA4 and a store can all be right and still disagree. This is what each one counts after Meta's March 2026 attribution change, how to split the gap into named causes, and which number to trust for which decision.