Is your shop agent-ready? · Article 1 · 2026
It is easy to test which shop an AI assistant recommends
On the evening of 2 November 2026 we asked three ordinary shopping questions, in German, from a German account, the way your customers do. Once in ChatGPT, once in Google's AI Mode. The answers tell you more than most studies on agentic commerce, because they show what already happens: the assistant recommends a product and a retailer, names a price, and on request puts the item in the cart. Which retailer gets named and at what price is decided by things most retailers do not have on their list.
The experiment
Three questions of the everyday kind:
- "Which 18-volt cordless drill under 150 euros do you recommend, and where do I buy it?"
- "Quiet 60 cm dishwasher, energy class B, up to 700 euros, which models and where?"
- "Men's neutral running shoes, size 44, what do you recommend and where do I order?"
Each question went to ChatGPT (paid account, web search on) and to Google's AI Mode on google.de. For the first question we also opened the Google Shopping tab and asked the idealo app inside ChatGPT. All from Germany, all on the same evening. This is not a study. It is a look through the window, and for what we want to show, it is enough.
What the assistants answered
ChatGPT answered three times with the same structure: one recommended product, one retailer, one price, one link to the product page. For the drill, an Einhell kit at BAUHAUS for 145 euros. For the dishwasher, a table with three machines and three retailer prices, among them a manufacturer's own shop and two specialist retailers, with the note "check the retailer price again". For the running shoes, three shoes with one retailer each, price and delivery time: Intersport, Foot Locker, Runmarkt.
Google's AI Mode answered differently. It named models with price ranges and retailers as a category: "Big chains like OBI, Bauhaus or Hornbach usually have the common models in stock." For the dishwasher it recommended checking "idealo.de or billiger.de" for current prices. It did not name a single retailer with a concrete offer and price.
Two things stood out. First: none of the six answers showed product cards or an offer carousel of the kind described in US reports on ChatGPT Shopping. The recommendation was text, with links. Second: a good part of the sources both assistants cited were price comparison sites. ChatGPT cited Geizhals and idealo, AI Mode cited Check24, billiger.de and idealo. The retailer itself appeared when the assistant had been able to read its product page.
The sentence that explains everything
In a second run, this time in the browser on chatgpt.com, ChatGPT recommended a Bosch kit at Alternate for "125 to 130 euros including shipping" and added, verbatim:
"The available retailer prices are from September, however; I could not directly confirm today's final price."
The price in the answer was six weeks old, and the assistant knew it. That is the core of the problem, and it affects every retailer whose prices move: the assistant works from a snapshot. At some point it read a page, pulled a price comparison or processed a feed, and that is the state it recommends from. Whether your price is different today, whether the item is still in stock, whether shipping has changed, it does not know. It cannot know, as long as it sees your shop the way a human does: as a page it has to read.
Where the snapshot comes from was examined in autumn 2026 by Peec AI, a Berlin company that measures visibility in AI answers: in Europe, ChatGPT sources the products in its answers mostly from Google Shopping and from price comparison sites it queries in the background. These background searches, "query fan-outs" in the trade, are the actual mechanism. From one customer question the assistant makes several search queries, collects the hits and builds the recommendation from them. Whoever is readable for those background searches gets found.
The route decides the price
For the same Einhell kit we got three prices on the same evening, depending on the route the request took:
In the ChatGPT app with web search: 145 euros at BAUHAUS, read from the product page. In the browser: a different kit at Alternate, with prices from September. Through the idealo app inside ChatGPT: the same Einhell kit at Proshop for 115.61 euros, with the note "The total prices come from the idealo offers retrieved just now."
The difference is the kind of connection. Web search reads pages and indexes, which is the past. The idealo app asks idealo through an interface, which is the present. Whoever has an interface delivers fresh prices. Whoever only has a page delivers whatever the assistant read at some point. And whoever is listed on idealo and cheap there wins the app route, even if the assistant never met their own site.
"Please buy that for me."
Finally we asked ChatGPT to buy the recommended kit. What happened next took one minute and twenty seconds: ChatGPT opened its own browser in the cloud, went to alternate.de, searched for the kit, picked the variant with two batteries, put it in the cart and reported back: "The Bosch UniversalImpact 18V-60 is in the cart: 127.89 euros including shipping. Expected delivery: 4 to 5 November. Not ordered yet: please add your address and payment method in the open checkout."
No standard, no interface, no consent from the retailer. The agent operated the shop the way a human does, through the user interface. Alternate saw a browser visit and a cart without a customer account. Nothing about it distinguishes the agent from a human, and nothing about it lets the retailer show the agent anything other than what every visitor gets: no customer price, no stock per location, no clarifying question.
That is the normal case today. According to measurements by the security vendors HUMAN (April 2026) and Forter (October 2026), around 70 per cent of agent activity on the web runs through browsers operating interfaces built for humans. The standards for the direct route, first among them the Universal Commerce Protocol from Google and Shopify, exist and are live in more than 19,000 shops, almost all on Shopify and Wix. In Germany, Austria and Switzerland we found exactly one shop with its own platform serving such a profile on 2 November. Until the interfaces are there, agents buy through the front door.
What this means for retailers with their own platform
You cannot derive market shares from three questions. You can derive four things, and they hold regardless of how large the assistants' share of revenue is today:
Whoever is readable gets named. The assistant recommends what it could read: product pages without hurdles, price in the text, complete structured data, a sitemap that works. That sounds like search engine basics, and it is. Hardly anyone checks it against agents today, and when someone does, it is on the homepage. The day after our test, we retrieved 100 product pages each from Hornbach and toom, the way an agent reads them. At Hornbach, the price was in the text on 99 pages, with GTIN and stock in the structured data. At toom, without JavaScript, the page showed nothing but the heading. What we learned about 34 shops in the process is in article 5 of this series.
The price comes second-hand. As long as the assistant takes your price from the page or from a comparison site, it is as old as the last retrieval. If you adjust prices daily or per channel, the assistant recommends you at a price that no longer exists. Or it recommends the competitor whose comparison-site entry is fresher.
Price comparison sites have become your suppliers. idealo, Geizhals, Check24 and billiger.de appeared in our six answers more often than the shops themselves. They give the assistants structure and freshness that retailer pages do not. Whoever is not listed there, or listed with wrong data, is represented in the AI answer accordingly.
The agent already buys, and you do not notice. Every cart filled by a browser agent looks like a human in your numbers. You do not know how many there are, you cannot offer them anything of their own, and you cannot recognise them when they do something you do not want. The first step is therefore not a new sales channel. It is sight: who comes, with which credentials, and what they get.
What we make of it
We call the goal an Agent Channel: an interface for agents next to the interface for humans, through which your shop selectively serves what it wants to serve, with a cockpit in which it sees who retrieves what. The first level, Agent-Ready, starts with the reading path from this article and ends with an agent catalogue carrying fresh prices. The second level, the Store Agent, answers questions and makes offers by your rules. The next articles in this series are about that.
Sources
- Own measurement, 2 November 2026: three shopping questions in ChatGPT (app and chatgpt.com, paid account, web search) and Google AI Mode (google.de), Google Shopping DE, idealo app inside ChatGPT; screenshots on file. Prices and wording as in the answers, translated from German.
- Peec AI, study on the origin of products in ChatGPT answers (autumn 2026; coverage by Search Engine Land).
- Own product page check, 3 November 2026: 100 product pages per shop from the sitemap, without JavaScript, checker identifying itself under its own name (details in article 5).
- HUMAN Security, State of Agentic Traffic (April 2026); Forter, network data on agents (October 2026). Both vendor studies.
- Universal Commerce Protocol: ucpchecker.com and our own retrievals of
/.well-known/ucpat DACH retailers, 2 November 2026. - Cloudflare Radar, share of bots in page requests, own query 2 November 2026 (for article 3 of this series).
We have no ties to the brands, retailers or comparison sites mentioned. We chose the products, the assistants picked the brands and retailers in their answers.
What does an agent see on your product pages?
We look at it together with you, on your own pages.