Home › Agentic Commerce
Agentic Commerce on Your Terms
AI assistants and agents find you, read your range and buy from you, at your prices and by your rules. For that, your shop has to serve two readers, people and agents. We show you where you stand today and build the path from there, stage by stage, with a result that belongs to you.
What is happening today
Three observations from our own measurements:
- Assistants recommend what they could read. We asked ChatGPT and Google AI Mode ordinary buying questions, in German, from a German account. The assistant recommended one product and one retailer, namely the one whose page it could read. It took the price from the product page or from a price comparison site, never from the shop system. For a cordless drill, ChatGPT added: “The retailer prices are from September.” Depending on the route, the same assistant gave three prices for the same device on the same evening, 29 euros apart between the highest and the lowest.
- Agents are already buying. On the sentence “Please buy this for me”, ChatGPT opened its own browser in the cloud, put the device into an online retailer's cart and stopped before address and payment. It took one minute and twenty seconds. The retailer saw a visit like any other, without a customer account. According to studies by HUMAN Security, Forter and other vendors, most agent activity on the web runs this way, through the interface built for people.
- Every website gets two readers. On Cloudflare's network, over four weeks, bots made 74 % of HTML page requests from Germany and 65 % worldwide (our own query in Cloudflare Radar). According to EHI, 11 % of retailers in the DACH region use agentic commerce.
One path, two entry points
Agentic Commerce follows the three steps in which agents work: find, read, buy.
- Find. AI assistants name you and quote your prices correctly. We measure how ChatGPT, Google AI Mode and Perplexity name you and connect that with your data from pricing, PIM, stock and CDN logs. More under AI Visibility.
- Read. Your shop is readable for agents and hands them data selectively. Every request gets a tier, and you decide who gets what: list price and availability in public, more depth per credential. That is Agent-Ready, described under Agent Channel.
- Buy. The first stage is an MCP server with a UCP profile: agents use it to search, check price and stock and add to the cart. The Store Agent builds on it. It answers the questions of your customers' agents and makes offers by your rules. Your pricing calculates the prices, never the language model. More in the article The Store Agent.
Where you start depends on how you sell today. With a shop or a public catalogue, as a retailer or as a manufacturer with a range online, you start at the first step.
In pure direct business there is no public shop an agent could read. Your customers are registered, and RFQs and orders arrive by mail. Here you start at the last step, behind the login. The Store Agent prepares parts lists, creates them in your system and prepares quotations and contract terms. Anything that touches price, delivery date or legal terms is signed off by a person. When your customers' purchasing agents arrive, they connect to the same service through an MCP server, signed in as the customer.
What raises your mentions in AI answers
These three measures are demonstrably linked to mentions in AI answers. We put them in place on your existing systems.
- Offer data that is right everywhere. Price, availability and delivery time are the same and current on the product page, in the feeds for Google and Microsoft and in the data for ChatGPT. OpenAI names availability, price and quality as criteria by which ChatGPT selects retailers. We supply the feeds directly from your PIM and your pricing.
- A domain that assistants can read. Among German shops, the most visible fifth on Google receives 72 % of all citations in AI answers. We let the assistants' fetchers through, deliver price and availability in the HTML and check continuously that bots and people see the same facts.
- Videos in which your products appear. Of all the factors studied, mentions on YouTube are the most closely linked to mentions in AI answers. Good videos can be produced with relatively little effort as a project: agents prepare topics, scripts from your product data, editing and descriptions, and your team speaks and signs off. That is how we work with Agentic Engineering.
On top of that come mentions through your own data analyses, reviews from real buyers, answer pages built from your product data and product texts for questions in natural language. More under AI Visibility. We measure what a measure achieves before and after; we give no forecast.
Stages and Proof of Value
The build-out has four stages. The first two are safe today and pay off:
- A, the basics: allow search and audit crawlers, the price in the HTML, complete structured data, current feeds, one bot policy from one source and measurement from the CDN log. Many shops still lack this.
- B, agent catalogue: an interface for agents with search, product and cart, described in the Universal Commerce Protocol, with quotas per operator. Almost nobody in the DACH region has it.
- C and D are the outlook: customer prices and terms for agents with a customer account, then checkout with payment authority. The standards for these are in draft or run only in the US so far.
A and B are the Proof of Value. It runs for 90 days and tests two claims that we measure separately. In weeks 1 and 2 we build the read path on one product family, with a second family as control; the claim is visibility. In weeks 3 to 8 follows the agent catalogue on your existing systems; the claim is the agent path. Throughout, we measure pass-through per bot from the CDN logs, retrievals and citations of your domain, the mention rate as a range, calls to the endpoint, and enquiries and carts that come from AI.
Every week has stop criteria. Every stage has a fixed price and acceptance criteria, and the result is yours even if you stop there. Legal questions, such as channel pricing or prices only after login, we clarify together before the offer.
Why Dewline, and what stays with you
Product data and prices are our core, and they belong together: prices can only be set correctly when the product data is right. That is exactly what agents need, complete product data and the right price at the time of the request. Real-time pricing per channel has been running at one of our references since 2020. For a retailer with its own platform we run PIM, pricing and feeds from one integration layer, so that price and stock changes reach the endpoint, the edge and the feeds at the same time. We know both sides, PIM, BMEcat, ETIM and feeds on one, UCP, MCP and agents on the other. Your suppliers' price lists we read and map with Process Automation.
Agents that work with purchase prices and price floors run on models with you. The cockpit and the agent catalogue are built in your project and belong to you. We use measurement data from visibility platforms, and the edge stays with your CDN partner.
Read more
- AI Visibility: the find step, with the article series on visibility in AI answers
- Agent Channel: the read and buy steps, with the article series “Is your shop agent-ready?”
Sources
Our own tests in ChatGPT and Google AI Mode, 2 November 2026; our own query in Cloudflare Radar, 5 October to 2 November 2026; HUMAN Security, April 2026; Forter, October 2026; EHI, September 2026; OpenAI, help article “Shopping with ChatGPT Search”; SISTRIX with LEAP, “KI & SEO im deutschen E-Commerce 2026”; Ahrefs, correlations on AI brand visibility, December 2025.