Stratosphere · 04
The item master gets a new reader
On a shopping agent with two doors, seven per cent who prefer the retailer's AI on objectivity, five funds instead of seven — and why visibility in retail hangs on the readability of the record, not on the brand.
dm has opened two doors, and they lead into different years. One is announced: the service “dm-mitgedacht” is due, according to the “Lebensmittel Zeitung”, to start early next year as an intelligent shopping list; automated purchase handling is meant to kick in when customers explicitly ask for it.1 The other is already open: at mcp.dm.de the drugstore chain gives AI agents access, via the Model Context Protocol, to product data, search and store availability.
The server was queried read-only on 17 October. It knows exactly five tools: product search, product details, store search, availability, and a function for product tiles. All five read. None puts anything in a basket or triggers an order or payment.2 That is a finding about the offer, not a stated rule. It shows what the agent gets from dm today: a record, not a shelf. According to the paper, the buyer responsible for the online business, Marcel Rieser, warns that anyone who already struggles to maintain their master data risks being pushed into the background by agentic commerce; he is not quoted verbatim.
That is the core of this piece. Visibility in retail is decoupling from brand communication and re-attaching to the readability of the record: what the machine cannot read, it does not propose. The sections that follow test the thesis against what is publicly documented, and end with what a business can check on Monday.
Visibility in retail is decoupling from brand communication and re-attaching to the readability of the record: what the machine cannot read, it does not propose.
What the agent reads
Peter Kenning, a professor at Heinrich Heine University Düsseldorf, uses the term “agentic commerce” for a form of trade in which consumers delegate purchasing decisions to artificial intelligence. Marketing therefore has to become “bilingual”, he said according to the release a trade report draws on: companies will have to address not only people but machines. Visibility would then arise less from emotional brand communication than from technical readability and data quality.3 This is a statement, not a measurement, and it comes second-hand; the original release was not found.
Retail expert Marilyn Repp is more concrete. For her, agentic commerce does not begin with the autonomous purchase, but “the moment an AI decides which three products people get to see at all”. Hence: “product data is becoming the new shelf space.”4 (Both translated from the German.) In the same interview she holds herself back: traffic mediated by AI systems on retailers’ sites is still relatively small. A shelf slot is scarce because it is physical. This one is scarce because an assistant perhaps shows only two or three options, and can only show what it has understood.
The customer who is not one
Whoever reads is not automatically whoever buys. On 4 August 2026 the Ninth Circuit decided a version of this question in Amazon.com Services LLC v. Perplexity AI Inc. (No. 26-1444). It concerned an allegation of unauthorised access, triggered by a shopping assistant acting on behalf of users on Amazon.com. The court vacated the preliminary injunction and remanded; it is not a ruling on the merits. The operative sentence: “It is the user who ‘accesses’ Amazon’s computers, with the help of the Assistant to carry out specific acts on Amazon.com.” The accessing party, on this reading, is whoever instructs the assistant, not the assistant.5
Two limits belong in that sentence. The court rules “on the facts before us” and leaves open whether Perplexity, on a different record, could control the assistant. And footnote 5 states that Amazon can still regulate access through private terms of service; that terms of service and bot detection lose their legal footing is not in the decision. The source is the copy of the decision at Courthouse News, not the court’s own server.
What follows is a reading, not something the court says: the agent is not a customer, but it stands where one stood, and acts for them. How many shop this way is best answered by a vendor report, with every reservation that implies. Salesforce, a seller of commerce software, writes in its State of Commerce report that use of agentic search as the first step of the shopping journey grew 200 per cent year over year; the basis is 3,450 commerce professionals, 4,690 consumers and behavioural data from more than 1.5 billion shoppers. Traditional search fell 15 per cent between August 2025 and May 2026, while new channels such as AI assistants grew 38 per cent.6 A growth rate is not a market share. Closer to our subject is another figure from the same report: only 27 per cent of organisations say their customer data is fully unified.
What no record reports
On 20 July 2026 The Coca-Cola Company presented a new visual system. The global release names the “bold red and white palette” as the core of the brand; Zero Sugar is to get “its own distinct visual cues”, including a black Dynamic Ribbon on cans and a black cap on PET bottles. The rollout covers more than 200 markets.7 A shared red for the original and Zero appears only on the Coca-Cola Switzerland page of 24 July: a “shared design language” joined by the bold red. “The new look changes only the exterior”; recipes and pack sizes remain unchanged.8 (Translated from the German.) For other markets this is not documented.
Both sources speak of packaging, not records; whether EANs or packshots are adjusted anywhere, neither documents. What follows is therefore an inference by 44up, not a statement from a source: if recipe and pack size stay the same, there is no reason to change the EAN, and a record kept under the EAN has no reason to report the change. The product looks different; the record says what it said yesterday. For people at the shelf that is a footnote. For a reader that knows only the catalogue, it is a product it can describe wrongly with full conviction.
Record Catalogue Agent Proposal
● ───────────── ● ───────────── ● ○
absent
Seven per cent
Which AI respondents prefer on objectivity has been measured by KPMG, fielded by EHI Retail Institute GmbH. Taking the smartphone as an example, 72 per cent would seek help from an agent when searching for information; for a binding step such as payment, that drops to 45 per cent. 79 per cent want to be able to intervene at any time. And 64 per cent prefer general generative AI when it comes to objectivity and independence from commercial interests, while only 7 per cent prefer retailer or manufacturer AI.9 The press release gives no sample size; 500 respondents appear only in a Lebensmittel Praxis trade report.
YouGov shows the same pattern on 15 October. Two thirds of potential Black Friday and Cyber Monday shoppers (66 per cent) believe AI tools can find the best offers, more than customer reviews and online communities (63 per cent). Search is readily delegated, the close is not: only 10 per cent would let AI complete a purchase after their approval, 8 per cent would let it carry out a purchase fully automatically. The subgroup is over 300 people in a sample of over 1,000 adults in Germany.10
The series has asked since its first piece whose claims are believed. These surveys do not answer that; they measure two other things that are easily confused: KPMG a preference on one criterion, YouGov confidence in a capability. Neither measures trust in a retailer’s statement about its own product. Both point the same way, however, and that is our reading: the authority that is listened to is moving from the brand to an intermediary reader. It knows about a product what the record says. Whoever makes a claim now has, for the first time, a counterpart that looks it up.
The reader in-house
On 14 October Rossmann described in a press release how it makes company data usable by chat in natural language. The solution is part of the data strategy started two years ago; logistics comes first, for which “quality-assured data products” were built. Christian Metzner, managing director IT & HR: “The biggest bottleneck in using data is today often not the availability of data but access to it.” Logistics, it says, had “deliberately decided to invest in the fundamentals and reliable data products”.11 (Translated from the German.) This is a self-description without impact figures, and it concerns an internal reader. The point is the sequence: data products first, then the chat. Whether the reader is a colleague in logistics or an agent acting for a customer, it reads a record and fails on its gaps.
The rules, too, are moving around this reader. Section 5b(3a) UWG, in force since 27 September 2026, provides that where a service compares products and provides information on ecological or social characteristics or circularity aspects, the comparison method counts as material information.12 That concerns comparison services of this kind, not agents in general; reading agents as such is an interpretation.
What a record now has to deliver
For Monday there are three checks, none of them a project.
- Put your customers’ questions about your own products to the assistants. In MIT Sloan Management Review, researchers who interviewed 15 senior leaders describe a case: the team of a New Zealand investment platform (pseudonym Stakewise) watched ChatGPT become a channel of roughly 700 referral sessions per month. The responses they found “confident, fluent, and frequently wrong”; one widely repeated incorrect claim stated that the platform offered five (not the actual seven) funds.13 That is a method example, not a market finding.
- Check which attributes an agent filters on, and whether they exist as maintained fields in your item master. dm’s product search returns up to 15 hits, with GTIN, price and attributes such as vegan, gluten-free and purchasable. A filter that meets a field nobody maintains proposes nothing.
- Treat every change that leaves the EAN unchanged as one that no record will report on its own. If the identifier stays, the entry stays silent until someone updates it.
Record Catalogue Agent Proposal
● ───────────── ● ───────────── ● ───────────── ●
read
The agent is not an intruder to be kept out. It is a customer to be served — or one that cannot be.
The series has looked at the same subject for four pieces. The first described a data layer that does not hold together, the second a seal that substitutes for data, the third the ways a proof can fail. This fourth describes what happens when the reader of the record becomes a machine: The agent is not an intruder to be kept out. It is a customer to be served — or one that cannot be. Whoever cannot serve it is not on the list the human finally sees.