Shopping online is currently shifting away from the product page and into the chat. OpenAI has launched its own payment function that can be used by AI agents: ChatGPT can use it to buy products on behalf of users, with approval given via biometrics – i.e. via fingerprint or facial recognition on the end device. In parallel, Otto boss Boris Ewenstein describes in an interview how the shopping platform is changing from a catalogue to AI-supported interfaces and how it intends to remain relevant despite new checkout channels.
Both reports show the same development from two directions: the platform side is building purchase completions directly into the assistant, while the merchant side is rebuilding its buying experience around AI. For online shop operators, this creates a new mandatory task – their data and processes must no longer work only for people and search engines, but also for software that shops on their behalf.
What is shifting technically
Until now, the shop was the stage: product images, filters, reviews, shopping cart. When an agent – i.e. an AI system that independently executes several steps – makes the selection and triggers the purchase, this stage loses weight. What becomes decisive then is what is machine-readable: product attributes, availabilities, prices, shipping costs, return conditions.
That sounds like an old topic, but it is a new level of quality. A human visitor forgives an incomplete size chart or a price that appears only as a graphic in the banner. An agent does not: what it cannot read out in a structured way does not factually exist for it. Shops with well-maintained feeds, clean structured data and documented interfaces simply have better chances here of even being considered.
The practical test dampens expectations
At the same time, restraint is appropriate. A practical test around Black Friday comes to the conclusion that ChatGPT, Gemini and comparable systems are still far from working as reliably as their confidently formulated answers suggest when searching for the best price, the right product and the right retailer. That is precisely the critical point: the interface appears competent, but the data situation behind it often is not.
For shop operators, this results in a double consequence. First: anyone relying today on agentically mediated revenue should treat it as an experiment and not as a channel with a reliable forecast. Second: part of the unreliability arises on the merchant side – through outdated prices in feeds, contradictory variants or availabilities that are not accurate in real time. In this environment, data quality is no longer a hygiene factor but a visibility factor.
The merchant perspective: gap to stationary retail
The Otto argument starts at a different point. Ewenstein describes AI as a means of closing the gap to the buying experience in stationary retail – i.e. advice, classification and orientation that a filter form does not provide. In this reading, the path from the printed catalogue through the classic web shop to dialogue-based interfaces is a continuation, not a caesura.
The second half of the statement is interesting: the question of how a platform remains relevant when checkout processes increasingly take place on third-party channels. That is exactly the strategic risk that the OpenAI payment function makes visible. If the purchase completion happens in the assistant, the shop becomes a fulfilment backend – with less contact with customers, less data and less opportunity to place brand and advice.
What shop operators can prepare now
No finished roadmap can be derived from the three reports, but a sensible sequence of preparatory work can:
- Consolidate product data: a leading source for attributes, prices and stock instead of distributed maintenance in the shop, feed and marketplace export.
- Take structured markup seriously: machine-readable information on price, availability, variants and conditions directly in the markup.
- Document interfaces: a stable, versioned API for catalogue, stock and order – even if it is used only internally today.
- Decouple checkout: separate order creation, payment and fulfilment in such a way that a purchase completion also works when it does not come through your own interface.
- Think about returns and service too: anyone who does not trigger the purchase themselves must still map returns, withdrawal and support cleanly.
None of these measures is sensible exclusively for AI agents. They also pay into marketplace integrations, classic price comparisons, mobile applications and internal processes. That makes them a comparatively low-risk investment in a field whose further development is open.
Open questions
At the current state, much remains unclear: how widespread agentic purchases will actually become, what role merchants will play in selection and placement, and what warranty and liability look like for a purchase triggered by software. At least the biometric approval shows that human consent remains a fixed part of the process for now.
For web and software projects, this means: the actual work lies not in connecting to a particular platform, but in clean data, clear interfaces and a checkout that can be detached from the frontend. Anyone who creates these foundations can react to new channels without rebuilding the architecture every time.
And those who don't make it are simply overlooked by agents – regardless of how good the product range is.