When I launched MarketOnline in 2005, the “visitor” was a human being. They looked at the photos, read the description, hesitated, came back a week later. Our whole job was to persuade that person: nice banners, big buttons, the simplest possible checkout. Years later, with Retargeting, I built tools that followed exactly that person’s behaviour in order to show them the right product at the right moment.
Now a different kind of visitor is showing up. It has no eyes, ignores banners and doesn’t get excited by “only 3 left”. It’s an AI agent, an assistant the person tells “find me some running shoes, size 43, under 100 euros, delivered by Friday”, and which searches, compares and, more and more often, buys on their behalf. For any online shop the question becomes very simple and very uncomfortable: how do you sell to a robot?
What has actually happened over the last year and a half
I wrote about AI agents as the next step after chatbots two years ago, when they were more promise than product. Since then, the big players in payments and AI have started laying the pipes. In April 2025, Visa announced Visa Intelligent Commerce and Mastercard announced Agent Pay, both initiatives for payments initiated by AI agents. In September 2025, OpenAI launched Instant Checkout in ChatGPT for purchases from US merchants, built on the Agentic Commerce Protocol, an open protocol developed with Stripe. Also in September 2025, Google announced the Agent Payments Protocol (AP2) for agent-initiated payments.
Don’t rush to conclude that tomorrow half the orders of a Romanian shop will come from robots. Much of this started in the US market, and availability elsewhere is a separate conversation. But the direction is clear: the companies that control payments and AI assistants are building the infrastructure through which an agent can pay safely on a person’s behalf.
The new “visitor” reads your data, not your design
An agent doesn’t see a product page the way we do. It looks for structured information: what the product is, what it costs including delivery, which sizes exist, whether it’s in stock, when it arrives, how to return it. If that information is buried in an image with text on it, or in a description written in a hurry, then for the agent your product simply doesn’t exist or, worse, gets misread.
This is the part I recognise from classic e-commerce: the same discipline that helps you in Google Shopping and on price comparison sites helps you with agents. I’ve written before about SEO in the age of AI-generated answers, and the logic is the same, just pushed one step further: it’s no longer enough to be quoted in an answer, you need to be chosen in a transaction.
How to prepare: product data comes first
If I had to pick one thing a Romanian shop should start on Monday, it would be this: clean up your product data. Concretely:
- Complete attributes. Size, colour, material, dimensions, weight, compatibility (“works with iPhone 15”, “fits model X”).
- Real stock and delivery times. An agent that orders an “in stock” product that isn’t actually in stock won’t pick your shop a second time. Stock sync becomes a matter of reputation, not just logistics.
- Clean feeds. The product feed you already send to Google or comparison sites becomes your main shop window. Clear titles, correct categories, EAN codes where they exist.
- Schema markup on your pages. Structured data (Product, Offer, availability, price) on every product page, consistent with the feed. If the feed says 40 euros and the page says 44, you have a problem.
- Machine-readable policies. Delivery costs and times, return conditions, warranty, written clearly and in a predictable place, not in a 14-page terms and conditions PDF.
Honest prices, fast checkout, APIs that don’t fall over
Hidden costs were always a bad idea. With agents they become a suicidal one. A human can be fooled once by a processing fee that appears at the very end; an agent compares the total price and, if it finds a gap between what you declared and what you charge at checkout, drops you from the list. If you want to win, the displayed price has to be the price paid.
The same goes for checkout. A human may forgive a slow form or a payment error. An automated system will simply try somewhere else. That means a simple, stable checkout and, in the medium term for larger shops, APIs through which stock, price and orders can be queried and placed without scraping the page. You don’t have to build it all tomorrow, but it’s worth asking your e-commerce platform what it plans to do about agentic commerce protocols.
Brand and reviews matter more, not less
There’s a fear that agents will turn everything into a price auction. I don’t buy it. People will very often say “buy it from a shop I trust” or “from where I bought last time”. A good agent takes into account the user’s history, reviews, return rates, how quickly parcels arrive. Reputation becomes a signal a machine can weigh. Bought reviews, on the other hand, are a ticking bomb.
Measuring sales that come through assistants
This is a real headache for marketing. If an agent searches, compares and places the order, you lose the ad click, the session, the abandoned cart and the whole story we used to follow in analytics. Attribution gets harder. What you can do now: tag orders coming through integrations and assistants separately, track referral traffic from AI platforms, ask customers at the confirmation step how they found you. I wrote about how I use AI in marketing in marketing my own businesses.
The risks: fraud, wrong orders and who picks up the bill
Say you run a car parts shop. An agent orders a set of brake pads for a different model than the customer’s car, because the data was ambiguous. Who is responsible? The customer, who gave the instruction? The company that built the agent? You, with your unclear product page? I’ve written about this question in broader terms in who is accountable when the algorithm gets it wrong, and agentic commerce makes it very concrete.
On top of that come fraud risks (fake agents, abusive automated orders), consumer protection questions (does the right of withdrawal apply the same way, whoever pressed the button?), data privacy and payment security. The legal answers will become clearer over time. Until then I would keep it simple: unambiguous product pages, clear return policies, an internal procedure for orders that look wrong, and a conversation with your payment processor about how it identifies agent-initiated transactions.
What doesn’t change
A good product, a fair price, on-time delivery and hassle-free returns. Looking back at what has changed in e-commerce since MarketOnline, technologies come and go, but these four things remain the foundation. An agent can bring the customer to you. It can’t keep them if you don’t deliver what you promised.
My take: early, but not too early
Agentic commerce is still in its early days. Adoption will be gradual, with overblown enthusiasm in some months and disappointment in others, as with any new technology. But I believe the winners will be those with clean data, honest prices and a genuine reputation. And the good news is that all of these are useful anyway, even if no robot ever bought anything.
Romanian shops have no reason to wait. Feeds, product pages, delivery and return policies can be put in order right now. When assistants start buying at scale in markets like Romania too, the ones who prepared will already be on the list.
What do you think: would you let an AI assistant shop on your behalf? And if you run a shop, have you started looking at your product data through a robot’s eyes?