Agentic legibility: a problem online and offline merchants cannot solve alone
What do a flower shop in Rome, a restaurant in Copenhagen and an online designer furniture store have in common?
Not much, except that all three take payments, at a terminal or through an online checkout.
But as AI agents start to shop on people's behalf, they're about to have one more thing in common: agentic legibility.
By that I mean whether an AI agent can read a business well enough to recommend it, or to buy from it.
As I write, online retailers are the ones paying the most attention to how AI agents read their products and present them to human shoppers.
The concern is real. But it also makes it easy to miss how deeply agents are going to matter for businesses with a till and a door.
Take a practical example. An American on vacation in Rome wants two dozen red roses delivered that evening to the rental where he and his wife are staying on Piazza Navona.
Until yesterday his best move was to open Google and search. How does that go?
The first results are ads. Under them, portals that are very good at ranking and promise same-day delivery, though who actually makes the bouquet and what margin the middleman takes isn't something the page tells you.
Google Maps, then? A real option, if the shop answers the phone in English and understands where to leave the flowers.
And there's something else the search can't help with. He wants it to be a complete surprise.
His to-do list as a customer: sorting, comparing, shortlisting, buying, checking it actually arrived.
For most people, that's a couple of hours' work, if every sign-up along the way goes smoothly, of course.
Until now the problem was that the customer had to speak the seller's language.
Now the seller has to speak the agent's.
A florist doesn't usually have a catalogue.
The supply changes through the seasons, and SKUs and GTINs aren't part of that world at all.
Which doesn't mean the business shouldn't be readable, and actionable, by an agent. It should.
What the florist has to be able to say to a machine is, first of all, that he's a florist.
Then where he is exactly, whether he delivers tonight, how far, at what price, and what extras he offers. What about a singing delivery man?
None of that is a product catalogue, and that's the point.
A restaurant has a sharper version of the same problem. Its inventory is a table at eight, for four people, with one allergy and a step at the door.
What he needs isn't a way for agents to check out and pay. It's a way to be recommended by them, and far more businesses need that.
So who's in a position to help all of these businesses, and not only the online stores, become legible to an agent?
It turns out the answer might be your payment processor.
Nobody expects the florist to do this himself. He's not going to read a specification, and nothing in his week suggests he should.
Whoever solves this for him is likely to be somebody he already pays for something else, and three European payment companies have moved in that direction this year.
Worldline said it first, in its full-year results in February: it told the market it was ready to link large language models to its payments API, for merchants who want to sell through agents.
Adyen announced a product suite of its own in June: a feed that gives agents a merchant's products, prices and stock in real time, and a cart that passes what the agent picks into the checkout the merchant already has. At launch the suite was open only to enterprise merchants in the United States. That feed is the product half of discovery, and Adyen built it in house.
And in between, on 3 March, Nexi Group, the Milan-based payment company, signed a memorandum of understanding with Google Cloud, committing to two protocols: the Universal Commerce Protocol (UCP), which handles the whole purchase from finding a product to paying for it, and the Agent Payments Protocol (AP2), which is meant to prove that the buyer really authorised the payment.
Then, on 17 September, Nexi announced a partnership with ReFiBuy, a company founded in 2025 and based in Raleigh, North Carolina, that closed a $13.6 million seed round in May 2026.
The joint release says the work complements the Google Cloud track at the product discovery layer, the part where agents find what to buy. Six months earlier Nexi had described discovery and payment as one lifecycle.
I asked Nexi's press office which of the two was accurate. Matteo Abbondanza, Head of Media Relations Italy, replied the same afternoon, on the record, and the answer matched the March description:
"Both are within the same programme, and they address the two phases of the value chain: search and payment."
Nexi also told me the agreement is go-to-market, with no revenue share at this time.
On who the partnership is for:
"The partnership is active in the Nordics, Germany and Italy. The segment is mid-size merchants that hold catalogue data in the CMS and in product catalogue management systems."
Not the florist, then, and not the restaurant. It starts with mid-size businesses in three named markets that already run a catalogue system.
The florist needs this as much as any larger business, only nobody is providing it to him yet.
A simple machine-readable description of the business, reachable and not tied to whichever platform he happens to use, would solve the first part of the problem. And in my view, at scale, it need not cost his payment processor much, in money or in effort, since it already knows what kind of business he runs and where it is. Knowing that gets an agent only so far, though: somebody still has to keep the rest current.
But that's my view of where this goes, not what the agreement is about. And the agreement has a geography that any European player has to work with.
UCP was launched by Google in January, developed with Shopify, Etsy, Wayfair, Target and Walmart, endorsed by more than twenty other companies, and released as open source.
Google's own checkout built on it is currently open in the United States, Canada and Australia, and the protocol's roadmap names India, Indonesia and Latin America as the next markets.
Europe is not on it. And as far as I could find, nobody has counted which European merchants support it.
ReFiBuy publishes a quarterly index with Digital Commerce 360, scoring how ready retailers' catalogues are for AI shopping agents, including whether a store supports UCP.
But it scores the retailers in Digital Commerce 360's Top 1000, which by definition is a ranking of North America's largest online retailers.
So I went and measured it on this side of the Atlantic.
Twenty-five certified merchants in Italy and twenty-five in Denmark, plus six US stores as a reference group, measured on 19 September 2026. The method and the data are public, and linked at the end.
The sample comes from the public member lists of two paid certification schemes that vouch for online shops, which makes these businesses unusually well equipped rather than typical. If anybody in this population had adopted a new convention, it would be them.
Fifteen of the fifty-six serve a UCP discovery document that passes the checks in the published method: the file that tells an agent which services the store exposes through the protocol. All fifteen are on Shopify. None of the other forty-one does.
The census measures one thing: whether the merchant's own domain serves that file. It does not say whether an agent can find the shop some other way.
Eight of the fifteen are in Denmark, four in Italy and three among the six US stores. In Denmark and Italy they are exactly the shops in the sample that run on Shopify.
Those forty-one run mostly on WooCommerce and PrestaShop in Italy, and on Magento, WooCommerce and a few local platforms in Denmark: the kind of CMS Nexi's press office named.
So what should a merchant in that position do?
The most careful and practical answer I've read came from Thomas Walter at Merkle, on 3 September: work on being found, and don't hand over the checkout.
He was also honest about the catch. For a curated or specialist retailer, the protocol can't say why the store is worth choosing over a cheaper one. But the price of staying out could be invisibility.
The flower shop in Rome, the restaurant in Copenhagen and the online furniture store still don't have much in common. But the payment processor behind all three could easily be the same one.
Nexi counts its reach in terminals, about three million of them, according to its own investor page. That sounds like the wrong unit for a piece about product data, and here it is the right one: behind every terminal there is a contract.
In this sample, a merchant had a machine-readable description for agents only where the platform shipped it, and that meant Shopify. If even certified online shops depend on their platform for it, a florist with no online platform has no default provider at all.
For him, and for many of the businesses behind those terminals, the description would have to come from somewhere else, and the natural candidates are the ones already under contract with these merchants and already holding a file on each of them.
A payment processor is one, but not the only one: a company that has sold directory listings to small shops for decades sits in the same position, and arguably closer to the problem, since being found is what it already sells.
A web platform that already holds data on almost every business could do the job too, but if it runs an agent of its own, it may have a reason to put that agent first.
The processor, by contrast, earns on the transaction whichever agent brought the buyer, so on paper, at least, its incentive is to make the florist legible to all of them.
For most of these businesses, the purchase is the last thing that moves, and the recommendation is the first.
Long before an agent places an order with the florist, it has to be able to tell somebody that the shop exists, where it is and whether it delivers tonight.
Back to our three merchants.
Soon they will share one more thing besides a payment processor: how legible they are to agents, and who makes them so.
Agentic legibility is not yet something they can switch on by themselves, and whoever provides it could have a say in which agents can find them.
Sources:
- Five written questions were put to Nexi media relations on 18 September 2026 and answered the same day, on the record. The exchange was in Italian, and the quotations are the author's literal translation.
- Quotations from the announcement are from the joint Nexi and ReFiBuy release of 17 September 2026.
- The Worldline statement is from its FY 2025 results press release of 25 February 2026; the Adyen product description is from its press release of 16 June 2026.
- The AI1000 and the scope of the Digital Commerce 360 Top 1000 are from the two companies' announcements of July 2026.
- Thomas Walter's argument is from his piece at Merkle of 3 September 2026, "The Universal Commerce Protocol: How European Commerce Leaders Should Respond".
- Nexi's scale figure and its unit are from the company's investor relations page.
- ReFiBuy figures are from the company's own materials and from reporting of its May 2026 seed round.
- The UCP launch details are from Google's developer blog of 11 January 2026, and protocol documentation was read on 19 September 2026.
Method, sample and full results: https://doi.org/10.5281/zenodo.22893539.
This is the first of three pieces on the census.