Companies such as Visa, Mastercard, Google and AWS are already laying the rails that would let AI agents, the kind of AI that can carry out tasks online for people, pay for things themselves. On 4 November, Google let shoppers in Canada buy directly from some retailers in AI Mode and the Gemini app, and nine card-issuing banks in Taiwan took part in the first wave of tests that Visa announced in May. But which business finally gets paid is decided before the payment, and the transaction that Mastercard completed in Taiwan worked exactly that way: the AI first offered recommended options, and the customer chose one before it paid. That picking doesn’t have to wait for payments to open up, and it has already started in Taiwan: on 9 November, Google launched a Chinese-language shopping feature for Gemini in Taiwan, where customers describe what they need as if they were chatting and it recommends a few products straight away, listing prices, reviews and where to buy. So our position is that you shouldn’t wait for the payment rails to reach Taiwan: the information that AI can read and trust when it picks a business needs to start building up now.

What happened

There are really two sets of rails, and they need to be looked at separately.

The first is called x402, a standard that lets AI agents pay other programs automatically. The 402 in its name is the web’s code for “payment required”. An AI agent asks for some data, the other side replies that it costs money, and the agent pays, asks again and gets it, with nobody filling in a form. The Linux Foundation announced on 14 July that the foundation overseeing this standard was fully up and running with 40 companies and organisations on board, including Visa, Mastercard, Google and AWS.

This set of rails is further from your business. x402 payments mostly aren’t customers buying from businesses: an August report from CoinDesk says that the buyers are often AI agents paying a few cents at a time for data, computing power or access to an online tool. A report that Visa funded and produced with the data firm Artemis found that, as of 21 April, the money moving over x402 went from about US$136 million to about US$15 million once wash trading and testing were filtered out. The report itself is still upbeat about these rails, but the money paid over them has little to do with the business in your shop, and the rails that you should be watching are the other set.

The second set is the one that has to do with customers buying things. It runs through the card networks and AI platforms, and testing has already started in Taiwan. Mastercard announced in March that it and DBS Bank (Taiwan) had completed the first authenticated AI agent transaction in Taiwan, which was an airport transfer booking. Visa announced in May that nine card-issuing banks in Taiwan, including Cathay United and E.SUN, would take part in the first wave of its programme, testing transactions initiated by AI agents in a controlled environment. Taiwan Mobile also announced a partnership with Visa in June to let holders of its co-branded credit card buy things and pay through an AI agent at its online store; as of 12 November, we can’t find any sign that this has opened for use.

In Canada, it has already reached shoppers. Google opened up on 4 November so that shoppers in Canada can buy directly from some retailers in AI Mode and the Gemini app. In Taiwan, what Google launched on 9 November is a Chinese-language shopping feature for Gemini, which recommends products, puts together price and spec comparison tables and includes links to buy, and the announcement doesn’t mention paying within the conversation. Gemini’s help page does already list Taiwan among the places where its shopping features are supported, and the same page describes tapping “Buy” in the conversation and checking out with Google Pay, but it doesn’t say in which countries that button is available; as of 12 November, we can’t find any news of direct checkout coming to Taiwan.

Why it matters to you

A September roadmap from the Market Intelligence & Consulting Institute, part of the Institute for Information Industry, puts the “agent economy”, in which AI agents can negotiate and transact on their own, at 2028 to 2030. That may sound two or three years away, but when the time comes, for AI to pick you for a customer it will at least have to be able to read about you, and the reviews that customers write about you and the articles that other websites write about you build up one at a time. They won’t all appear in that one year.

When a customer asks Gemini which product it recommends, or asks AI which business it recommends, today, the AI likewise picks out a few for them first, in the same order as that Mastercard airport transfer, and the only difference is that at the end they click the link or make the phone call themselves. Say you run an airport transfer company in Dayuan, Taoyuan, and suppose a customer in Toufen, Miaoli tells an AI agent, “Book me a car from Toufen to Taoyuan Airport, leaving at 4am next Wednesday, five people and five suitcases.” We haven’t found any platform that has said publicly how AI agents choose a business, but for the agent to put you among its recommendations, at the very least it has to be able to read that you have a seven-seater, run through the night and go as far as Toufen. You can change the words on your website today, but whether customers have written about you on blogs and in Google Maps reviews is something that other people build up little by little, and it can’t be produced at short notice.

What to do about it

Act now, starting with the step that comes before payment. x402 and the digital wallets that let AI pay by itself aren’t things that you need to adopt, because that’s an engineering project, and if someone uses this news to sell you on it, you can just hear them out. What you should look after first is the information that AI can read when it picks a business: the facts on your website and what’s written about you elsewhere. Whenever these rails reach Taiwan, how much AI can read about you will depend on how much has built up by then, and that needs to start now.

If you just ask AI yourself now and then, it’s hard to see this building up change anything. This month a customer writes a new post about you, and next month when you ask AI, the answer may well be much the same, so you can’t tell whether you’re heading in the wrong direction or it’s just too soon.

So get a clear look at where you stand first. Set up a demo with one of our consultants and tell them the kind of conditions that your customers give, like the five people and 4am above. They’ll put those to ChatGPT or Gemini there and then, and who AI picks today, and whether it picks you, is clear at a glance. The same session also goes through your website’s GEO audit score: that’s geoweb.tw’s free audit, which checks how well AI can understand your site.

The earlier you start, the better. If you want it done over the long term, the service is called managed GEO: we revise your website’s content, we tidy up what other websites say about you, and we regularly put the same set of questions to ChatGPT, Gemini and other leading AI engines and keep the answers to compare, which is how the trend becomes visible.


Further reading