July 27, 2026

Agent Spend Gets a Number, an Aggregator Walks Into ChatGPT With 500 Million Products, and Asia Funds Its Own Rails

Week of July 21-27, 2026. PHD and WARC published the first market sizing with a 2026 column rather than only a 2030 one, and named the UK the third-largest agentic market in the world. Shoppable became the first commerce infrastructure provider in ChatGPT’s plugin directory, bringing a universal catalogue of 500 million products and a multi-retailer cart into the conversation. Michaels published early numbers for the assistant it built with Google Cloud in six weeks. Ant International closed a $1.2 billion round earmarked for agentic commerce. And European merchants have until Sunday on the AI Act disclosure rule, with no grace period on that part.

Last week we covered commerce becoming the most attacked industry on the internet, agent readiness getting a public score, and x402 getting a standards body. The question that week was which agents you let through the door. This week the question moved one step further in: once they are through, who is doing the selling, and on whose terms?

Because the answer is no longer obviously you. The week’s biggest story is an intermediary with half a billion products walking into the largest AI surface on the internet without asking any merchant’s permission. The counter-story is a retailer deciding it would rather own the conversation itself.

One housekeeping note before the news, because it has a date attached: if you sell in the EU and you run an AI assistant or chatbot on your store, the AI Act disclosure obligation applies from Sunday, 2 August, and unlike the watermarking requirement it gets no extension. Details in the European section, and it is action item one.

Here’s what happened.

Agentic commerce gets a 2026 number, not just a 2030 one

On July 21, Omnicom’s PHD and WARC published “From Abundance to Agents: How the Delegation of Choice is Transforming Marketing.” Every previous forecast in this space has been a 2030 headline. This one has a current column. Consumer spending facilitated by AI agents is estimated at roughly $944 billion for 2026, which is 1.3 percent of all global consumer spending, and the research projects $3.35 trillion by 2030, or 3.8 percent. The concentration is severe: ten markets take nearly 68 percent of agent-facilitated spending by 2030, with the United States at $1.1 trillion or 31.9 percent of the global total, China at roughly $505 billion or 15.1 percent, and the United Kingdom third at $131.2 billion or 3.9 percent. Growth is not spread evenly across categories either. Telecoms and utilities is the report’s standout, forecast to grow more than sixfold, with travel and transport and financial services close behind. Lucinda Barlow, international CMO at Uber, added a useful brake: agents will influence human behaviour, but people still value their own decision-making.

Why it matters for merchants: The global average is the least useful number in the report. What matters is that agentic spend concentrates twice over, by geography and by category, and both concentrations are steep. Three markets take just over half of the 2030 total between them. A handful of categories take most of the growth, and they share a profile: routine, repeat, comparison-heavy, low emotional attachment, the purchases people are actively glad to stop making themselves. If that describes your catalogue, your exposure is already well above 1.3 percent and the 2030 curve applies to you at full force. If you sell considered, tactile, high-emotion goods, it does not, and you should stop budgeting as though it does. The practical version of this is not a strategy exercise. It is a look at your own order data: what share of your revenue is replenishment or repeat purchase, and what share is discovery of something the customer had not thought about? The first number is your agentic exposure. The second is your defence.

An aggregator walks into ChatGPT with 500 million products

On July 22, Shoppable announced that its universal checkout plugin is live in the ChatGPT Plugin Directory, giving US users access to a universal catalogue of 500 million products with multi-retailer product discovery and cart-building inside the conversation. Shoppers add the plugin from the directory and start a prompt with “@Shoppable.” The company says it is the first commerce infrastructure provider to offer multi-retailer discovery and cart-building through the directory, built on fifteen years of universal checkout work and four US patents. To pay, the shopper is redirected to Shoppable, where a single click buys from multiple retailers at once. Founder and CEO Heather Udo framed the logic plainly: the cart should come to the consumer, wherever they already are. The plugin reaches ChatGPT’s roughly 900 million weekly active users, though it is limited to shoppers in the United States for now.

The context makes this bigger than a plugin launch. OpenAI spent the first half of 2026 backing away from running checkout itself and pushing merchants toward brand-owned apps instead, and as of March around 100 firms had built ChatGPT apps but only one, Instacart, could complete a checkout inside ChatGPT. Shoppable does not clear that bar. It routes around it: everything up to the cart happens in the conversation, and the payment happens one redirect away. Which turns out to be enough.

Why it matters for merchants: This is the intermediary layer arriving at scale in the biggest AI surface there is, and it cuts both ways. On the upside, your products can now be found and carted inside ChatGPT with no integration, no protocol decision, and no developer. On the downside, your products can now be found and carted inside ChatGPT with no integration, no protocol decision, and no developer.

You may be listed, compared, and sold through a party you did not sign with directly, in a catalogue you do not control, next to competitors you did not choose, using whatever price and stock data the aggregator happens to hold. That is the affiliate and marketplace problem all over again, except the shelf is a conversation and you cannot see it. So find out whether your products are in that catalogue and whether the data is current, because a stale price or a wrong stock status there is a lost sale plus a support ticket you will not be able to trace. Then decide deliberately whether being aggregated is a channel you want or a leak you want to close.

Worth noting alongside this: the plumbing under AI shopping surfaces is due for a version bump. The Model Context Protocol’s 2026-07-28 specification, whose release candidate locked in May, is targeted for final publication on 28 July and introduces MCP Apps, which lets a server ship interactive HTML that the AI host renders in a sandboxed frame. In commerce terms that is the difference between your product being described in a paragraph and your product appearing as a real carousel with variant pickers and a buy button. Whoever controls that surface controls what you look like on it.

Michaels takes the other path and builds its own agent in six weeks

On July 21, Michaels unveiled Ask Mike, an AI shopping assistant built with Google Cloud’s Gemini Enterprise for Customer Experience, live on Michaels.com and the retailer’s iOS and Android apps. Shoppers describe a project or a result they want in ordinary language instead of entering product names and working through filters. The early numbers are the interesting part. Since a quiet launch in May, Ask Mike has driven nearly 75,000 conversations, and more than 60 percent of interactions are product discovery rather than service questions. Heather Bennett, president and chief customer officer, says shoppers who engage with Ask Mike convert at more than double the rate of those using traditional site search, and 27 percent of interactions end in a product click or an add to cart. Google Cloud’s Paul Tepfenhart noted the build went from concept to production in about six weeks.

Why it matters for merchants: Put this next to the Shoppable story and the two available strategies are stated cleanly. Sell through somebody else’s agent, which is fast, free, and costs you control. Or run your own, which costs you effort and keeps the customer. Michaels is evidence the second option is no longer a twelve-month platform programme: six weeks, on a managed enterprise product, by a specialty chain rather than a hyperscaler. The conversion claim is the number to take to a budget conversation, but the 60 percent discovery split is the one to internalise, because it says these shoppers are not asking where their order is. They are asking what to buy, which is the highest-value conversation in retail and the one that used to happen on a category page you controlled. It reinforces what Salesforce reported earlier this month, that retailers running their own shopper agents grew sales 59 percent faster than those who did not. The honest caveat: a first-party assistant only reaches people already on your site. It does nothing for the shopper who never leaves ChatGPT. Most merchants will end up doing both, and the ones who do it well will make sure the product data feeding the aggregator is the same data feeding their own assistant.

The rails stop being an American project: Ant raises $1.2 billion for agentic commerce

On July 21, Ant International closed a Series A of approximately $1.2 billion, with existing investors including Ant Group and Alibaba Group participating alongside international institutions, earmarked for cross-border payments and agentic commerce solutions. Ant International runs operations across Asia, Europe, the Middle East, and Latin America, and connects more than 150 million merchants to over two billion user accounts. The money is aimed at merchant payment, treasury, and credit technology for businesses trading across borders. It also arrives on top of a protocol Asia already has. In April, UnionPay International released its Agentic Payment Open Protocol framework, covering agent identity management, intent verification, and authorisation standards across its global acceptance network, with a first live agent-initiated transaction completed in Hong Kong on 2 April. Its stated design principles are regulatory controllability, security through agent identity and traceable intent, clearly allocated liability for disputes, and low-cost compatibility. UnionPay put it back on display at the World AI Conference in Shanghai this month.

Why it matters for merchants: Two weeks ago the x402 Foundation looked like it might become the single open handshake for agent payments, and it may still. But UnionPay has a functionally overlapping framework under different governance, and Ant now has $1.2 billion to build agent-era payment infrastructure across four continents including Europe. That is not a footnote for European merchants: Ant is already a live acquiring and wallet relationship for a lot of cross-border sellers, and UnionPay acceptance is real in European travel and luxury retail. The read is not that you should pick a side. It is that the same pattern, verified agent identity plus provable intent plus pre-allocated liability, is now converging from three independent directions, which is strong evidence the pattern itself is settling even though the implementations will not interoperate for a while. Practically, it sharpens the question for your payment provider past the one we asked last week. Not “are you in the x402 Foundation,” but “how do you tell a delegated agent purchase from card-not-present fraud today, and does that still work when the agent runs on infrastructure that was not built in California?” This is the trust gap we covered in June, now with more parties racing to close it.

The European angle: the disclosure rule lands Sunday, the expensive half gets until December

Two things worth pairing. The first is the PHD and WARC ranking, which puts the UK third globally and ahead of every other European market. The second landed on July 22, eleven days before the AI Act’s transparency regime starts to bite. Law firm Addleshaw Goddard published an analysis of the Commission’s guidelines confirming that from 2 August 2026 providers must ensure interactive AI systems tell people they are dealing with an AI at the latest at the point of first interaction, and that AI-generated content carries machine-readable marking, with chatbots on online platforms explicitly in scope. The part merchants keep missing is that the two obligations have different clocks. There is a grace period to 2 December 2026 for implementing the technical marking and detection measures, but it applies only to systems already placed on the market before 2 August, and it does not touch the disclosure obligation, which starts on the day. The guidance also notes that AI-generated product advertising can fall within the deepfake provisions if it could mislead consumers about how a product actually looks or performs.

Why it matters for merchants: Split the work in two and the panic goes away. If you run an AI assistant, chatbot, or agent-facing helper on your store, the “you are talking to an AI” notice is due Sunday, there is no extension for it, and it is a copy change and a design decision. Do that this week. The watermarking and machine-readable marking of AI-generated imagery and copy is the expensive part, and if the system went live before 2 August you have until 2 December, which is a November project rather than a July one. Check that condition carefully, because anything you launch in September gets no relief. And take the advertising note seriously: if you generate lifestyle imagery showing a product in a state it does not actually ship in, that is now a disclosure question and not only a returns question. Then step back to the WARC ranking. The UK sitting third worldwide while continental markets trail is not a compliment to Britain so much as a record of where the US platforms shipped first, a head start we have tracked since June. The strategic point for continental merchants is that the demand arrives whether or not the surfaces launch locally, because your customers are already using US-hosted agents in English. Being unreadable to those agents is a decision you are making by default. We have written before about how far most catalogues are from ready, and that gap is the thing the deadline does not fix for you.

What moved this period

Development	What happened	Why a merchant cares
PHD + WARC market sizing	$944B agent-facilitated spend in 2026, 1.3% of global total	Agentic commerce now has a 2026 number, not just a forecast
Shoppable in ChatGPT	500M-product catalogue, multi-retailer cart, US only	You can be sold in ChatGPT by a party you never signed with
Michaels “Ask Mike”	75k conversations, 2x site-search conversion, 6-week build	First-party agents are now a quarter’s work, not a year’s
Ant International raise	$1.2B Series A for cross-border and agentic commerce	Agent payment rails are not only being built in the US
MCP 2026-07-28 spec	Final publication targeted 28 July, adds MCP Apps	Product carousels render inside the chat, not just text
EU AI Act Article 50	Disclosure due 2 August, marking deferred to 2 December	The cheap half is due Sunday, the expensive half is not

What merchants should do this period

1. European merchants: ship the AI disclosure notice before Sunday. If a customer can talk to an AI on your site, they have to be told at the point of first interaction, from 2 August, with no grace period. This is a small change and a hard date. The technical marking of AI-generated content is the separate, expensive obligation, and it runs to 2 December, but only for systems already live before 2 August. Confirm which of your systems qualify, get the notice shipped this week, and schedule the watermarking work for November. (See our earlier coverage of the Article 50 obligations and EuroCommerce’s objection to their scope.)

2. Measure your category’s exposure from your own order data. The 1.3 percent global average tells you nothing about your business. Pull the split between replenishment or repeat purchases and genuine discovery purchases over the last twelve months. The first bucket is what agents take first, and if it is most of your revenue you are exposed well above the average today. The second is what agents are worst at, and it is where your merchandising still earns its keep.

3. Find out whether you are already in somebody else’s catalogue. Shoppable is one aggregator with 500 million products and it will not be the last. If you sell in the US, add the plugin and search your bestselling SKUs to see who is offering them, at what price, with what stock status. If you are outside the US you cannot run that test yet, so ask your distributors and affiliate partners which universal catalogues they feed, because that is how your data gets there. Stale data in an aggregator is a lost sale you will never see in your analytics.

4. Take the Michaels build to a budget conversation. Six weeks from concept to production on a managed enterprise product, with a claimed doubling of site-search conversion, is now the reference point when someone tells you a first-party assistant is a multi-year programme. The prerequisite is the same as everything else in this space: clean, complete, machine-readable product data. If that is not in place, that is the project, and the assistant becomes straightforward afterwards.

5. Sharpen the question you ask your payment provider. Last week the question was whether they had joined the x402 Foundation. This week it should be operational: how do you distinguish a delegated agent purchase from card-not-present fraud today, and does that hold for agents running on non-US infrastructure? With Visa, Mastercard, the x402 members, UnionPay, and a newly funded Ant all building toward verified agent identity and provable intent, your provider should have an answer this quarter.