Two platforms. The same AI agent. Opposite decisions, inside the same fortnight.
Meta launched Muse, a personal AI agent, and it went to number one on the US App Store free chart — above ChatGPT. Meta's stock rose 11 per cent the following Monday. Shopify had already switched Muse on by default across its stores. Then, on the Sunday, Amazon cut it off. Users trying to shop through Muse on Amazon.com saw a message telling them that continued access by an unauthorised AI agent violated Amazon's Conditions of Use.
Twelve days from launch to being locked out of one of the two largest commerce platforms it touched, while the other built agentic checkout for it. I wrote in the WebMCP post that the most important visitor to a site is about to be an agent acting on behalf of your customer, not a person clicking through it. This is that prediction arriving as a live commercial split, on the same product, in the same week.
What Amazon said, and what I think is actually going on
Amazon's stated reasons are specific, and I take them at face value. Meta didn't seek authorisation. The agent doesn't identify itself when it browses. And Amazon says it captures customer credentials and scrapes order history as it goes — behaviour that falls outside the 2026 rules Amazon has set for automated systems, which require an agent to identify itself and to stop when asked. That's a real security and authorisation complaint, and I have no reason to doubt it.
It is not the whole story though. What Amazon said and what I think is driving it are two different things. In 2025 Amazon earned US$8.27 of advertising revenue for every US$100 of merchandise sold. That advertising revenue depends on someone looking at the shelf — sponsored placements, browsing, comparison, the whole retail-media layer that sits on top of the product listing. An agent that goes straight to the SKU it was told to buy and skips the shelf entirely produces none of that. It's not a customer in the way Amazon's business model needs a customer to be. It's closer to a competitor's checkout running on Amazon's infrastructure for free.
Shopify's economics point the other way. Shopify's entire revenue is US$3.07 per US$100 of merchandise sold — it's paid on the transaction itself, not on attention sold around it. An agent that completes a purchase is doing exactly what Shopify is paid to enable. So Shopify built Shop Pay agentic checkout with Meta, announced the day after Amazon's block, and its shares rose 7 per cent on the news.
Same agent, same week. One platform is paid to sell attention and the agent doesn't buy any. The other is paid to process transactions and the agent generates one. The decisions aren't really about Muse at all. They're about what each business is paid to do.
Why I think the others follow anyway
Muse going to number one, above ChatGPT, is the clearest signal yet that a personal agent is the right shape for this — not a better search box, not a smarter recommendation engine, but something you hand a job to. People reached for it immediately. I think the other large providers ship their own, and I think the recent evidence against that says less than it appears to.
OpenAI retired in-chat Instant Checkout in March 2026, about five months after launch, after Walmart found conversion inside the chatbot running roughly three times below what its own site achieved. That is a real result. It measured checkout inside a chat window, with one particular design, at a point when nothing comparable to Muse existed to set expectations. A conversion rate tells you an implementation did not work. It does not tell you people don't want an agent to do the shopping. Those get conflated, and they are not the same finding.
Anthropic went a different direction entirely. Claude Commerce Agents, released 2 September 2026, is an Apache 2.0 blueprint for merchants to run their own shopping and operations agents — handing off to the store's own checkout rather than offering a consumer wallet that competes with it. Accenture, Visa, Mastercard and Shopify are named adopters. Read as a verdict on consumer agents, that looks like a retreat. Read as a position, it is closer to a hedge: if consumer agents do win, every merchant running one of those blueprints still needs something on the other end of the conversation.
The form factor is settled. Who owns it is not. A competitor sitting at number one above your own app is a difficult thing to leave unanswered for long, and the withdrawn bet was withdrawn on a narrower question than the one Muse has just asked. I expect more personal agents, not fewer, and I expect the next wave to handle checkout better precisely because the first one did not.
The identification question is already answered
One thread running under all of this: can a site actually tell what's requesting a page — a person, or an agent, and if an agent, which one? That question already has an answer. Web Bot Auth, the IETF draft standard led by Cloudflare, lets an agent cryptographically prove its identity. Google has signed its crawlers with it since June 2026. Shopify since May. OpenAI signs its agentic requests too.
The infrastructure for an agent to say who it is, credibly, already exists and is already in production use by several large platforms. Amazon's complaint that Muse didn't identify itself isn't a gap in the technology — it's a gap in whether Meta used it. Which means the interesting question for anyone running a website isn't whether agents should identify themselves any more. That's settled. It's what your site does when one shows up and says so.
This isn't only a shopping story
All of the coverage has been about consumer retail — Muse, Amazon, Shop Pay. But the same two questions apply to any business that sells to other businesses. Ask them directly rather than filing this under "interesting AI news."
First: if your buyer sends an agent to place an order, request a quote, or check stock, can it actually complete that job on your site — and which of your flows would it fail on? The agent-readiness audit I wrote in August is built to answer exactly that. Since then the answer has acquired a price.
Second, and this is the one Amazon's decision doesn't map onto cleanly: are you paid on attention, or on the transaction? Almost every B2B operator I can think of is paid on the transaction, the relationship, or the contract — not on advertising sold around a product listing. Which means an agent completing a purchase on your site is a customer, in the same way it's a customer on Shopify. Blocking it, the way Amazon did, is a defensible move for a business whose revenue depends on eyeballs on a shelf. It's a strange move for a business that gets paid the moment the order goes through.
The audit itself is cheap and quick. In August it was preparation for something coming. Two platforms have now priced it, in opposite directions. Which of the two does your business actually resemble?
Amazon and Shopify did not reach different answers because one understands agents better. They reached different answers because they are paid differently, and each read its own economics correctly. The businesses that handle this well over the next few years will be able to say plainly what their systems are for, and to decide on that basis what gets to act on their behalf.
If you want to work out where your own site sits on this split, start a conversation with us.