Amazon Blocks Meta's Muse: What the Fight Over AI Shopping Agents Means for Your Business

Two weeks ago Meta launched Muse, its entry into AI shopping agents: a personal assistant that shops, books and negotiates on behalf of its users. On Sunday night Amazon locked it out. Anyone who asks Muse to buy something on Amazon.com now gets an error message instead of a parcel, and the wording is blunt: continued access by an unauthorized AI agent violates Amazon's Conditions of Use.

On the surface it looks like a quarrel between two giants, and partly it is. But the block is also the clearest signal so far that the rules of agentic commerce are being written right now, in public, by whoever moves first. AI shopping agents are no longer a demo on a conference stage. They sit in millions of pockets, and the question of who lets them in has direct consequences for anyone who sells online.

In this post we look at what actually happened, why Amazon did it, and what the standoff means for businesses that run their own stores and digital products. Let's sink in!

What Actually Happened Over the Weekend

Meta released Muse on 8 September 2026, and within a week it was the number one free app on Apple's App Store. The agent works by driving a browser on the user's behalf, which lets it operate on services that never offered it an API. That includes Amazon, where Muse could browse products, read a user's order history and complete purchases like a very fast, very patient human.

On the evening of 21 September Amazon switched that off. Shoppers who sent Muse to Amazon started seeing a notice that unauthorized AI agents violate the site's Conditions of Use. Amazon's spokesperson put the position on the record: third-party applications that make purchases on behalf of customers "should operate openly and respect service provider decisions about whether or not to participate."

Meta declined to comment on the block itself. Earlier the company said Muse "has no visibility into people's passwords or payment methods" and that credentials are stored securely. Amazon clearly does not find that reassuring, and it did not wait for a longer debate.

Why Amazon Shut the Door

Amazon's stated complaints are specific. Meta never told Amazon that Muse would access the platform. The agent does not identify itself while browsing, so it is indistinguishable from a normal customer session. And according to Amazon, it appears to capture and store customer credentials, which creates privacy and security exposure on account pages and order history that Amazon would end up answering for.

There are quieter business reasons too. Amazon sells its own AI stack through AWS, and it has its own shopping agents: Alexa for Shopping launched in May, alongside the Buy for Me feature. An external agent that picks products for the customer also skips something Amazon cares about a lot, which is advertising. Amazon made $68.6 billion in ad revenue in 2025, up 22 percent on the year before, and an agent that jumps straight to the best-matching product never looks at a sponsored listing.

And there is plain operational risk. Muse is considered one of the more accurate agents of its generation, with one of the lower hallucination rates on the market, but a low error rate is still pretty far from zero. When an agent orders the wrong size or the wrong quantity, the return, the refund and the angry customer land on Amazon and its sellers, not on Meta.

The Court Case That Changed the Rules

The interesting part is why Amazon is blocking with error messages rather than lawyers. It already tried the legal route. Amazon sued Perplexity over its Comet browser, which shopped on Amazon in a similar way, and in March 2026 it won a preliminary injunction. Then in August the Ninth Circuit overturned it, and on 10 September the court declined to reconsider.

The reasoning matters for everyone in e-commerce. The court took the view that under the federal anti-hacking statute it is the user, not the AI company, who accesses the system. If a customer with a legitimate account tells an agent to buy something, that is the customer acting, just with a different tool. After that ruling, suing agent makers became a much weaker weapon, so platforms are falling back on technical blocks and terms of use.

Amazon has form here: it has also blocked shopping agents from Google and OpenAI. And it is starting to formalise the policy. From 28 October Amazon Japan will require automated agents to identify themselves in their HTTP requests and will prohibit using Amazon data for AI training. That is a preview of the kind of rulebook other markets will likely get.

Four Camps and No Rulebook

Right now agentic commerce has four competing visions and no shared standard. Amazon blocks outside agents while building its own. Google promotes open standards that would let any agent identify itself and transact. OpenAI built a checkout protocol together with Stripe. Meta simply wants its agent inside platforms it does not own, and it is finding out what that costs.

Consumer trust is not evenly spread either. In one survey only 8 percent of US consumers said they would trust Meta with their passwords, against 30 percent for Google. Whoever wins the standards fight, users will still choose agents partly on reputation, and merchants will feel that choice in their traffic.

The strange footnote is that Amazon and Meta stay close partners elsewhere. Amazon products have been purchasable inside Facebook and Instagram since 2023, and in April 2026 Meta signed a multibillion-dollar deal to run AI workloads on Amazon's Graviton chips. This is not a cold war between the companies. It is a fight over one specific thing: who controls the moment a purchase happens.

Should Your Store Welcome or Block AI Shopping Agents?

Amazon can afford to slam the door because it has its own agents, its own ad business and more customer habit than anyone else. Most businesses are not in that position. For a typical online store, an agent that arrives with a paying customer behind it is simply a sale, and blocking it means turning away revenue to protect nothing in particular.

That does not mean the concerns are imaginary. Agents can distort your analytics, trip up A/B tests, hammer your search, and place orders that a confused customer later disputes. If your margins depend on upsells, bundles or ads shown during browsing, an agent that goes straight to checkout changes your economics the same way it changes Amazon's, just at a smaller scale.

The practical point is that this should be a decision, not an accident. Many bot-protection and firewall tools quietly block automated browsers by default, which means some stores are already refusing agent orders without anyone having chosen that. Check what your protection layer actually does, look at your logs for agent traffic, and decide on purpose which way you want to lean.

Getting Your Digital Products Ready Either Way

Whatever policy you pick, the preparation looks similar, and most of it is just good engineering. Agents read websites the way accessibility tools and search crawlers do, so clean semantic HTML, working forms and stable checkout flows matter more than visual polish. A product page built from unlabeled divs with prices rendered in JavaScript five seconds after load is hard for an agent, and honestly it is hard for Google too.

Structured data is the cheapest win. Proper Schema.org markup for products, prices, availability and shipping gives agents a reliable source to read, which lowers the odds of a wrong-item order that turns into a support ticket. A machine-readable product feed does the same job even better, and if agent traffic becomes meaningful for you, a simple ordering API means agents stop scraping your checkout entirely.

Then there is the housekeeping. Update your terms of service to state your position on automated purchasing, whichever it is. Separate agent traffic in your analytics so your conversion numbers stay honest. And keep an eye on the identification standards emerging from Amazon Japan's rules and Google's proposals, because self-identifying agents will be the norm soon, and your site should know how to greet them.

Final notes

The Muse block will not be the last move. The courts have leaned one way, the platforms are pushing the other, and somewhere in between a standard for agent identity and agent checkout will settle. Businesses do not need to predict the winner. They need their stores to work well for humans, stay legible to machines, and have a deliberate answer ready when an agent knocks.

At Davydov Consulting we build web platforms and e-commerce systems that hold up as the channel mix shifts, whether the buyer is a person, an assistant or something in between. If you are wondering how agent traffic is treating your own store, that is exactly the kind of question worth answering before it answers itself.