Amazon shut an agent out, and Shopify let one in. No merchant on either side made the call.
On September 8, 2026, Meta began rolling out Muse, a personal AI agent, in the United States. The same day, Shopify added Meta to Agentic Storefronts as an AI channel. Products were shared with Meta by default through Shopify Catalog. Direct checkout was on by default for eligible stores, US buyers only, with no fee beyond standard payment processing. No merchant installed anything, approved anything, or was asked anything.
Twelve days later, Amazon blocked the same agent. It said Meta never told it Muse would access the store, that the agent does not identify itself while browsing, and that it appears to capture and store customer credentials. The next morning, Shopify CEO Tobi Lütke announced a deeper arrangement: agentic checkout with Shop Pay across all Shopify stores. Mark Zuckerberg confirmed it the same day. Shopify shares rose about 7%.
The coverage read this as a split between an open platform and a closed one. That is the surface, and the surface is the less interesting layer.
Both platforms made the same kind of decision. Each answered, on behalf of every brand underneath it, whether an agent may buy there. Amazon said no for its sellers. Shopify said yes for its merchants, by default, twelve days before anyone wrote about it. The channel decision has moved above the brand.
What the default actually turned on
Shopify Catalog syndicates a defined set of fields to AI channels: title, description, options, images, price, availability. That list is the whole surface. An agent's read of a brand is that list and nothing else. Not the theme, not the photography, not the product page copy that went through three rounds of review.
So the question a merchant should have asked on September 8 is not whether to allow the channel. It is what the channel reads when it arrives. Color, material and fit that live inside a description paragraph are invisible as attributes. They are prose to a human and noise to a parser. The same product with those values in structured fields is a different product to an agent, and it is the only version of the product Muse will ever see.
Turning it off is available under Sales channels, then Agentic, and takes up to seven days to process. It does not remove the brand from agent-mediated discovery. Products stay findable through crawling and indexing the same way they are in conventional search. The toggle governs syndication, not visibility. There is no setting for being read badly.
The protocol story is less settled than the announcements
American Banker reports that Muse searches Shopify's catalog against a shopper's prompt and completes checkout through the Universal Commerce Protocol, the full-lifecycle checkout standard Shopify and Google co-developed. Meta's own launch ran Stripe's Link wallet, with Shop Pay described as coming. Shopify has said Shop Pay already works in Muse. The September 21 announcement reads as the Shop Pay path being formalized rather than invented, and Meta declined to comment.
Hold that against the adoption record. A scan published by Originality.ai in May 2026 found 26 public sites carrying Universal Commerce Protocol files out of more than three million scanned, four months after the protocol's debut. A standard endorsed by twenty-plus retailers and implemented at twenty-six addresses is currently describing platform-to-platform plumbing, not merchant adoption. Shopify merchants inherit compliance because Shopify implemented it. That is a genuine advantage, and it is not the brand's own infrastructure. A brand that leaves Shopify, or sells anywhere else, leaves the compliance behind.
Amazon's block is the same decision pointed the other way
Amazon's position is not anti-agent. Its own Buy for Me feature purchases from external brand sites, and Amazon notes that it identifies itself and lets brands opt out. The objection is to an unannounced agent operating inside customer accounts.
The legal context matters here. In August 2026, the Ninth Circuit vacated the injunction Amazon had won against Perplexity, on the reasoning that it was users, not the agent vendor, who were accessing the site. Amazon lost the computer-access argument and moved to the lever it still holds, which is terms of use enforced server-side.
That is the shape of the next two years. Not litigation over whether agents may shop, but platform-level admission decisions: fast, unilateral, and made for everyone underneath.
What it costs, and when
Meta direct checkout carries no fee today beyond standard payment processing. Zuckerberg has said he expects Muse to earn a share of the commerce completed through it, with a large free-compute allowance seeding adoption first. The comparison is already on the table: The Information reported that Shopify merchants pay OpenAI a 4% fee on sales through ChatGPT checkout.
Free distribution that later carries a take rate is the ordinary shape of platform economics, and it is not a reason to stay out. It is a reason to model it. The planning question is whether a brand's margin structure survives a mid-single-digit platform fee on a channel it did not choose to enter and cannot currently measure.
The measurement problem is the one nobody toggled
Shopify surfaces Meta channel performance inside its agentic settings. Most merchants have never opened that screen.
A brand that cannot separate agent-sourced orders from direct orders in its own reporting cannot price the take rate when it arrives, cannot tell whether structured attributes changed its selection rate, and cannot tell a board what share of revenue now comes through agent-mediated buying. The channel opened on September 8. The ability to see it did not open with it.
What to do this week
Three things, in this order.
Open Sales channels, then Agentic, and decide deliberately what is already true by default. The decision may well be to leave everything on. Leaving it on knowingly is a different position from never having looked.
Audit whether the attributes that decide selection exist as structured fields rather than description prose, product by product. That is the Agent Catalog Read question, and it is answerable this afternoon on a sample of ten products.
Confirm that agent-sourced orders are visible by channel in your own reporting and not only in the platform's.
On September 8, a channel opened for roughly a million merchants, and none of them decided to use it. The brands that do well in agent-mediated buying will not be the ones that picked the right platform. They will be the ones whose product data was correct before a platform picked for them.