In a single Wednesday afternoon keynote, Meta built more of agentic commerce than the payments industry assembled in the previous two years. At Connect 2026, Mark Zuckerberg and chief AI officer Alexandr Wang announced that Muse, the personal agent launched in early September, can now reach the entire Shopify product catalog agentically, complete purchases with Lync and Shop Pay, lean on expanded PayPal support, and connect directly to Best Buy, Gap, Sephora, Walmart, Wayfair and Expedia, with Instacart groceries coming soon. Muse will live on Meta’s smart glasses, where it can, in the company’s words, “help them buy products they see.” It will soon run a user’s Mac, operate any app on it, and keep working after the user walks away. It is getting its own email address.

Meanwhile, the adoption numbers are no longer hypothetical. Muse has topped the Apple App Store ahead of ChatGPT, with more than 2.5 million downloads in its first two weeks according to Sensor Tower data shared with CBS News. Over the same launch period, ChatGPT recorded 3.1 million global downloads and Anthropic’s Claude about 200,000. Meta’s market positioning is unambiguous. “The centerpiece of our vision for what we’re building is Muse,” Zuckerberg said, predicting the agent will grow into “the personal superintelligence that billions of people around the world are going to use.”

The transaction stack is shipped, in other words. What Connect did not assign to anyone is the layer underneath it: whether the thing being bought deserves to be bought.

What Connect actually shipped, layer by layer

GoBuy’s framework, published last week, decomposes agent trust into five layers: Accessibility, Understanding, Evidence, Transaction and Accountability. It is striking how neatly the Connect announcements tile four of the five.

Accessibility. Retail connectors to Best Buy, Gap, Sephora, Walmart and Wayfair, plus Expedia and Instacart, mean Muse reaches inventory that was previously locked behind apps and interactive front ends. Meta also opened its connector platform to third-party developers and received more than 1,500 applications in under a week, which Wang called “a generational opportunity to start building for a new platform.”

Understanding. Catalog access via Shopify means structured product data: prices, variants, availability, machine-readable from the source. This is the layer most legacy retail still fails, and it arrives in Muse essentially for free.

Transaction. Shop Pay, Lync, PayPal and a Stripe partnership give Muse checkout rails across what Wang described as “almost anything on the internet.” This is the layer the payments industry spent 2026 building, from Visa and Mastercard agent-pay products to the bank principles papers we covered yesterday. It is now, functionally, a solved problem for a top-of-App-Store agent.

Accountability, partially. Mashable reports each Muse agent runs on a dedicated Muse Secure VM, a cloud computer separate from the user’s Mac, and that Muse requests approval before completing sensitive actions such as sending emails or making purchases. Approval gates are the classical accountability control. It is not decision provenance, but it is a start.

Four layers, one keynote, one company. Then there is the fifth.

The evidence layer is the one nobody owns

Evidence is the question of whether the signals around a product support the listing’s claims: whether reviews were written by customers or commissioned by sellers, whether the seller has a real track record, whether the price history is honest. On September 23, the day of the keynote, no Connect announcement addressed it. Not a partnership, not a connector, not a data source.

This would be a gap under any circumstances. It is worse now, for one specific reason: Amazon has blocked Muse.

Amazon confirmed to CBS News that it has barred the agent from its marketplace, citing security and user-experience concerns, extending the access war we covered when it started. Whatever the corporate motives, the structural consequence is that the internet’s largest review corpus is now unreadable to the top consumer agent in the West. Muse’s purchase decisions will be made on the open web.

The open web is where review enforcement is weakest. Amazon, for all its problems, operates the most aggressive fake-review enforcement apparatus in commerce; the company says it blocked hundreds of millions of suspected fake reviews in 2025 and took legal action that shut down more than 100 fake-review facilitation websites, as Reputon’s 2026 review-checker analysis documents. Independent Shopify storefronts, aggregator sites and the long tail of retail have nothing comparable. An agent that cannot read Amazon and shops the open web instead is an agent making purchasing decisions inside the least-policed review environment in e-commerce. The walled garden’s moat was also, incidentally, its fraud filtering.

The fee problem: when the agent operator bills the purchase

The most consequential sentence of the keynote was not about connectors. It was the business model. “We’re standing behind this by making Muse free for a huge number of tokens,” Zuckerberg said, “with the expectation that over time we will profit by taking a small fee from transactions.”

Set aside intent; look at structure. A token-subsidized agent whose revenue is a function of completed purchases has a financial interest in purchases completing. Every design decision downstream of that fact, from how aggressively Muse proposes a purchase to how it ranks transactable versus non-transactable inventory, now sits under a billing event. The merchant whose products are inside the connector network is one payment rail away from the agent’s user. The merchant who is not does not exist.

The next day, half a continent away, Mastercard’s Gaurang Shah described to PYMNTS exactly the dynamic this creates. Consumers are already using agents for “discovery, ranking, comparing,” but “they are definitely stopping at the recommendations.” Merchants, Shah said, want to be “discoverable,” “trusted” and “transactable,” but want to keep their pricing power and customer relationships. PYMNTS CEO Karen Webster named the endgame: a shopping agent whose simplest optimization is “find the cheapest acceptable product” turns agentic commerce into “a race to the bottom,” and merchants already argue that liability “should rest with whoever the agent broker is.”

Note that phrase: agent broker. The industry’s own language has conceded that the operator of the agent is not a neutral instrument. A broker with a fee interest in closing is a salesperson. The difference is that a human salesperson’s customer can see the salesperson. Muse’s user sees an avatar named Jolly.

The connector gold rush is the new shelf

The 1,500 connector applications in under a week are the tell that developers understand what Muse connectors really are: placement. When the agent is the front end, connector presence is the new page-one ranking. A connector is also, structurally, a marketing artifact: it is built with the merchant, it carries the merchant’s own product claims, and it has no obligation to third-party verification. Nothing about a Best Buy or Sephora connector tells Muse whether the product’s reviews are authentic, whether the discount is real, or whether the seller’s history supports the listing’s promises.

Combine the three forces now in motion on this single platform: a curated connector shelf that determines what the agent can see, a fee model that rewards the agent for completing transactions, and no independent evidence layer anywhere in the decision path. That is not a trust architecture. It is a conversion funnel with a friendly face.

Consumers sense it, which is why they stop at the recommendation

The usage data explains why this matters commercially rather than just ethically. Per IBM’s Institute for Business Value, cited by CBS News, 41 percent of consumers already use AI assistants to research products and 31 percent use them to hunt deals. A Coresight survey of more than 1,000 consumers found 30 percent would use AI tools to compare prices for holiday shopping and the same share to find deals or promo codes. GlobalData’s Neil Saunders called Muse’s momentum “the acceleration of agentic shopping” while adding the qualifier that defines the market: “People are still a little bit nervous about allowing AI agents to do the purchasing on their behalf.”

That nervousness is not a UI problem or a payments problem; those are solved or being solved. It is an evidence problem. Yesterday’s Global Payments report quantified it: 42 percent of consumers fear the agent simply buys the wrong thing, and 26 percent demand proof that AI systems cannot be manipulated. The Global Payments survey put a hard $100 ceiling on what Americans will let an agent spend in most categories. Consumers are pricing the downside of a choice they cannot personally examine, and the price is low because nothing verifies the choice for them.

Lift the ceiling and you lift the revenue on every transaction Meta hopes to tax. The evidence layer is not a compliance nicety bolted onto agentic commerce. It is the constraint on the whole P&L.

What a real evidence layer looks like

The test for any agentic platform is simple to state: does the agent consult anything the seller does not control before paying? If the answer is no, every recommendation is an ad with extra steps.

A real evidence layer has three properties:

  • Independence. Scores derived from the review corpus, seller history and price behavior, computed by a party with no stake in the transaction closing. GoBuy’s Smart Score, a 0-to-100 rating built on review authenticity rather than review volume, is exactly this: the difference between a product with ten thousand purchased reviews and a product with four hundred earned ones.
  • Brevity with accountability. Agents do not browse; they consume ranked evidence. GoBuy returns only the top seven verified products per category, not thousands of sponsored slots, with score provenance attached so the decision can be audited after the fact. Products scoring 80 or above for 90 sustained days carry the GoBuy Verified badge, which is precisely the kind of durable, slow signal a fee-motivated connector cannot fake in a launch week.
  • Open access for agents. All of it is exposed over MCP at gobuy.ai/api/mcp, so any agent, Muse connector developers included, can consult independent product evidence before checkout. The docs live at gobuy.ai/agent-docs. For the human half of the workflow, the GoBuy Chrome extension overlays the same trust panel directly on Amazon product pages, filtering manipulated reviews and surfacing the authentic ones.

Meta opened Muse’s connector platform to 1,500 eager applicants. Not one of them is incentivized to tell Muse the truth about a product. That is the slot an independent evidence connector fills, and until something occupies it, the most-used shopping agent in the West will be buying almost anything on the internet on the strength of claims made by the people selling it.

Before your agent buys, make it check. GoBuy is the trust layer before purchase: Smart Score 0-100, fake reviews filtered, authentic reviews weighted, only the top 7 products shown. Start at gobuy.ai, or wire the evidence layer into your own agent at gobuy.ai/agent-docs.