The most consequential commerce number published this month did not come from a retailer, a payments network, or a regulator. It came from a social media company’s earnings call.

On Wednesday, October 21, Meta reported third-quarter results of $64.6 billion in revenue, up 26 percent from $51.2 billion a year earlier, with net income up 264 percent to $9.9 billion. All of that was expected. What was not expected was the sidebar about shopping. As CNET reported, Mark Zuckerberg told investors that Meta Muse, the personal AI agent launched in early September, has surpassed 10 million engaged users, and that more than a quarter of them are already using it to shop, with an average purchase size of over $200. TechCrunch confirmed the figures and confirmed shopping is a top activity, not a novelty edge case.

Pause on what that means. Within roughly seven weeks of launch, an AI agent has become a purchasing channel with a user base larger than most national retailers’ loyalty programs, moving baskets bigger than the average e-commerce order, at volumes that broke the assumptions of the payments infrastructure underneath it.

The payment rails bent and did not break. The question this piece asks is the one nobody on the call asked: what is the agent reading when it decides what to buy?

What Zuckerberg actually told investors

Four disclosures from the call deserve to be quoted precisely, because each one marks a milestone for agentic commerce.

The scale number. Muse has more than 10 million engaged users, within weeks of an early-September launch that topped the U.S. free-app charts almost immediately. Over a quarter of those engaged users shop through it. Run the arithmetic conservatively: two and a half million shoppers at $200 per purchase is $500 million of monthly volume if each buys just once a month. That is not a reported figure; it is what Meta’s own numbers imply at the floor of the range, before repeat purchasing, before the holiday quarter, before the smart glasses integration Zuckerberg said arrives next month.

The payments shock. “Plaid told us that they hadn’t seen activity like they’re seeing with their connector on any other service before,” Zuckerberg said, referring to the financial infrastructure company that connects Muse purchases to bank accounts. Volume ran so far past forecasts that Meta had to work with Stripe to raise credit limits on the agent’s payment rails just to keep demand flowing. When was the last time a shopping feature, seven weeks old, required a payments partner to re-provision limits?

The strategic framing. Zuckerberg called Muse’s shopping usage “product-market fit” and predicted Muse will eventually earn more than Facebook, Instagram, WhatsApp, Threads and Messenger combined. He also sketched the end state: “It’s not just a kind of single-player game where you have your Muse, but making it so that your Muses can work with your friends’ and your families’ Muses… And if you want to buy something, it’s not just navigating to that business’s website; it’s talking to their agent.”

The distribution multiplier. Muse arrives on Meta’s smart glasses next month, which moves the agent from desktop novelty to an always-present layer over daily life, including the aisles of physical stores.

Every one of those four disclosures is about the transaction: scale, payments, strategy, distribution. None is about the evidence the agent consults before it spends someone’s $200.

The court already answered the question Amazon is fighting over

Amazon banned Muse from its site in early September, arguing that outside applications making purchases for customers “should operate openly and respect a retailer’s decision about whether to participate,” as Newsweek recounted. This is the second act of a fight Amazon has already run once. In November 2024, Perplexity launched Buy with Pro, and its chief business officer Dmitry Shevelenko insisted the feature was “just acting as an agent on behalf of the user.” Amazon answered Perplexity’s Comet browser with a technical block, which was routed around within 24 hours, then a cease-and-desist, then a lawsuit arguing Comet was “covertly” posing as a human customer. Amazon won a preliminary injunction in March 2026. In August, a federal appeals court vacated it.

The appellate reasoning is the part every operator in this market should read twice. The court held that Perplexity was not accessing Amazon’s website. Perplexity’s customers were, using the agent as an instrument. Tell an agent to buy something and you are the one buying, not the agent.

That ruling settles the consent question in favor of the agent ecosystem, and it quietly settles something larger: liability follows the human. The person who delegated the purchase owns the purchase, including every defect of judgment the agent made along the way. Newsweek’s analysis names the resulting gap precisely: every purchase involves information and the assurances behind it, and “an agent has nothing at stake in a transaction. When it gives a false answer, no one has broken their word because no one gave it.”

There is real money attached to the information half of that pair, by the way. Amazon’s advertising business, the machinery that monetizes human attention inside its storefront, generated $19.8 billion in the second quarter of 2026 alone. The blockade is not primarily about bots misbehaving. It is about who controls the evidence layer at the moment of choice, and who gets paid to rank it.

The corpus those agents are actually reading

So the agent is buying, the human is legally the buyer, and the question collapses onto the evidence. What is the evidence made of? Reviews. Still.

BrightLocal published its 2026 Local Consumer Review Survey the same week as Meta’s earnings call, and its headline finding is the quiet counterpart to Zuckerberg’s numbers: 97 percent of consumers still read reviews to guide purchase decisions, and 41 percent now read them every single time they look for a business, up sharply from 29 percent a year ago. The average consumer consults six different review sites. BrightLocal co-founder and CEO Myles Anderson called the dependency “permanent” and “cemented”: “Even in a world where people are more aware and more frustrated by the scourge of fake reviews, 97% of consumers still lean on reviews to guide their purchase decisions.”

Read those two datasets together and the picture is uncomfortable. Millions of purchasing decisions per month are now delegated to agents whose primary evidence base is the same review corpus that consumers themselves distrust, in an ecosystem where one industry compilation priced review fraud at $770.7 billion in global consumer harm for 2025. A human skimming reviews applies skepticism, fatigue and pattern recognition. A language model compresses the corpus into a confident recommendation. Contaminated evidence does not get discounted by the machine. It gets weighted.

There is a second, subtler finding in the BrightLocal data: Google’s share as the discovery surface for reviews fell from 83 percent to 71 percent in a single year, while AI tools like ChatGPT surged into third place as a source of local business recommendations. The reader is already changing faster than the corpus is being cleaned.

Business buyers, the cohort with procurement departments and compliance budgets, show the same split between information and assurance. Forrester surveyed nearly 18,000 of them and found 94 percent used generative AI during their most recent purchase, up from 89 percent the year before. Gartner surveyed 645 business buyers and found that while 67 percent prefer buying without a sales representative, 69 percent bring a human representative back in specifically to check what the AI told them. They trust the machine’s information roughly as little as a salesperson’s pitch, 51 percent to 49 percent, and they use it first anyway.

That is the division of labor for the next few years: agents supply the information, humans carry the risk. But Muse’s numbers say the consumer side is skipping the second half of that pattern. A shopper who tells Muse to “buy the good ones” is not convening a review committee. The agent is the committee.

The assurance gap, priced

Put the three datasets in one frame:

  • 10 million-plus engaged users, more than a quarter shopping, $200-plus average purchase: delegated purchasing at consumer scale, arriving years ahead of forecasts, with Stripe re-provisioning limits to keep up.
  • An appeals court ruling that the human is the buyer: every mispurchase, every manipulated recommendation, every hallucinated product claim lands on the person who delegated, not the platform that ranked or the agent that chose.
  • 97 percent review dependence, 41 percent reading every time, $770.7 billion in annual review-fraud harm: the evidence layer is the most heavily used and least verified input in the entire stack.

Payments solved the hard part of agentic commerce in a single season: identity, authorization, limits, settlement. Consent architecture is spreading fast, from Walmart’s Agent Profiles to agent wallets. What no earnings call, court ruling, or retail announcement has produced is an independent answer to the only question a buyer ultimately cares about: is this product what it claims to be, and are these reviews written by people who actually bought it?

A store cannot be that answer. A store’s ranking is a commercial surface; Amazon monetizes it at a $19.8 billion quarterly run rate, and every retailer with a media network is copying the model. The agent vendor cannot be that answer either, because the agent is the one whose word needs checking. The assurance layer has to sit outside both, or it is not assurance. It is marketing with better grammar.

Building the check into the loop

This is the specific, narrow problem GoBuy exists to solve: the trust check that runs before the transaction, independent of the store and the agent.

Concretely, that means review authenticity filtering that happens upstream of any score, so purchased praise and bot-written five-stars are excluded before they can be averaged into a rating. It means a Smart Score from 0 to 100 that weights review quality, not review volume, because a thousand low-content ratings should not outvote two hundred verified, specific, recent ones. It means showing a curated top-7 set instead of an infinite ranked feed, because an endless scroll is an advertising surface and seven vetted options is a decision. And it means an MCP endpoint at gobuy.ai/api/mcp that an agent like Muse, or any agent built on the model context protocol, can call in the same decision loop where it calls a catalog or a checkout: consult the trust layer before spending the customer’s $200, not after.

Products that hold a Smart Score above 80 across a 90-day window carry the GoBuy Verified badge, which is earned by sustained review quality and revocable, not purchasable. For the humans still making their own calls during this transition, the GoBuy Chrome extension overlays the same trust panel directly on the product page, so the person and the agent are at least reading the same evidence.

None of this slows the transaction down. It is one additional call in a loop that already spans authentication, catalog search and payment authorization. What it changes is who stands behind the recommendation: someone whose score cannot be bought, in a world where, as the court established, the buyer bears everything.

What to watch in the next ninety days

Three events will define the first agent-mediated holiday season, and each one stress-tests a different layer:

  • Muse on glasses in November moves agent shopping off the desk and into physical retail. Watch whether purchase volume accelerates faster than trust behaviors mature.
  • The Amazon-Muse escalation. The appeals court ruling was narrow; the blockade continues; Amazon’s own assistant carries sponsored answers in the main search bar. Watch for the first sponsored placement inside any agent-native result set. That is the moment the referee problem fully arrives in agent form.
  • Holiday-quarter disclosures. Meta will report Q4 in January. If Muse shopping grows and every disclosure is still about payments and scale, with nothing about evidence quality, that gap is the story.

The transaction layer of agentic commerce scaled in seven weeks. The evidence layer is still where it was in 2024: enormous, essential, mostly unverified, and now read primarily by machines. Until that changes, the fastest-growing purchasing channel in retail is a very confident reader walking through the world’s largest unfiltered comment section with your credit card.

Somebody has to check the products. That is the layer we run.

Before your agent checks out, have it check first: gobuy.ai for humans, gobuy.ai/agent-docs for agents, MCP at gobuy.ai/api/mcp.