Four days before Cyber Week opens on November 24, Amazon is running two agentic commerce policies at the same time, and they point in exactly opposite directions. On the retail front, Amazon has blocked AI shopping agents from Google, Meta, and OpenAI from its consumer website and filed suit against Perplexity over its Comet browser’s autonomous buying features. On the cloud front, announced at re:Invent 2026 and reported the morning of November 21, AWS Marketplace is doing the opposite: it has opened agentic purchasing through the x402 payment protocol, wired transaction handling through Amazon Bedrock AgentCore, and begun selling data products through MCP servers that agents can query and buy without a human in the loop.
Same company. Same month. Same technology class. One business unit treats autonomous buyers as a threat to be litigated out of the funnel. The other treats them as customers worth building new rails for.
The contradiction looks like hypocrisy, but it is more useful read as a map. Agent commerce is being permitted first where listings are standardized, prices are contractual, and the buyer is effectively a procurement system: software, APIs, data. It is being resisted hardest where the shelf is an ad auction, the reviews are unaudited, and the ranking is the moat. That is the consumer product page, and the week’s best survey data explains what happens when agents are sent into it anyway.
The cloud front: where agents are invited to buy
The AWS Marketplace announcement is the most complete agentic purchasing stack any major platform has shipped. Mya Papolu, product marketing manager for AWS Marketplace and Partner Services, described the change plainly: “Buying software and data through AWS Marketplace just got easier, for you and the AI agents working alongside you.”
The mechanics matter, because they show what a serious agent purchasing channel assumes. Sellers must register with AWS Marketplace and integrate the x402 software development kit before any agent can buy from them. Transactions settle in US dollars through AWS billing, which means existing procurement controls, approval workflows, and audit logs apply to agent-initiated spend. The x402 protocol itself was introduced by Coinbase and Cloudflare, handed to the Linux Foundation earlier this year, and is now supported by AWS, Google, Stripe, and several blockchain networks, with Brave Software adding x402 support to its search API in November 2026.
AWS Marketplace is also expanding beyond software into data-as-a-service, with listings reachable by human teams and agents alike through the website, APIs, and MCP servers. The initial catalog is telling: Moody’s Credit MCP server for financial research, Reuters Geopolitical News for AI-structured reporting, US building permit data from Shovels, and a conversational acoustic corpus from Uber AI Solutions. Every one of those products is a standardized, verifiable asset with a defined license and a deterministic price. An agent that buys the Moody’s server knows precisely what it bought.
That is the precondition Amazon’s cloud side accepts and its retail side refuses to guarantee: that what the agent sees in the listing is what the buyer gets.
The retail front: where agents are stopped at the door
The consumer side of the same company spent this month doing the reverse. Amazon has blocked AI shopping agents operated by Google, Meta, and OpenAI from its retail site, and its lawsuit against Perplexity over the Comet browser escalated the blockade from robots.txt to federal court, as Forbes framed it this week in a piece asking whether agentic commerce can scale at all without the largest retailer’s participation. Amazon did not respond to a request for comment on the difference between its retail bot restrictions and its AWS Marketplace agentic commerce program.
The stated stakes are traffic and advertising economics. The unstated stakes are trust. A retail product page is a synthesized persuasion surface: sponsored placement in results, sponsored brands on the shelf, review scores that aggregate verified and unverified purchases alike, and star ratings that no independent party audits before display. An autonomous agent does not skim past the sponsored tile the way a distracted human does. It either ingests the ranking as ground truth or it needs an outside source of truth. Amazon blocking agents protects the surface from a reader that would take it literally.
Visa builds the identity layer, 140 companies at a time
While Amazon split itself in two, Visa spent the week assembling the plumbing everyone else will run on. On November 17, Visa expanded its Agentic Ready program from issuers to the acceptance ecosystem: merchants, payment service providers, acquirers, and B2B platforms. The expansion launched with more than 140 participants, including Adyen, Antom, Billtrust, Checkout.com, Fiserv, Global Payments, PayPal, and Stripe, joining the 175-plus issuers already participating across 45 markets.
The program’s purpose is to test how agent-initiated transactions can be recognized, authenticated, and processed through infrastructure that already exists. Rubail Birwadker, SVP of Growth Products and Partnerships at Visa, put the problem in one sentence: “Agentic commerce is already taking shape, but scaling it requires more than innovation. Merchants need confidence in who they’re transacting with, and payment providers need to know these transactions can move securely and reliably through existing infrastructure. No one company can solve that alone.”
Note precisely what is being verified. Agentic Ready, like Web Bot Auth and the permission layers that shipped earlier this month, answers: is this agent allowed to act, and can its money move? Nothing in the 140-company program answers the question that comes before the transaction: is the product it is about to buy any good, and is the price honest?
The error data: agents fail on evidence, not on payments
The sharpest numbers of the week came from Salesforce, whose Consumer Shopping survey of more than 2,500 holiday shoppers across the US, UK, Canada, and Australia was released November 12 and re-analyzed on November 20 by Ecommerce.com. The survey found 31 percent of consumers have already tried a personal AI agent, Meta’s Muse leading at 12 percent adoption, with 28 percent of agent users dispatching tasks multiple times a day. But the error taxonomy is the story:
- 27 percent of personal agent users say their agent has bought something that was too expensive
- 22 percent have had an agent blocked from completing a purchase
- Only 32 percent report no major errors with their agent at all
- Most shoppers abandon an AI assistant after just two or three mistakes, and when AI makes a bad recommendation, 32 percent of shoppers blame the AI itself
Run the arithmetic on the middle line: 68 percent of people who have actually used a personal agent have experienced at least one major failure. And look at what the failures are. “Too expensive” is a value judgment failure, which means an evidence failure: the agent could not tell whether the product justified its price, which is impossible to do when the quality signals it consumes are unaudited. “Blocked at purchase” is an access failure, which is Amazon’s policy showing up in the data. Neither of the top failure modes is a card decline or a token misfire. The payment layer, the piece 140 companies are now jointly testing, is not what is breaking.
Trust asymmetry compounds it. Only 19 percent of personal agent users have deployed an agent at checkout, 73 percent of all shoppers are uncomfortable giving an agent payment information, 49 percent are not interested in agents buying on their behalf at all, and just 9 percent of agent users grant unrestricted financial access. Another 36 percent have set up a separate credit card purely for their agent, a behavioral hedge that says the same thing the error data does: people expect agents to make mistakes that cost money.
Duran Inci, CEO of Ecommerce.com, reading the same dataset, landed on the human constant underneath it: “Shoppers who hand a purchase to an agent still want the same thing they always wanted, which is the right product at the price they expected.”
The numbers that say build anyway
None of this friction is slowing the direction of travel, and Salesforce’s Cyber Week 2026 forecast, built on the Shopping Index’s view of 1.5 billion shoppers and 1.5 trillion page views, is the scale backdrop:
- Cyber Week (November 24 through 30) is expected to drive a record 357 billion dollars in global online sales, 84 billion in the US
- 75.8 billion dollars of it, 21 percent, will be influenced by AI and agents, up 13 percent from 67 billion in 2025
- AI-driven traffic to retail sites is up 104 percent year over year, and 20 percent of holiday shopping traffic will be nonhuman
- Shoppers arriving via AI referrals convert at more than twice the rate of typical ecommerce visitors
- 39 percent of shoppers have used a brand’s own shopper agent; 80 percent reported a good experience, 75 percent of that group felt more loyal afterward, those shoppers spend 8 percent more per order and buy 23 percent more items
- Brands deploying AI shopping agents show seven times faster sales growth than brands without them, and one in three retailers will have shopper agents live by Cyber Week
Caila Schwartz, Head of Agentic Commerce Shopper Insights at Salesforce, distilled the operating advice: “Be the brand that AI can find and humans can trust. Shoppers are starting with AI this season, so make sure your product data is accurate, your prices and inventory are current, and your own agent is ready to answer the hard questions.”
Why agents overpay: they can verify price but not worth
The single most revealing statistic in the Salesforce survey is buried in the adoption data: the top use case for personal agents is price comparison. Not checkout, not discovery, not support. Price. That is the same pattern we traced earlier this week: price has become machine-verifiable in real time, discount claims are now auditable against history, and DHL found nearly nine in ten Black Friday shoppers using AI precisely there.
Agents optimize toward what they can verify. A price is a number, checkable against a tracker, a lookback window, or a competitor’s feed. A product’s quality is a 4.8-star aggregate built from a review corpus nobody audits, where fake and incentivized reviews are priced by the market and detection runs behind manipulation. So agents compare the one thing they can prove and guess at the one thing they cannot, and 27 percent of their users report the guess was too expensive. Overpaying is what an evidence gap looks like when you close your eyes and buy anyway.
This is the exact seam between the infrastructure layer that shipped this week and the trust layer that has not. Visa authenticates the agent. AWS standardizes the listing. Poppy gates the permission. What verifies the product?
The missing layer has to sit before the checkout
At GoBuy, that question is the product. GoBuy is the trust layer that runs before the buy: fake and manipulated reviews are filtered out, authentic ones are weighted up, and every product gets a Smart Score from 0 to 100 computed from the quality of its review evidence, not the quantity. Quantity is the metric that review farms inflate; quality is the metric they cannot. Instead of a thousand undifferentiated listings, GoBuy surfaces only the top seven products in a category, because the point of an agent is to compress the shelf, not to re-create it.
That intelligence is exposed where agents already live. The GoBuy MCP server lets any MCP-capable agent, the same architecture AWS Marketplace just legitimized for data purchases, query filtered review analysis and Smart Scores before committing a single dollar of its user’s budget. Products that hold a score of 80 or above across 90 days earn the GoBuy Verified badge, a standing, auditable claim an agent can cite rather than a star average it must take on faith. And for the human still holding the card, which on this week’s data is 73 percent of shoppers, the Chrome extension injects the trust panel directly onto the Amazon product page, before checkout, where the decision actually happens.
Cyber Week will set records with or without the trust layer: 357 billion dollars does not wait. But the shoppers who abandoned their agents after two mistakes, the 27 percent who overpaid, and the 68 percent who logged a major error are the retention problem of agentic commerce, and none of it traces back to the card network. Payment is solved. Identity is solving itself across 45 markets. The open frontier is product evidence, and the first platforms to make agent purchases verifiably trustworthy will own the 21 percent of Cyber Week that AI already touches.
Ask what your agent reads before it buys. If the answer is a star rating, ask again at gobuy.ai, and give your agents the evidence endpoint they are missing at gobuy.ai/agent-docs.