By Thorsten Meyer | October 26, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A reporter at The Verge asked Meta’s new Muse agent a simple question: where are granola bars cheapest per piece? Muse came back with Sam’s Club, Walmart and Costco, and then admitted something important. Amazon might have had the lowest price, but Amazon had blocked it, so it couldn’t check.
That is the AI shopping agent story of autumn 2026 in one scene. Your next purchase may well be made by an AI. But only where the shop lets it in. Walmart has just built a front door for agents. Amazon has locked its doors and put up a sign. And behind the technical fight sits a much older question: who owns the customer, and who gets paid when a machine does the buying?
Key takeaways
- Meta’s Muse made agent shopping mainstream fast. Meta says more than 10 million people in the US and Canada set up Muse within about six weeks, shopping is the most common use, and the average purchase is over $200. These are Meta’s own figures from its Q3 call on 21 October.
- Two strategies are now visible. Walmart is opening an MCP-based Agent Gateway and per-agent profiles with spending limits. Amazon blocks Muse, requires every agent to identify itself in its Conditions of Use, and pushes its own assistant instead.
- The real fight is over the customer relationship. Amazon earned more than $68 billion from ads last year, per GeekWire. Meta wants “a percent of the value” Muse creates. Whoever owns the agent sits between the shopper and the shop.
- Agents mostly research, they rarely check out. In HUMAN Security’s September data, 81% of agent requests went to product and search pages and only 2.4% to checkout and payment.
- Affiliate sites face a quiet problem. Amazon’s Associates terms pay commission when a customer clicks a tracked link. An agent that reads a review and buys elsewhere, or inside Amazon’s own assistant, leaves no click to credit.
- For users: set limits before convenience. Approval on every purchase, a spending cap, a one-time card, and knowing whose return policy applies matter more than the demo.
Products Worth Considering
Sonos Era 100 Wireless Smart Speaker, Black
Apple iPad Air 13-inch (M4), 128GB, Wi-Fi 7, Space Gray
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
Six weeks that started a shopping war
The timeline is short, and every step is documented:
- 8 September: Meta launches Muse, a personal agent that “can open a browser, fill out forms, and negotiate on their behalf”. It runs on Muse Secure VM, a dedicated computer in Meta’s cloud, and pays through Link by Stripe, which issues a one-time-use card.
- 20 September: Amazon cuts Muse off. Users see a pop-up: “Continued access by an unauthorized AI agent violates Amazon’s Conditions of Use, to which our customers have agreed.” Amazon says Meta never told it Muse would shop there, that the agent does not identify itself, and that it “appears to capture and store customer credentials”. Meta’s launch post says Muse “has no visibility into people’s passwords or payment methods”.
- 23 September: At Connect, Meta announces retail connectors with Best Buy, Gap, Sephora, Walmart, Wayfair and Expedia. Mark Zuckerberg says Meta expects to profit “by taking a small fee from transactions”.
- 28 September: Shopify lets browser agents complete checkout through three structured tools instead of screen-scraping.
- 6 and 9 October: Sierra and Meta announce the Personal Agent Protocol (“Poppy”) with Walmart, Shopify and Stripe, then publish a draft and 35 more design partners, including OpenAI, Visa, Mastercard, PayPal and Target.
- 21 October: Walmart announces Agent Profiles and the Walmart Agent Gateway. The same evening Meta reports its numbers.
Others are closing doors too. The Verge reports that JPMorgan Chase’s Jamie Dimon called outside agents in customer accounts “too risky”, that Muse hit Target’s bot check one day and got through the next, and that users struggle with Delta and Amtrak.
How an AI agent actually buys
There are three routes, and they explain the whole conflict.
1. The browser, like a human. The agent opens pages built for people, clicks, signs in with your credentials and checks out. This works almost anywhere, which is exactly why shops object: they cannot tell the agent from you. Muse does this from a virtual machine in Meta’s cloud. Perplexity’s Comet does it from your own computer, a detail that mattered in court (more below).
2. The shop’s own agent door. Walmart’s Agent Gateway uses MCP, the Model Context Protocol, so approved agents can search inventory and buy without clicking through pages. Shopify offers a hosted MCP server and, since September, WebMCP checkout tools called get_checkout, update_checkout and complete_checkout.
3. Shared protocols. Several are competing to become the standard:
| Protocol | Backers | What it covers | Status (Oct 2026) |
|---|---|---|---|
| Personal Agent Protocol (Poppy) | Sierra, Meta, Walmart, Shopify, Stripe + 35 design partners | Who the agent is, what the customer allowed, one session across website, API and company agent | Draft 0.1 (9 Oct); payments left to extensions |
| Universal Commerce Protocol (UCP) | Google, Shopify; its site also lists Etsy, Wayfair, Target, Walmart, Amazon and Microsoft as co-developers | Shopping journey from discovery to cart, checkout and orders | Published; built on REST, JSON-RPC, MCP, A2A and the Agent Payments Protocol (AP2) for payment mandates |
| Agentic Commerce Protocol (ACP) | OpenAI, Stripe | Checkout between agent and merchant | Open source, Apache 2.0 |
Agentic shopping
Three ways an AI agent buys for you
Every fight between shops and AI agents comes down to which of these three routes the agent uses, and whether the shop agreed to it.
The browser, like a human
The agent opens pages built for people, signs in with your login and clicks through checkout. Works almost anywhere, which is why shops object: they can't tell the agent from you.
Examples: Meta's Muse (from a virtual machine in Meta's cloud), Perplexity's Comet (from your own computer)
The shop's own agent door
Approved agents search and buy through a structured interface instead of clicking through pages. The shop sets the rules.
Examples: Walmart Agent Gateway (MCP, beta), Shopify MCP server and WebMCP checkout tools
Shared protocols
Common rules for who the agent is, what you allowed, and how the checkout and payment are confirmed.
Examples: Personal Agent Protocol (draft), Universal Commerce Protocol, Agentic Commerce Protocol
The protocols competing to be the standard
| Protocol | Backers | Covers | Status |
|---|---|---|---|
| Personal Agent Protocol (Poppy) | Sierra, Meta, Walmart, Shopify, Stripe + 35 design partners | Agent identity, customer permission, one session across site, API and company agent | Draft 0.1 (9 Oct 2026) |
| Universal Commerce Protocol (UCP) | Google, Shopify; site also lists Etsy, Wayfair, Target, Walmart, Amazon, Microsoft | Discovery, cart, checkout, orders; payment mandates via AP2 | Published |
| Agentic Commerce Protocol (ACP) | OpenAI, Stripe | Checkout between agent and merchant | Open source; ChatGPT's own checkout scaled back in March 2026 |
What a careful agent purchase looks like
- You give the task and a limit, for example a $50 cap.
- The agent identifies itself to the shop and gets only the access you approved.
- It prepares the cart and shows you an approval card.
- Payment goes through a token or one-time card, so the agent never sees your card.
- The shop’s normal returns apply if the agent got it wrong.
Sources: Meta (8 Sep 2026), Walmart (21 Oct 2026), TechCrunch on Shopify (28 Sep 2026), Sierra (9 Oct 2026), ucp.dev, agenticcommerce.dev, OpenAI (24 Mar 2026). Purchase steps: Walmart’s agent terms and Meta’s Muse description.
Powered by Thorsten Meyer AI
A reality check on protocol three: OpenAI launched Instant Checkout in ChatGPT on ACP in September 2025. In March 2026 it stepped back: “the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we’re allowing merchants to use their own checkout experiences while we focus our efforts on product discovery.” Building the pipes is easier than getting merchants through them.
Strategy one: the open gate
Walmart’s plan starts from a real problem. Today an agent signs in with your account, so its purchases and yours look identical. Agent Profiles, “launching in the next month”, give each agent its own profile inside your account. You can require approval for every purchase or allow buying within a limit, “for example, $50, $100 or $200”. Walmart compares it to adding a cardholder to your credit card.
The terms are the interesting part. Agents must not act beyond what the shopper intends. They do not receive card details and cannot change payment methods. Approval is on by default, Walmart’s returns process applies “including if an agent doesn’t act as intended”, and agents that break the rules get blocked. Walmart also promises that agents see the same prices humans see: “We price products, not people.” That is Walmart’s own commitment, and an agent-era price audit would be the way to test it.
Why open up? Walmart sells practical, repeat items, the category where shoppers trust AI most, according to University of Virginia professor Luca Cian in Axios. And an “everyday low price” store benefits when an agent’s job is to find the cheapest option.
Strategy two: the walled garden
Amazon’s position is written into its Conditions of Use (last updated 14 August 2026). The “Agent Terms” say no agent may use Amazon unless it identifies itself in every request with the user-agent string “Agent/[agent name]”, does not mimic human typing or solve CAPTCHAs, answers truthfully if asked whether it is a computer, and stays out when Amazon says so. Amazon’s robots.txt fully disallows about 100 named crawlers and agents when I checked on 26 October, including GoogleAgent-Shopping, ChatGPT-User, Claude-User, Perplexity-User and meta-externalagent.
Amazon’s argument is consent, not a hatred of agents. Its spokesperson said apps that buy for customers “should operate openly and respect service provider decisions about whether or not to participate”, and pointed to food delivery apps and travel agencies, which usually work with the agreement of the restaurants and airlines they sell for. Its own agent, Buy for Me, shops on other brands’ sites but identifies itself and lets brands opt out, per GeekWire.
One twist I did not expect: the Universal Commerce Protocol’s own website lists Amazon among its co-developers, next to Google, Shopify, Walmart and Target (checked 26 October). I found no Amazon statement on when outside agents could use it. So the wall is about permission, not about agents as such.
And Amazon has its own agent. Alexa for Shopping, which connected Rufus with Alexa+ in May and retired the Rufus name, can monitor prices and buy automatically when an item hits a target. Andy Jassy told analysts in February that 300 million customers used Rufus in 2025 and that third-party “horizontal” agents “get a lot of the product details wrong, they get a lot of the pricing wrong”. In October 2025 he had said “I do think that we will find ways to partner.”
Retailers vs. AI agents
The open gate and the walled garden
Walmart builds a front door for outside AI agents with limits and rules. Amazon blocks agents it hasn't agreed to and pushes its own assistant.
Walmart: the open gateAnnounced 21 October 2026
- Outside agents
- Welcome through an MCP-based Agent Gateway (beta) and per-agent profiles, "launching in the next month"
- Your control
- Approval on every purchase by default, or a spending limit, for example $50, $100 or $200
- Payment
- Agents don't receive card details and can't change payment methods
- When the agent errs
- Walmart's returns process applies
- Prices
- "We price products, not people" (Walmart's commitment)
- Own assistant
- Sparky, plus apps in Muse, ChatGPT and Gemini
Amazon: the walled gardenConditions of Use updated 14 August 2026
- Outside agents
- Muse blocked since 20 Sep 2026; about 100 named crawlers and agents fully disallowed in robots.txt
- The rule
- Agent Terms: identify as "Agent/[name]" in every request, no human mimicry, no CAPTCHA solving, leave when asked
- Stated reason
- Consent and security: Amazon says Muse doesn't identify itself and appears to store credentials (Meta says Muse cannot see passwords)
- Own assistant
- Alexa for Shopping (formerly Rufus): price alerts and auto-buy at a target price
- Its own agent elsewhere
- Buy for Me identifies itself and lets brands opt out
- Standards
- Listed as a UCP co-developer on ucp.dev; no public timeline for outside agents
“Whoever owns the agent owns the relationship with the shopper, and every retailer knows it.”Luca Cian, University of Virginia, to CNBC
Sources: Walmart (21 Oct 2026), Amazon Conditions of Use, Amazon robots.txt (read 26 Oct 2026), GeekWire (20 Sep 2026), GeekWire (13 May 2026), ucp.dev, Meta Q3 2026 call, CNBC (3 Oct 2026).
Powered by Thorsten Meyer AI
The turn: this is a fight about the customer, not the code
Follow the money and the positions make sense. GeekWire notes Amazon generated more than $68 billion in ad revenue last year, a business that depends on people seeing sponsored products on Amazon’s pages. An agent that compares five shops and buys never sees those ads. GlobalData analyst Neil Saunders told Axios that Amazon fears agentic shopping could undermine its advertising business.
Meta’s side is just as clear. On the Q3 earnings call, Zuckerberg said Meta wants to “build a model where we get a percent of the value that we are helping to create”. CFO Susan Li acknowledged that some businesses are “understandably focused on protecting customer relationships”, and predicted a move “away from agents attempting to navigate every digital experience like a human” towards authenticated APIs and agent-to-agent deals.
Luca Cian summed it up for CNBC: “Whoever owns the agent owns the relationship with the shopper, and every retailer knows it.”
The law has nudged this into contracts. In Amazon’s case against Perplexity, the Ninth Circuit ruled on 4 August that the user, not Perplexity, “accessed” Amazon under the federal anti-hacking law, because Comet ran through the user’s own computer. The court left open claims based on terms of service and noted that agents whose company servers talk to the website directly could be judged differently. That is why Amazon’s pop-up cites its Conditions of Use, not hacking. Whether a cloud-hosted agent like Muse falls on the other side of that line has not been tested.
What the numbers say, and what they don’t
The adoption figures are real but mostly self-reported:
- Meta: more than 10 million people set up Muse in about six weeks; about 25% use it on any given day; shopping is used by over 25% of weekly actives; the average purchase is over $200. Meta gave no total purchase value on the call.
- HUMAN Security (5 October): agent traffic across its customers’ sites rose 209.7% from August to September. Muse took 40.3% of agent traffic in its first month. E-commerce received about 60%. But 81% of agent requests went to product and search pages and only 2.4% to checkout and payment.
- Adobe (16 April): traffic from AI sources to US retail sites grew 393% year on year in Q1 2026 and converted 42% better than other traffic in March. Note: these are people clicking from chat assistants, not agents buying.
- Consumers are cautious: only 16% are comfortable letting an AI find and buy products for them (Coveo, via Axios), and 59% of US adults call AI-driven price gouging a “major concern” (Morning Consult, via CNBC).
What the numbers say
Lots of agent traffic, little agent checkout
Meta's adoption figures are impressive, but they are Meta's own. Independent traffic data shows agents spend most of their time reading product pages, not paying.
Agents mostly research, they rarely pay
Where agent requests went in September 2026, share of all agent requests on HUMAN customers’ sites.
Agent traffic +209.7% vs August. E-commerce received about 60%.
Who sends the agents
Share of agent traffic, September 2026.
People clicking from chat assistants
AI-referred traffic to US retail sites.
Shoppers are cautious
US surveys cited by Axios and CNBC.
Meta’s figures come from its earnings call and are not audited; Meta gave no total purchase value. HUMAN measures only its own customers’ sites. Adobe counts people arriving from AI tools, not agents buying.
Sources: Meta Q3 2026 earnings call (21 Oct 2026), HUMAN Security (5 Oct 2026), Adobe (16 Apr 2026), Axios (21 Oct 2026), CNBC (3 Oct 2026).
Powered by Thorsten Meyer AI
What this means for companies
Retailers: decide on purpose, before your bot defences decide for you. Target told The Verge that agent access depends on “safeguards” and can differ by page. The RealReal learned the hard way: The Verge used Muse to re-add items to a cart every 20 minutes, keeping them on hold for other shoppers for hours. The company then tightened its bot defence. Practical steps: publish agent terms, require self-identification, cap cart holds and retries per account, and offer a structured route (MCP, UCP or a Poppy discovery file) for agents you want.
Brands: your product data is now read by machines first. Adobe found product detail pages scored only 66% on its AI visibility check, so roughly a third of that content is not ready for language models to read. Price, availability, sizes, return rules and compatibility need to be in structured feeds, not in images or scripts.
Publishers and affiliate sites: this one hits close to home, because this site publishes affiliate reviews too. Amazon’s Associates Commission Income Statement pays when “a customer clicks through a Special Link” and orders within a 24-hour session, or within 89 days for items added to the cart in that session. Its policies also forbid creating sessions “by way of a robot or software program”, and the current version (updated April 2026) requires agents that use Associates content to identify themselves. I found no Amazon rule on paying commission when a user’s agent reads a review and then buys. Under the published terms, no click means no commission. HUMAN’s data shows agents mostly research, which is the job review sites do. If the agent reads your comparison and buys through Amazon’s own assistant or a retailer’s gateway, your work informed the purchase and nothing tracks it. Proposals to pay sources that “informed the agent’s decision” exist, from vendors like Rye, but nothing at Amazon. Diversify: direct audiences, newsletters, and retailers and networks that pay on agent traffic.
What this means for users
Convenience vs. price transparency. Agents are good at chores: re-ordering, price drops, comparisons. But an agent only shows what it can reach. When the biggest store is blocked, “cheapest” means “cheapest among the shops that let it in”.
Payment authority. Use approval for every purchase until you trust the agent, then a spending cap. Prefer one-time card numbers (Meta says Link issues them for Muse). Walmart’s terms say agents never see your card. Agents also expose accounts, so treat credentials as covered in my piece on AI agents holding our keys.
Returns and liability. Walmart says its normal returns apply even if the agent misfires. Meta says Muse purchases through Link get free coverage for damaged or lost items, price drops and returns on eligible purchases. On Amazon, the pop-up reminds users they agreed to the Conditions of Use, and the Ninth Circuit treated the user as the one accessing the site. If you send an unwelcome agent, you are the one breaking the store’s terms.
Honest limits
Much of this is announced, not measured. Walmart’s Agent Profiles are not live yet. Poppy is a draft. Meta’s user and purchase figures come from an earnings call and are not audited. HUMAN measures traffic on its own customers’ sites. Adobe’s figures are AI referrals, not agent purchases. Amazon’s claims about credential capture and Meta’s security claims are statements by the parties, not findings. Nobody has published how much affiliate revenue agents have shifted so far. For the bigger picture on personal agents, see Claude Duet and the race for always-on agents.
My bet: in a year, the question won’t be whether agents shop, but which shops sign which agent contracts. Watch who publishes agent terms, and who gets paid when an agent buys.
Sources
- Meta, Introducing Muse (8 Sep 2026) and Q3 2026 earnings call transcript (21 Oct 2026)
- Walmart, Making It Possible for Agents to Shop on Behalf of Customers (21 Oct 2026)
- Amazon, Conditions of Use, Agent Terms (updated 14 Aug 2026); robots.txt (read 26 Oct 2026); Associates Program Policies (updated 14 Apr 2026)
- GeekWire, Amazon blocks Meta’s Muse AI assistant (20 Sep 2026) and Alexa for Shopping (13 May 2026)
- Sierra, Sharing a draft of Personal Agent Protocol (9 Oct 2026)
- TechCrunch, Muse at Connect (24 Sep 2026) and Shopify agent checkout (28 Sep 2026)
- OpenAI, Buy it in ChatGPT (Sep 2025) and Powering product discovery in ChatGPT (24 Mar 2026); Agentic Commerce Protocol; Universal Commerce Protocol
- Axios, Walmart welcomes AI agents while Amazon blocks Muse (21 Oct 2026); CNBC, Meta’s Muse can shop and check out for you (3 Oct 2026); Fortune, Zuckerberg on Muse and Stripe credit limits (22 Oct 2026)
- The Verge, AI agents are ready to shop (22 Oct 2026) and The RealReal update (26 Oct 2026)
- Modern Retail, Jassy on partnering with third-party agents (31 Oct 2025); Customer Experience Dive, Jassy on first-party agents (6 Feb 2026)
- Cooley, Ninth Circuit rules on AI agent access under the CFAA (6 Aug 2026)
- HUMAN Security, State of Agentic Traffic, September 2026 (5 Oct 2026)
- Adobe, AI traffic surge, retail sites not machine-readable (16 Apr 2026)
- Rye, Affiliate commissions in agentic commerce (vendor analysis)
Researched and drafted with AI assistance, including Anthropic’s Claude.
