Deep Dive — Meta wants a cut of what Muse buys, not your eyeballs

"We believe that Muse will make you money," Mark Zuckerberg told the room at Meta Connect on September 24, "and we are standing behind this by making Muse free for a huge number of tokens with the expectation that over time we will profit by taking a small fee from transactions." Read it again next to what the rest of the industry is doing and it lands as something close to heresy: the world's largest advertising company, announcing that its hottest new product will not carry a single ad. No banner under the chat. No sponsored row in the agent's recommendations. Just a cut of what the agent buys on your behalf.

Three weeks after launch that posture is being stress-tested from every direction at once — merchants who refuse to be shopped, analysts who don't believe it will last, and the plain fact that Meta has never in its history left money on the table it could reach. The bet underneath the quote is that trust is now a monetizable asset, and that an agent holding your inbox, your logins and your card credentials can only charge for transactions if it never once sells the person doing the transacting.

What the plan actually is

Muse shipped on September 8 as a US-only personal agent: free up to roughly 100 million tokens a week, with subscription tiers at $20 and $100 a month that chief AI officer Alexandr Wang told Axios exist mainly to cover compute for power users. At launch there was no advertising in the product, Wang said, only "commerce opportunities" being explored. The mechanism was visible from day one — the agent runs in a dedicated virtual machine with its own browser, and, as we reported when Meta's Muse gave a stranger a YouTuber's home address, that trust surface has already been tested in public.

At Connect the commerce machine arrived. Wang announced PayPal support plus retail integrations spanning Walmart, Best Buy, The Gap, Sephora, Wayfair and American Eagle, with Expedia for travel and Instacart for groceries; earlier in the month Meta had already wired Stripe's one-tap checkout into purchases and, per the Wall Street Journal, agreed to let Muse complete purchases at Shopify-powered stores through Shop Pay. The stack for a take-rate business is assembled. What has never been published is the rate itself — "a small fee" is the entire disclosure.

The distribution numbers explain why Meta can afford the experiment. Sensor Tower counted 560,000 daily active users eleven days after launch; the app took the top spot on the US App Store free chart ahead of ChatGPT within a week, and Meta shares rose 11% in a single session as analysts at Wells Fargo and KeyBanc raised their price targets on the uptake. Millions of downloads of a free agent that people hand their email and payment credentials to is not a product looking for revenue today. It is a product looking for a habit.

Why "no ads" is a strategy, not a charity

The case for the ad-free agent was laid out this week by MBI Deep Dives, and it starts from the user's side of the screen: "If I am going to hand over my emails, logins, and payment credentials to an agent, I need to believe it is working for me and not for the highest bidder." An agent that recommends with one eye on the sponsor is an agent you stop trusting with the keys — and against vertically built rivals (DoorDash's own ordering agent, Airbnb's new AI search, every brand's in-app assistant), Meta's edge is exactly the general-purpose trust that a single sponsored placement would burn. Investors MBI talks to mostly expect ads to arrive anyway; the counter-argument is that "no ads" is the strongest possible counter-position, not least against Meta's own history.

There is a second, more subtle layer, and Meta's own engineering blog spells it out: "Muse doesn't share your conversations or the data in your Virtual Machine with Meta ad systems. That said, there are some legitimate scenarios where how you use Muse can influence the ads you see." When Muse browses a clothing designer's site on your behalf, that visit can surface as your activity — retargeting follows. So Meta occupies a position no one else can: it can monetize a personal agent through advertising without ever showing an ad inside the agent. The feed on Instagram and Facebook gets smarter because your agent's errands exist; the chat window stays clean. As one analyst put it on X, Meta is "one of the very few companies that can monetize a personal agent through advertising without ever showing an ad inside the agent itself."

The take rate has a veto clause

Here is the structural weakness in the plan, and it is already visible. Advertising revenue does not require the advertiser's permission to be earned from your own surfaces; transaction revenue requires every merchant in the chain to participate. The first big merchant to answer has been Amazon, which blocked Muse from shopping its site — users now see a notice that "continued access by an unauthorized AI agent violates Amazon's Conditions of Use" — after failing to get Meta to exclude the store voluntarily. Amazon's stated complaints are security and transparency, but its interest is legible: more than $68 billion of ad revenue last year depends on people browsing Amazon's pages and seeing sponsored products, and an agent that buys by SKU without the shelf is a cut of the merchant relationship that skips the ad auction.

We traced Amazon's strategy — keep outside agents off the shelves while paying for presence inside ChatGPT — in Amazon blocked AI agents for a year. Now it sells ads inside one. Meta is running the mirror-image experiment: open the agent to commerce, charge the transaction, show nothing. If the Amazon standoff is any guide, every large retailer now holds a quiet veto over Meta's take rate. A fee nobody has to pay is a fee that gets negotiated merchant by merchant — and the merchants with the most leverage are precisely the ones with their own ad businesses to protect.

The market still doesn't believe it

Eric Seufert's read, from the same Connect coverage: Zuckerberg has said Meta wants to monetize the agent directly, and "I just fundamentally don't see a path forward for agentic monetization at scale that doesn't rely on ads." The honest arithmetic problem is that advertising has been, for practical purposes, all of Meta's revenue for two decades — there has never been a second engine to feed — while transaction take rates attach to a slice of purchases that routes through the agent, in categories where merchants may refuse, at fees competing platforms will undercut. Subscriptions cover compute; they do not cover the tens of billions a year Meta spends on AI.

Then there is the demand-side doubt, which the BBC put bluntly last week in a piece about what agents do to the web: agents don't look at ads, and the sites that live on them are already straining. If the agent becomes the place where discovery happens — "I need a gift, find one and order it" — the whole click-funded layer between merchant and customer shrinks toward the checkout, and Meta's fee attaches exactly there. The optimistic version is that Meta collects at the point of sale what publishers and affiliates lose upstream. The pessimistic version is that the transaction layer ends up thinner than the ad layer it replaces, because agents comparison-shop by default and every fee you add is a fee the agent is prompted to route around.

There is a further irony in the same week's reporting: OpenAI is building the ad-funded alternative — an ad business that reached a $1 billion annualized run rate within months of launch — inside the assistant Meta has declared ad-free. Two labs, two assistants, two incompatible theories of how an agent gets paid, both launched inside the same season.

What to watch

• Whether a fee rate is ever published. "Small fee" with no number, no revenue share, no participating-merchant terms is a press release, not a business model. The first disclosure filing or partner document that names the number settles how serious it is.
• Whether Amazon budges — or Muse routes around it. The resolution of the standoff (direct negotiation, a Buy-for-Me-style reciprocity deal, or a world where the agent simply shops elsewhere) tells you who holds pricing power in agentic commerce.
• Whether ads land inside Muse in the first year. Seufert's challenge is falsifiable: one sponsored placement in the agent, and the trust thesis was a launch strategy rather than a strategy.
• The Confidential VM deadline. Meta has promised a version by year-end in which it cryptographically cannot read the VM, with outside auditors already reviewing the design. That turns "we don't feed your data to the ad system" from a policy into a proof — and makes the no-ads promise considerably harder to quietly break.

Yesterday we looked at how nobody can price what an agent buys from the buyer's side — tokens metered, budgets blown, the unit of sale in flux. Meta's answer is a third option that shifts the meter entirely: don't sell the customer intelligence at all, take a percentage of what the intelligence moves. If it works, it is the cleanest business model in the industry. If it doesn't, the reason will be that the most powerful commerce platform on earth simply said no.

Would you trust an agent with your inbox and card if its recommendations never paid it — and would you still trust it if they did? Tell us in the comments.