Meta’s new AI agent Muse needed just 22 days to cross an estimated 5 million downloads.
That sounds like a product story. Look closer, though, and it starts to look like a distribution story—and potentially a preview of how the next phase of the AI race gets won.
According to Sensor Tower data reported by 9to5Mac, Muse reached 5 million U.S. downloads considerably faster than ChatGPT, Grok, or Claude did after their respective mobile launches. It also spent 12 straight days atop the U.S. App Store. Sensor Tower estimates that only 23 apps have crossed 5 million U.S. downloads within 22 days, and most were games.
But Meta had something those AI launches didn’t: Facebook, Instagram, WhatsApp, a giant advertising machine, and the ability to repeatedly put Muse in front of people who never went looking for an AI agent in the first place.
Sensor Tower estimates Muse received as much as 50% of Meta’s daily house-ad impressions during a two-week stretch in September. Meta also promoted the app outside its own ecosystem, while Sensor Tower ranked Muse among the top 15 U.S. brands by advertising impressions during the week of September 21.
That makes Muse an unusually clean test of a question the AI industry is going to keep running into:
What happens when the company with the best distribution doesn’t necessarily need the best chatbot to win?
- Meta can put an AI agent almost anywhere
- Five million downloads still don’t tell us whether Muse won
- The prize is sitting between you and the checkout button
- Amazon already said no
- Muse has already shown how messy “permission” can get
- Then there’s the Meta-sized privacy question
- Distribution may be the new model advantage
Meta can put an AI agent almost anywhere
Muse launched September 8 as what Meta calls a “personal AI agent”: software designed to act on your behalf instead of simply replying to prompts.
Meta says Muse can open websites, fill out forms, send emails, book travel, negotiate, remember personal context, and continue working after you close the app. Before sensitive actions such as sending an email or making a purchase, the agent is designed to request approval.
That difference matters.
Chatbots compete for questions. Agents compete for tasks. And some of the most valuable tasks end with a transaction.
Muse already supports payment workflows, while Meta has announced additional integrations and connectors across commerce and business software. Meta is also trying to move the agent beyond a standalone app and into the devices and services people already use.
Meta’s real advantage here may be painfully simple: it already owns a lot of those surfaces.
Google can weave Gemini into Search, Gmail, Android, and YouTube—a distribution strategy we saw play out at Google I/O. Meta can make a similar play through Facebook, Instagram, WhatsApp, Threads, its advertising network, and eventually its glasses.
The consumer AI race is starting to resemble less of a model benchmark and more of a fight over who owns the front door.
Five million downloads still don’t tell us whether Muse won
Meta’s launch numbers deserve some skepticism for another reason: nobody should confuse an install with a habit.
Different mobile intelligence companies were already producing materially different Muse estimates before the 5 million milestone. Sensor Tower, Apptopia, and Appfigures all produced different cumulative estimates during Muse’s early launch. These services use different methodologies, so none should be treated like an audited user count.
There are also meaningful differences between Muse’s launch and those of the AI products it is being compared with. ChatGPT arrived before consumer generative AI was a familiar category. Claude and Grok were initially limited to iOS during parts of their early mobile launches, while Muse arrived on iOS and Android with Meta’s promotional machine already behind it. Sensor Tower itself noted those launch-condition differences.
The better numbers will come later: how many people are still using Muse after 30 or 90 days, what tasks they repeatedly delegate, how often those tasks succeed, and whether people trust the agent enough to connect accounts and authorize purchases.
Early usage can tell us more than downloads alone, but even that evidence remains preliminary. The real test is whether Meta can turn a burst of attention into repeated delegation.
So the interesting question isn’t whether Meta “bought” five million downloads.
It’s what Meta can do with the attention once it has it.
The prize is sitting between you and the checkout button
An AI agent becomes economically much more interesting when it stops recommending what to buy and starts making the purchase happen.
Imagine asking Muse to plan a trip. It could search destinations, compare flights and hotels, interact with booking sites, remember your preferences, and eventually bring you a transaction to approve.
The user sees convenience.
Everyone else sees a new intermediary.
Meta has already signaled that commerce is part of the plan. Muse supports payment infrastructure, and the company is building integrations that allow the agent to interact with more services.
That introduces a much harder question than whether Muse gives good answers: who decides which businesses the agent puts in front of you?
If millions of consumers eventually stop visiting individual stores, travel sites, marketplaces, and apps directly and instead say, “Muse, find me the best one,” the agent starts becoming part search engine, part personal shopper, part browser, and part checkout lane.
For merchants, that creates an uncomfortable trade.
An AI agent with millions of users could become a powerful acquisition channel. But participating could also mean surrendering part of the customer relationship: less direct traffic, less control over merchandising, fewer opportunities to upsell, and potentially another commercial intermediary between the business and the buyer.
We’re already watching that conflict emerge.
Amazon already said no
Amazon blocked Muse from shopping on its platform in September, objecting to its unauthorized access and raising concerns around privacy, security, and customer credentials. The Verge reported the block.
Other companies are choosing a different path: controlled integration.
Rather than allowing an agent to roam freely across their services, businesses can build or approve specific connections that determine what the agent can access and how it interacts with their systems. Meta has been expanding Muse through exactly that sort of connector model.
That distinction matters.
We covered Amazon’s Muse block earlier at The Neuron because the dispute exposes a basic permission problem: if you authorize an AI to act for you, does the AI automatically inherit your ability to use every website and service you can access?
Amazon’s answer, at least for Muse, was no.
If more platforms insist on controlled access, the supposedly universal AI agent starts looking less universal. It becomes a patchwork of approved integrations, negotiated partnerships, blocked websites, application programming interfaces, commissions, and business deals.
And suddenly the best agent isn’t merely the one with the smartest model.
It’s the one allowed through the most doors.
Muse has already shown how messy “permission” can get
The trust problem isn’t theoretical anymore.
Tech YouTuber Matt Robb recently used Muse to help sell a keyboard through Facebook Marketplace. Muse accepted a $600 offer, shared Robb’s home address with the buyer, and arranged a pickup. Robb said he did not realize the agent had been authorized to take those actions until the buyer arrived at his home. Business Insider reported the incident.
There’s an important complication: Robb had previously selected an “Allow Always” permission.
Meta said Muse acted within the permissions Robb had configured. Robb, meanwhile, said he misunderstood what that permission allowed the agent to do. Reporting on the incident says Meta acknowledged the interface could make those implications clearer.
That makes the episode more revealing than a simple “AI ignored its user” story.
Agent safety depends not only on whether software technically obeys a permission setting, but whether a human actually understands what granting that permission means.
Clicking “Allow Always” for a chatbot sounds abstract. Clicking it for software that can negotiate with strangers, accept offers, send messages, share information, and arrange real-world meetings is something else entirely.
As agents gain more autonomy, the distinction between permission granted and permission understood becomes part of the product.
Then there’s the Meta-sized privacy question
The amount of trust required from users rises dramatically once an AI moves from answering questions to acting for them.
Meta says Muse runs inside a dedicated Secure VM, stores credentials separately so the agent cannot directly see them, uses another system called Sentinel to review actions, requests approval before sensitive activities, and provides an audit trail of what the agent has done.
There’s another promise with particular significance for Meta: the company says Muse conversations and data stored in its virtual machine are not shared with Meta’s advertising systems. Users can also disconnect services and opt out of having their interactions used to train Meta’s AI models.
Those boundaries matter because Muse is being launched by one of the most sophisticated advertising companies ever built.
As agents become more personal, they potentially encounter a different class of information from social media: what you want to buy, where you want to travel, which emails matter, what bills you want negotiated, which businesses you prefer, and what tasks you want completed.
Meta says it has built technical boundaries around that information. But the Marketplace incident demonstrates that privacy cannot be reduced to whether Meta’s ad systems receive agent data.
It also depends on what the agent itself is authorized to reveal, what users believe they authorized, and whether the interface makes the consequences obvious before the agent acts.
That is a much harder product problem.
Distribution may be the new model advantage
The AI industry spent the last few years obsessing over which model scored highest on which benchmark.
Muse points toward another competitive layer.
A company can build a capable model. A much smaller group of companies can immediately distribute an agent to millions of people, integrate it with commerce, put it inside communication tools, subsidize its acquisition costs, connect it with merchants, and eventually carry it onto a pair of glasses.
Sensor Tower’s broader research has already suggested that AI is becoming a new front door to shopping. Muse shows what happens when one of the world’s largest consumer internet companies decides it wants to own that door.
Five million estimated downloads do not prove Meta has succeeded. Retention could disappoint. Users may hesitate to delegate consequential tasks. Merchants could resist. And more incidents—or recurring confusion about what users have actually authorized—could quickly erode trust.
But Meta has already demonstrated something important: when it wants millions of people to notice a new AI product, it doesn’t have to wait for them to discover it.
It can bring the product to them.
The next phase of the agent race may therefore come down to a deceptively old-fashioned advantage: not who builds the smartest assistant, but who gets theirs in front of you first—and how much of the internet lets it through the door.
