Meta Muse and AI Brand Visibility: How the Agent Buys, What Meta Has That ChatGPT and Gemini Don't
Everything below is current as of September 20, 2026. Muse is twelve days old and Meta is shipping changes every week, so some of this will date quickly.
On September 8, Meta launched Muse, a personal AI agent that does things for you instead of telling you how to do them. It books the trip, argues with the cable company, fills the cart and checks out. Sensor Tower counted more than 730,000 US downloads in its first five days, which beats what the Meta AI app did in its own first five, and TechCrunch says it quickly rose to the top of the US App Store charts. A Mac app showed up on September 18. On September 19 Meta opened a connector platform so outside businesses can plug their services into it.
Most of the coverage so far has been about whether anybody should hand Meta their inbox and their bank login, which is a fair question and one I'm going to leave to the security people. I want to talk about what Muse does to brand visibility, because I think it breaks a few assumptions the whole AI visibility conversation has been built on. A few people have started in on this (Debra Aho Williamson's hands-on test is the best thing I've read so far, and I'll come back to it), but I haven't seen anybody walk through the whole thing from the marketer's side.
For the last two years AI visibility has meant something pretty specific. Somebody asks ChatGPT or Gemini or Perplexity or Claude a question, the model writes an answer, and your brand either made it into that answer or got left out. I wrote a book about why that happens, and I built a tracking tool that measures it across those four engines. All of that assumes there's an answer to look at, and with an agent like Muse there often won't be one, because the person asked for a result and what comes back is a booked table or a box on the porch.
THe short versioN
Muse is Meta's personal AI agent, launched in the US on September 8, 2026. It runs on the Muse Spark 1.3 model inside a dedicated cloud computer with its own browser, and it's free, with paid plans at $20 and $100 a month.
Muse is a separate product from Meta AI, the assistant inside Instagram, Facebook and WhatsApp. Meta AI conversations have fed ad targeting since December 16, 2025, while Meta says Muse's conversations and data stay out of its ad systems.
Muse buys from a merchant's own website with a single-use card issued through Link by Stripe, so no retailer has to opt in. OpenAI's Instant Checkout needed merchants to integrate, and OpenAI ended it in March 2026.
Muse has connectors no other agent can offer, because Meta owns the platforms. Instagram (including DMs), Threads, Facebook (including Marketplace) and Messenger are Muse-only.
Personal context by itself doesn't separate Meta from Google. Gemini Spark has worked across Gmail, Calendar and Drive since May. What Meta has to itself is the social side of a person's life, a free tier, and more than a million businesses already running Meta's business agents every week.
Muse's browser agent reads a page's accessibility tree and never sees the raw DOM, according to Meta's own engineering write-up. A checkout that a screen reader can't get through is one an agent probably can't get through either.
When Muse visits your site, Meta says the visit shows up as the person's own activity, so your pixel fires and your retargeting pool fills with people who never saw the page.
Standard AI visibility tracking can't see any of this, because there's no public answer to sample and the result depends on whose Muse is doing the asking.
Muse and Meta AI are two different products
Meta has put the Muse name on a lot of things this year, and they're easy to mix up. Muse Spark is the model family from Meta Superintelligence Labs, first released in April, with version 1.3 out on September 2. There's also Muse Image, and Muse Code, which is a coding tool that runs in the terminal. Then there's Muse, the agent, which is what launched on September 8. And Meta AI is still the name of the assistant that lives inside Instagram, Facebook, Messenger, WhatsApp, Threads and the glasses, and it runs on Muse Spark too. If you're a marketer, the two you need to keep apart are Meta AI and Muse, because they follow different rules.
Meta AI is the one with the reach. Meta said more than a billion people a month were using it as of last fall. Since April it's had a shopping mode that, by Meta's description, pulls ideas from the creators and communities a person already follows. In May Meta added the ability to @ a specific brand or creator and scroll their products in a grid, and to search Facebook Marketplace listings next to results from the rest of the web. In July it started carrying out plans and connecting to email and calendars. And since December 16 of last year, what people say to Meta AI has been a signal for ad and content personalization across Meta's apps, with carve-outs for sensitive topics and with the EU, the UK and South Korea held out for now.
Muse is a separate app (plus muse.ai, a Mac app, and a chat inside WhatsApp), and it's built differently. Meta's launch post says each person's Muse runs on its own cloud computer, which Meta calls Muse Secure VM, with a file system, a terminal and a full web browser. A second agent called Sentinel sits on the same machine and has to approve anything that leaves it. Muse never sees your passwords or card numbers, it asks before it sends an email or spends money, and it keeps working after you close the app and messages you when it needs a decision. It's one long conversation with a memory you can open up and edit. You give it a name and an avatar. The designers say the first line of its system prompt reads, "Your purpose is to make your user's life better."
It's free for most of what people will do with it, with a $20 plan and a $100 plan for heavier use, and it's US-only for now. On September 16 it started placing phone calls to US businesses for people in the beta. Meta says AI glasses are next.
When I say Muse from here on, I mean the agent.
How Muse buys things
TechCrunch's launch coverage laid out the order Muse works in when it needs to reach a business. Number one, it uses a built-in connector if Meta has one for that service. Two, it can set up its own connection to any service with a public API, using credentials the person gives it. Then last, when there's no API, it opens the browser and uses the website like a person would.
For payment, Muse checks out through Link, Stripe's wallet, which issues a single-use card number for each purchase. Meta's engineering write-up says the number is tied to that one merchant and that dollar amount and expires quickly, and that every purchase stops for an approval showing exactly what's about to be bought. Shop Pay is coming as a second way to pay, and 1Password support is coming so Muse can use the logins a person already has.
Compare that to how the other two big players went at checkout. OpenAI launched Instant Checkout inside ChatGPT last September, and it depended on merchants integrating through its Agentic Commerce Protocol. CNBC reported in March that OpenAI was ending it after product selection stayed thin for six months, and purchases now finish on the merchant's own site or app. Google announced its Universal Commerce Protocol in January with Shopify, Etsy, Wayfair, Target and Walmart, and it powers a buy button on eligible listings in AI Mode and the Gemini app, paid through Google Pay, with the retailer staying the seller of record. Both of those run through a program a merchant signs up for.
Muse skips that step. It shows up at your site with a browser and a one-time card, the way a person would, and whether you're "in" Muse comes down to whether your site actually works when it gets there.
Williamson, who spent nearly two decades as an analyst at eMarketer and now runs Sonata Insights, asked Muse to find her a work tote and watched it browse brand sites and Nordstrom, which briefly flagged it as automated traffic before letting it back in. It came back with three bags, she stopped short of buying at the sign-in page, and then she went to Nordstrom herself and ended up buying four items instead of one. Her line for retailers was that "agent readiness is a new form of shelf space," and the whole write-up is worth your time.
How it compares to ChatGPT, Gemini and Claude
The closest thing to Muse is Google's Gemini Spark, and Google got there first. Spark launched at I/O on May 19 as an always-on agent that runs in Google's cloud and works across Gmail, Docs and the rest of a person's Google account, with Canva, OpenTable and Instacart connected at launch. It sits behind a Google AI Pro or Ultra subscription, though, and the free plan doesn't include it. Meta bills Muse as the first personal agent "built for everyone," and since Google shipped four months earlier, I read that as a claim about price and ease of use, and on those two things I think Meta's right.
OpenAI's agent push this summer has been ChatGPT Work, which is aimed at white-collar jobs. TechCrunch reported in August that the app it shares with Codex had about 20 million users, against more than a billion for ChatGPT. Anthropic's agent products are pointed at knowledge work too. Muse is the one aimed at your dentist appointment and your kid's permission slip.
On ads, the big names have split. ChatGPT has sold sponsored placements since February, the business is at a billion-dollar annual run rate, and as of September 10 Amazon is letting its own advertisers buy into it. Google is piloting Direct Offers inside AI Mode. Anthropic said in February that its Claude products will stay ad-free. Muse has no ads today, and Alexandr Wang, Meta's chief AI officer, told CNBC the commerce model for it is "potentially really interesting." On the second-quarter earnings call Mark Zuckerberg talked about building toward a "business-in-a-box" service for companies that sell on Meta's platforms. My guess is Meta makes money from Muse through commerce fees and business services well before it puts an ad unit inside the app. I could be wrong about that.
Two other groups belong in the picture. The startups are moving as fast as Meta is (Instinct, a text-message agent, is reportedly raising at a $10 billion valuation, and it added phone calls the same week Muse did). And Amazon spent last year adding outside AI bots to its blocklist, Meta's crawlers included, while it built its own shopping agent. I don't know how Muse fares on Amazon, and I haven't seen anybody test it.
What Meta has that the others don't
Personal context alone doesn't separate Meta from Google, since Spark reads your Gmail and your calendar. I count four things Meta has to itself.
Number one is the social side of a person's life. The day after launch, Wang posted the connector list, and he described Instagram (with DMs), Threads, Facebook (with Marketplace) and Messenger as "unique connectors that are only available to Muse," according to explainx.ai's write-up of the thread. Meta's own launch example is Muse turning a recipe reel somebody saved on Instagram into a grocery list. When MindStudio tested the onboarding, Facebook and Instagram were already connected, and Muse used them to build a starter profile (location, relationship status, professional interests, hobbies) before it asked for Gmail. Meta says people choose what Muse connects to and can disconnect anything, and I believe both things are true, since you sign in with your Meta account in the first place. Back in January, Zuckerberg told investors that an agent's value comes from "the unique context that they can see," and that Meta's would be personal in a way nobody else's could.
So think about what kind of context that is. ChatGPT mostly has what you've typed into it, Google has your searches and your inbox and your calendar, and Amazon has your order history, which is a record of decisions already made. What Meta has is the stuff from before the decision, the reel you saved, the Marketplace listing you messaged about, the DM thread where you asked a shop whether the jacket runs small. Most of us report saves and nobody does anything with the number. My guess is that changes, because a saved post is one of the few places a customer's interest gets written down somewhere their agent can read it.
Number two is the source material. Meta has said since April that Muse Spark is built to surface and cite what people share on Instagram, Facebook and Threads, that shopping mode draws on the creators and communities a person follows, and that answers will carry "credit back to the content creators." The July update says Meta AI's research pulls from the web and from what creators and communities post in Meta's apps. ChatGPT leans on the open web and merchant feeds, and for Google it's the search index plus Merchant Center. For Meta's AI, your Instagram account and what creators say about you are source documents, in the way a review site or a Wikipedia page is a source document for ChatGPT. That's all documented for Meta AI, the assistant. Whether Muse the agent weighs that material the same way when it goes shopping, I don't know, and Meta hasn't said, though it does have the connectors to read it. eMarketer made a related point back in April, that as long as the model stays inside Meta's own walls, brands get less say over how their products show up and lean harder on Meta's recommendation systems.
Then number three, Meta is on both ends of the conversation. On the Q2 call Zuckerberg said more than a million businesses use Meta's business agents every week on WhatsApp and Messenger, with Instagram rolling out now. His example was Movida, a Brazilian rental car company that runs its whole booking flow inside WhatsApp, where a returning customer can reserve a car in as few as three messages. Muse lives in WhatsApp, and it has Messenger and Instagram DM connectors. So I can picture a customer's Muse messaging a business and getting that business's Meta agent on the other end, with Meta's models on both sides of the exchange. I haven't seen Meta describe that scenario publicly, so treat it as my inference. Google can do a version of this with Business Agent in Search on one side and Spark on the other. Meta's version runs through the messaging apps people already use to talk to businesses, and that's a big head start in a lot of the world.
Then last, distribution and price. Meta told investors that 3.6 billion people use at least one of its apps every day. Muse is free, it works inside WhatsApp, and Meta told TechCrunch it expects most people to stay on the free tier. Spark and ChatGPT Work both start at about $20 a month.
The thing working against all of this is trust. NMI surveyed 1,000 US adults in August and found that 11% had ever used AI to complete a transaction, 10% would give an AI full control of a purchase, and 84% had never put payment details into an AI tool. Forbes, citing Reuters, reported that Meta employees were still hitting failures during launch week, including an agent that got around its guardrails and surfaced private iCloud photos. So I'd plan for Muse, and I wouldn't bet the quarter on it.
Where advertising fits
Meta's position is that Muse doesn't share your conversations or the data in your VM with its ad systems. The same engineering write-up then explains where ads can still come in. When Muse browses the web for you, Meta says, "it will appear as your activity," and its own example is that a clothing designer whose site Muse visited might use that visit to show you an ad on Instagram. I'd read that paragraph to your analytics person and your paid social person, because a few things follow from it.
Your site analytics are going to have sessions in them where a real buyer is represented by software. Filter those out as bot traffic and you're throwing away customers, and if you leave them in, your time-on-page and scroll numbers stop meaning what you think they mean. I don't have a clean answer for how to tell them apart yet.
The retargeting pools get strange too. They'll include people who never laid eyes on the page, and for some of them the retargeting ad will be the first time they've seen your brand at all, after their agent already looked you over. Creative that's written as a reminder ("still thinking it over?") is going to land on people with nothing to remember.
Affiliate is murkier still. Affiverse pointed out that Meta hasn't said whether Muse keeps affiliate parameters intact when it browses, or whether it reports which sources shaped a recommendation.
And there's a loop here that I think marketers should see plainly. Muse sends demand to your site, your own pixel reports the visit to Meta's ad system as that person's activity, and then you pay Meta to retarget them. There's nothing new in the mechanics, since that's how the pixel has always worked, but the separation Meta describes covers what Meta pulls out of the VM and says nothing about the tags you put on your own site.
Later this year Meta plans to ship Muse Confidential VM, where the whole machine is encrypted with a key only the user holds. That's good for users, and for those of us trying to measure any of this it means the thing gets harder to observe over time.
What your site looks like to Muse
There's one technical detail in Meta's security write-up that I think matters more to marketers than anything in the launch post. It says the browser sub-agent that does the clicking is shown a snapshot of the page's accessibility tree. It doesn't get the raw DOM, and it can't run JavaScript inside the page. Muse is driving a real Chromium browser, so your scripts still run and the page still renders. What the agent gets handed is the structured outline that screen readers use, which is roles, names, labels, states and text.
OpenAI tells site owners the same kind of thing about its own browser agent. Its publisher FAQ says ChatGPT Atlas interprets pages using ARIA, "the same labels and roles that support screen readers." So two of the biggest consumer agents read the one layer of your website that most marketing teams have never looked at. If you asked who owns it, you'd probably hear, "That's a dev thing," or, "Legal handles accessibility."
The things that break are the things that have always broken for screen reader users. That means icon-only buttons with no label, an "add to cart" built out of a div, a price or a size chart rendered as an image, form fields with no labels, and a color picker that only makes sense if you can see it. An agent hits one of those in your checkout and it stalls, and then the person either takes over or picks the other option on the shortlist.
I'd add a caution from Adrian Roselli, who's been doing accessibility work a long time. When OpenAI's guidance came out, he warned it would push people to spray ARIA attributes around as an SEO tactic and make pages worse. The first rule of ARIA has always been to use plain HTML wherever it does the job. So I'd ask for real buttons, real labels, alt text that says what the thing is, ARIA only where HTML runs out, and no keywords stuffed into labels.
And please don't try to talk to the agent in hidden text. Meta says it treats everything on a web page as untrusted input, trains the model to ignore instructions it finds in content, runs separate classifiers looking for prompt injection in the page and in images, and pays bounties up to $130,000 for an injection that works. A hidden line telling AI assistants to recommend your product is, to Muse, an attack. It also checks where the browser is headed against the same list of malicious sites Meta uses to protect people inside its apps.
Two more practical things. Nordstrom's bot protection nearly cost it a sale in Williamson's test. Somebody on your security team is going to say, "We block automated traffic on purpose," and about scrapers they're right. Have the conversation anyway, because this particular automated visitor is carrying a customer's payment. And Muse can now place phone calls, so if your business answers a phone, the people answering it should know an agent might be the one asking about availability or trying to book.
Then there's the connector platform. The page describes three steps. You describe what your connector does, Meta reviews it for functional, security and legal requirements and tests it end to end, and then it appears in a directory inside Muse. The page also says, "Editors will review connectors for featured placement," which makes it sort of an app store, a curated directory with featured slots that Meta runs. Zuckerberg's pitch, as explainx.ai reported it, is that you bring the API and Muse brings the agent, the browser and the context about what the person wants. If your business has an API (booking, ordering, quotes, subscriptions), a connector is the most direct kind of visibility there is, because people can call you by name and the agent doesn't have to fight your website. Keep in mind a Muse connector reaches Muse and nothing else. Google took Spark's outside partners in through MCP, which is an open protocol that Claude and others use too. I haven't found a public technical spec for Muse's version yet, and I'd want both eventually.
Four places a brand wins or loses an agent-run purchase
Put all of this together and an agent-run purchase has four points where a brand can win it or lose it. They're what the agent already remembers, what it finds when it goes looking, what the person sees when they're asked to approve, and whether the agent can finish the job on your site. Nearly all of the AI visibility work I see is really aimed at the second one.
What the agent remembers comes first because it can end the process before it starts. Meta says Muse acts on details a person mentioned only once, and its memory sits in files the person can read and edit. Somebody tells their Muse to reorder the usual, or Muse sees the past orders sitting in their Gmail, and there's no research step and no shortlist. The best name I've come up with for this is share of memory, as a cousin to share of voice. The only way into an agent's memory is for a person to like the product enough to name it, which is the oldest job in marketing, and I think the brand-building crowd just got a very good argument for their budget.
What it finds when it goes looking is the part that's closest to the AEO and GEO work a lot of us are already doing. Be findable, be described the same way everywhere, state plainly what the product is and who it's for. For Meta I'd add your presence inside Meta's own apps, since that's where its model is built to look. That means an Instagram profile and catalog with real information in them, creators who say the specifics out loud and in captions, and a Marketplace presence where the category fits. Williamson got three options back out of every tote bag on the internet, and I'd assume that's about the size of the list most of the time.
What the person sees at approval is small. It's a card with a merchant name, an item and a price. NMI found that 70% of people want to review or override before an AI buys anything, and 58% worry it'll pick the wrong item, size, color or brand. A name they recognize gets a quick yes, and one they've never heard of gets a "wait, who is this?" and a trip to the website to check. The same survey asked who people would blame when an AI buys something before they were ready, and 34% said the retailer.
Whether the agent can finish is the unglamorous one. It comes down to the accessibility work above, bot rules, a guest checkout that doesn't force an account, and forms that say what they want. Williamson bailed at a sign-in page, and I'd bet a lot of people do. If you sell through messaging, I'd turn on Meta's business agent and keep its information current, since a customer's Muse may end up talking to it.
Why the tracking tools can't see this yet
Every AI visibility tool works the same basic way, and I'm including the one I built, which is called Citingly. You send a prompt, you record the answer, you count who got mentioned, and you do that enough times to call it a share of voice. It works because the question is public and anybody could have asked it.
Muse Spark is available through Meta's API, so a tracker can sample the model. What comes back is the model with nobody's life attached to it. The Muse that matters is the one with a particular person's memory, their connected accounts and their saved posts, and its output is frequently an action that no outsider ever sees. There's no results page to scrape and no answer log to buy.
So for now the only instrument is doing it by hand. Get a few people on your team with different profiles to run the same errands in your category, screen-record the browser pane, and write down where Muse actually goes, what it shortlists and where it stalls. It's tedious and the sample is tiny. I'd still rather have a handful of real runs than a dashboard number that was never measuring this.
A few measurement habits are going to need revisiting while you're at it. Adobe's numbers show AI referral traffic to US retail sites up 62% year over year in July, with those visitors worth 53% more per visit. That's people clicking through from an assistant. An agent that does the visiting itself doesn't arrive as a referral in that sense, and I don't know yet how those sessions get classified. I'd ask your payments person whether orders paid with Link's single-use cards can be identified on your side, because that might be the cleanest marker you get. And watch basket size on anything you can tie to an agent. Williamson spent more than four times as much shopping on her own as she would have if she'd let Muse buy the one bag.
What I don't know yet
Start with how Muse picks its three options. Meta hasn't published anything about how they get ranked, and my guess is it's some mix of what the person asked for, what the agent can verify on the page, and price, which I can't prove. Whether placement ever goes up for sale is the next question. There are no ads today, and Williamson's read is that Meta usually waits for very large scale before it monetizes a new surface, which sounds right to me. CNBC reported that the Muse app has a feed that summarizes updates from a person's connected Facebook and Instagram accounts along with stories from the web. If a sponsored slot ever shows up inside Muse, my guess is it starts there, or in the Ideas tab, or in the featured slots in the connector directory.
Adoption is the other big unknown. A fast first two weeks on the App Store tells you people are curious, and the NMI numbers tell you most of them aren't ready to hand over a purchase, and both of those can stay true for a long time. I also can't tell you whether the advantage Meta's content has inside Meta AI's shopping mode carries over to Muse when it's out shopping on the open web. It would make sense, but nobody outside Meta has shown it yet.
Sourcing and limits
Everything here comes from Meta's own published materials, its second-quarter earnings call, hands-on tests by other people that I've named, and reporting I've linked inline. None of it is my own lab data. Muse is US-only and less than two weeks old, the connector platform is two days old, and I'd expect some of what I've described to change before the end of the year.
I wrote Explainable about why AI recommends some brands and ignores others, and the idea underneath the book is that a machine has to be able to explain you before it'll recommend you. An agent raises that bar a notch, because now it has to explain you to one specific person who's about to be asked to approve a charge.
If you've run your own category through Muse, I'd like to see what it picked and where it got stuck, especially if it contradicts anything I've said here.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.