Meta launched Muse on September 8, 2026: a personal AI agent that books travel, fills out forms, completes purchases and negotiates on its owner's behalf, and keeps working after the app is closed. For a business, the change is that more first contact will come from software acting for a buyer. That software reads a website differently from a person, and the companies it can read clearly are the ones it will put in front of its owner.
What Meta Launched
Muse is a personal agent rather than a chatbot. Once someone shares a goal, it builds a plan and carries out the steps itself, asking for approval before anything sensitive such as a purchase. It runs on a new Meta model, Muse Spark, inside a dedicated virtual machine per person, with a separate system Meta calls Sentinel deciding what the agent may read and what it may change on the web.
| Detail | What Meta announced |
|---|---|
| Launch | September 8, 2026, on iOS, Android and the web, now available in the United States and Canada, with Meta's AI glasses to follow |
| Price | A free tier, plus subscriptions at $20 and $100 per month |
| What it does | Sends email, books travel, fills forms, buys, negotiates bills and sales, and turns saved content into plans |
| Payments | Checkout through Stripe's Link with single-use card details, Shop Pay announced as coming |
| Advertising | Meta says Muse conversations and the data in a person's virtual machine aren't shared with its ad systems |
| Business version | Muse for Small Business, September 29, connecting to Shopify, Stripe, QuickBooks, Slack and Meta ad accounts |
Adoption was fast. The app intelligence firm Apptopia estimated 2.8 million installs in Muse's first 12 days, and it reached the top of Apple's free iPhone charts in the United States and Canada. Meta hasn't published its own figures. Its VP of AI products, Vishal Shah, told reporters that roughly one in three early users had already connected Muse to a business account.
The Shortlist Gets Made Before a Person Looks
A search engine hands a person ten options and lets them choose. An agent does the choosing and hands back two or three, or simply completes the task. That moves the decision earlier: by the time the owner sees anything, the agent has already compared the options it could find and read, and dropped the rest.
This is the same shift AI answers started, carried one step further. An answer engine names a few companies in a paragraph. An agent acts on what it finds. A business that an agent can't read, or whose details it can't confirm, wouldn't be a worse option on its list. It wouldn't be on the list at all.
An Agent Reads Your Site Before Any Person Does
An agent works from what it can retrieve and verify. That puts weight on the same things that already decide whether a company is cited in an AI answer, which are covered in Generative Search Signals:
- Content That Exists Without JavaScriptPages that only assemble themselves in a browser are slow or blank to automated readers, and an agent working against a deadline moves on.
- Facts That Agree EverywhereThe same name, address, hours, policies and product details on the site, in its structured data and on every profile. An agent that finds two versions has a reason to trust neither.
- Plain Answers to Buying QuestionsDelivery times, return terms, availability, compatibility and lead times, stated on the page in words, not buried in a PDF or an image.
- Product Data Generated From the CatalogPrice, stock and variants in structured data that stays correct as they change, because an agent comparing stores will act on whatever it reads.
None of this is new work. It's the same foundation ordinary search already rewards.
Forms and Checkouts Will Meet Software Customers
Muse fills out forms and completes purchases. That means a contact form, a quote request or a checkout can be submitted by an agent on behalf of a real customer, and businesses will need to decide how to treat that.
Some are already pushing back. Resy has banned users whose agents repeatedly polled its reservation system, and has said it doesn't permit unapproved third-party agents on its platform, while offering approved booking through its integrations with ChatGPT and Claude. That points to where this is likely heading: agents admitted through sanctioned routes, and blocked when they behave like bots.
The payment side is already being built for it. Stripe and OpenAI's Agentic Commerce Protocol, Google's Agent Payments Protocol, Visa's Trusted Agent Protocol and Mastercard's Agent Pay all exist to let a verified agent pay on someone's behalf with limited, auditable permission. Muse's use of Stripe Link and its planned Shop Pay support put that into a mainstream consumer product. For a Shopify store, much of the plumbing would arrive through the platform rather than custom work.
The practical question for a business is narrower: would a legitimate agent submitting your contact form get through your spam protection, and would you know it had been an agent? An over-aggressive filter that blocks real people is a problem many sites already have without realizing it.
DTC Brands Will Feel This First
Muse is built to buy, and its first purchases will be consumer ones: a refill, a gift, the groceries from a saved recipe. That puts direct-to-consumer stores at the front of this shift.
An agent shopping for its owner would compare what a person skims: price per unit, the delivery date, shipping cost and return terms. A store that states those plainly on the product page is easy to buy from. One that reveals them at checkout loses the order to a store that didn't make the agent dig.
Checkout would matter in a new way. Muse pays through Stripe's Link, with Shop Pay announced, so a store on those rails can be bought from without the agent filling in a card form. Repeat purchases would likely move the same way: an agent that remembers its owner's preferences can reorder without a visit, which favors brands with clear subscription and reorder options.
It also changes where influence happens. Meta says Muse conversations aren't shared with its ad systems, so an agent-led purchase may never pass through the ad funnel a brand paid to build. What reaches the agent is the product data, the reviews and the policies it can read, and the content its owner saved, like the recipe reel Meta uses as its own example.
Published Terms Become the Starting Point
Meta's own examples include negotiating a lower bill and selling a car for more. An agent that negotiates starts from whatever terms it can find, so published policies, standard lead times and clear service descriptions would carry more weight than they do when a person skims them. Vague terms invite an agent to ask, and an agent will ask every vendor on its list the same question at once.
The Same Agents Are Arriving on the Business Side
Muse for Small Business, launched three weeks after the consumer product, connects the agent to a company's own tools: its Shopify store, its accounting, its ad accounts and its Instagram and Facebook analytics. Meta says it can analyze sales, campaigns and social activity, draft campaign content and emails, and that nothing publishes, sends or spends without the owner's approval.
Put the two together and some transactions would have software on both ends: a buyer's agent asking questions and a seller's agent answering them. The businesses best placed for that are the ones whose information is already organized, consistent and published, because that's what both agents would draw on.
What Doesn't Change
Muse is a consumer product, and it asks for approval before it spends money. A considered purchase, such as industrial equipment, a service contract or anything with a long sales cycle, will still end with people deciding. What changes is the research that comes first. Buyers who already ask an assistant to shortlist suppliers will increasingly ask one to do the comparing as well, and the suppliers it can read clearly are the ones that make it to the conversation.
The agents are also early and imperfect. Meta's own safety report says its model remains susceptible to prompt injection, where text on a page tries to hijack an agent, and Meta offers a bug bounty for finding one. Expect rules about how agents may act on the web to keep changing for some time.
What a Business Can Do Now
- Check What an Agent Can ReadOpen View Source on key pages and confirm the content is in the raw HTML, not only in what the browser builds.
- Reconcile Your FactsMake the site, its structured data, Google Business Profile and directory listings say the same thing.
- Publish the Answers Buyers Ask ForLead times, delivery, returns, service areas and specs, in text on a page.
- Test Your Forms the Way Software Would Use ThemConfirm legitimate submissions get through and that you can see where they came from.
- Measure Where You Stand in AI AnswersAgents and answer engines draw on the same sources. See how to measure AI visibility.
For stores, the related work is covered in how consumer brands get recommended in AI answers and ecommerce conversion optimization. For everyone else, it starts with AI visibility.
Frequently Asked Questions
What Is Meta Muse?
A personal AI agent Meta launched on September 8, 2026. It plans and carries out tasks for its owner, such as booking, buying, filling forms and negotiating. It asks for approval before sensitive actions. It's free, with paid plans at $20 and $100 per month.
How Is a Personal AI Agent Different From a Chatbot?
A chatbot answers questions. An agent acts: it opens websites, fills in forms and completes tasks across several steps, and Muse keeps working after the app is closed.
Should a Business Block AI Agents From Its Website?
Blocking abusive automated traffic makes sense, and Resy has done it. Blocking every agent outright would mean turning away customers who sent one. The better position is a site an agent can read and a policy for how agent submissions are handled.
Does This Matter for B2B Companies?
Yes, in the research stage. Muse is a consumer product, and considered B2B purchases will still be decided by people, but buyers already use assistants to shortlist suppliers. A company an agent can't read clearly is less likely to make that shortlist.