How to Show Up in Meta Muse: The 2026 Playbook for Brands and Products

TL;DR

• Meta launched Muse on 8 September 2026 as a personal agent that browses, fills forms, books and buys. Estimates reported by Bloomberg put it past 900,000 downloads in its first six days, and it reached number one on the US App Store inside its first two weeks.
• Muse reaches your business by three separate routes: a product catalogue, its own browser, and the connectors a user has approved. Each route has a different optimisation job.
• The browser route is the one most brands misread. Meta's engineering write-up says the browser sub-agent sees an accessibility tree snapshot of the page rather than the raw DOM, and has no ability to run JavaScript in the page context. Your page still renders in Chrome, so this is not a no-JavaScript test. It is a semantics test.
• Meta says its browsing "will appear as your activity", and the Meta documentation I checked publishes no Muse user agent, so there is no clean row waiting in your acquisition report. Amazon blocked the agent on 20 September partly because it does not identify itself.
• Do the unglamorous work: accessible page semantics, facts in real text, a clean catalogue feed, and a written decision on your own bot policy before someone else makes it for you.

On 8 September 2026 Meta shipped Muse, a personal AI agent that does not stop at answering. It opens a browser, fills in forms, sends email, books travel, lowers bills and pays for things. It keeps working after the user closes the app.

Twelve days later, Amazon locked it out.

That is the shape of the story so far, and it tells you more about the next two years of marketing than any keynote did. The buyer is starting to delegate. The thing doing the reading, the comparing and the checkout is not a person with a browser tab. It is a program with a task list, a payment card and a very literal way of reading your website.

This is a practical guide to being the brand that program picks.

How big is Muse, really?

Big enough to plan around, early enough that the work is cheap.

Those are installs, not habits, and a launch curve is not a market. The number I care about is the one underneath it. Adobe, tracking over a trillion visits to US retail sites, found AI-sourced traffic to US retailers up 393 percent year on year in the first quarter of 2026, and by March that traffic was converting 42 percent better than non-AI traffic. A year earlier it converted 38 percent worse. Adobe's numbers describe a channel that stopped being a curiosity somewhere around the start of this year.

Muse is what happens when that channel grows hands.

How does Muse actually find your product or your page?

Three routes, and they behave nothing alike. Most of the advice circulating about "optimising for Muse" collapses them into one, which is why it is useless.

Route one: the catalogue

If you sell on Shopify, this one already happened without you. Shopify's changelog entry of 8 September 2026, the day Muse launched, says Meta is now an AI channel in the admin and that "your products are shared with Meta by default through Shopify Catalog". Shopify's documentation adds that every available agentic channel has access to your products through Shopify Catalog by default, and that eligible stores can offer direct checkout on Meta surfaces.

One caveat worth holding onto: Shopify documents a Meta channel and Meta surfaces, not a Muse-specific route. The public documents I checked do not establish how much of that discovery or checkout traffic belongs to Muse rather than to another Meta surface.

What flows through is the usual catalogue payload: titles, descriptions, options, images, prices and availability. What that means in practice is that your product title and your variant naming are now retrieval copy. "Merino crew neck, mid weight, machine washable, navy" answers a shopper's stated constraints. "The Hartley" does not.

The settings matter as much as the mechanism. Controls sit under Sales channels then Agentic in the Shopify admin, and turning catalogue sharing off can take up to seven days to take effect. A detailed review of the Meta channel documentation reports that direct checkout is active by default for eligible stores, is deactivated by default on Shopify's Agentic plan, and excludes subscriptions, bundles, customisable items and B2B-only products, along with local delivery, in-store pickup and pickup points. Check your own admin rather than assuming. You remain the merchant of record for the order, the fulfilment, the returns and the complaint.

Route two: the browser

This is the route that covers everyone else, and it is where the misreading happens.

Meta's research team published how Muse is built. Each user gets a Muse Secure VM. A separate process called Sentinel is the sole permission authority for connector actions and network egress. The browser sits behind a broker that owns the Chrome DevTools Protocol connection, and Meta describes the agent's access like this:

Meta says the browser sub-agent "sees an accessibility tree snapshot of the page", not the raw DOM, and that it "has no ability to run javascript in the page context, no script verbs, no exec in the browser process, and Chrome devtools are disabled".

Read that carefully, because the popular summary of it is wrong.

This is Chrome. Your page renders, your scripts run, your framework hydrates. What is restricted is the agent, not the browser: it cannot inject or execute JavaScript of its own, and it does not get the raw DOM. What it gets is the accessibility tree, the same computed structure a screen reader consumes, made from roles, names, states, headings, links, labels, tables and lists.

So the test is not "does my page work with JavaScript off". The test is: after everything renders, is each fact present as exposed text, and is each control named?

Be precise about this, because it is easy to over-correct. Ordinary visible text is exposed in the accessibility tree even inside a generic div, so a price written as text is fine wherever it sits. What is not fine is a fact that never becomes text: a delivery promise drawn into a canvas element, a spec table baked into a promotional image with no alt text, a stock status conveyed only by the colour of a dot. And for anything interactive, the roles, names, states and keyboard behaviour are what an agent has to match against. A size selector built from unlabelled divs with click handlers gives it nothing to aim at.

A checklist that follows from the actual mechanism:

1. Look at the tree, not the screenshot. Open Chrome DevTools, find the accessibility pane, and read what your product page computes to. That view is closer to what Muse gets than anything else on your desk.
2. Use real elements. A button that is a button, a link that is a link, a label tied to its input, a table with header cells. Properly labelled custom controls work too, but only when someone did the labelling.
3. Name icon-only controls. Alt text, aria-label, visible label text, and keyboard operability. An icon button with no accessible name is a dead end for an agent and for every screen reader user you already have.
4. Get the facts out of pixels. Anything living only in an image, a canvas or a colour cue should also exist as text somewhere on the page.
5. Mind the timing. The agent reads a snapshot. A fact that arrives from a third-party widget several seconds after the rest of the page may not be in the snapshot when it is taken. Server-rendering the important ones is a timing decision as much as an accessibility one.
6. Put the constraint answers in words. Vegan, gluten free, ships to Ohio, fits a 2019 Golf, in stock in a 42. Agents match on stated constraints, and a constraint that exists only in a lifestyle photograph is doing less work than you think.

Adobe's machine-readability work says how far off the average is. Adobe's retail analysis scores individual product pages at 66 percent on machine readability, the lowest of any major page type. In a later Adobe report, whole-site scores by retail category run 63 percent for cosmetics, 56 for electronics, 51 for sporting goods and apparel, 48 for grocery and 47 for furniture and home. Adobe's tool measures readability for language models rather than Muse's particular interface, so read it as a direction, not a Muse score. The direction is unflattering: roughly a third of what a shopper sees is not reaching the software reading alongside them.

Route three: the connectors

Muse works inside the user's own accounts. Users choose which apps it connects to and how much access it gets, Meta says, and the email connector filters out one-time tokens, password reset links and login magic links before the agent sees them.

The practical consequence is that for users who connect their inbox, your order confirmations, booking emails and account notices are read by software on their behalf. A confirmation email whose reference number lives only inside a hero image is harder to extract reliably at the exact moment a customer's agent is trying to reschedule a delivery, and "harder" compounds across thousands of messages.

Plain structure, real text, one clear reference number, dates in an unambiguous format. Nothing exotic.

Why isn't there a Muse line in your analytics?

Because it is not presenting itself as a crawler. It is acting inside the user's session.

Meta's own security post says that when Muse browses, "it will appear as your activity", using the example that a designer whose site Muse visits might later show that person an Instagram ad. The Meta documentation I checked publishes no Muse user agent or referrer for site owners to key on, so there is no tidy row waiting in your acquisition report. That is an absence of a documented signal rather than proof that no signal exists anywhere, and the position may change, but plan for the absence.

Amazon's objection on 20 September rested on the same fact. Amazon said Meta did not tell it that Muse would access its store, that the agent does not identify itself when it browses, and that it appeared to capture and store customer credentials. Shoppers using Muse on Amazon met a popup: continued access by an unauthorised AI agent violates Amazon's Conditions of Use. Meta's answer is that Muse "has no visibility into people's passwords or payment methods" and that shared credentials go into secure storage, which you can read in Meta's own launch post.

Two large companies that partner on shopping and on compute are now arguing in public about whether an agent may walk through a logged-in session. Expect more of this, not less.

Since no Muse-specific identifier is documented in the sources I checked, measure the things you can see move.

• Accessibility-tree reachability. Take your top thirty pages, capture the computed accessibility tree, and list the facts that never become text and the controls that have no name. This tests one documented input to one route, so it is not a complete exposure model, but it is the part you can fix this week.
• Channel-attributed orders. In Shopify, agentic orders land attributed to the Meta channel rather than to a Muse-specific field. That aggregate can include other Meta surfaces, and Muse browser purchases can land elsewhere, so watch the line and do not report it as Muse revenue.
• Server-to-server events. Direct checkout on a Meta surface does not fire your on-page analytics or custom pixels, so events arrive server side. If your dashboards only listen to the browser, agentic revenue will look like it never happened.
• Session shape, held loosely. Very fast, very linear sessions with no hover and no scroll dithering are worth logging, but nobody has published a validated Muse signature and the same pattern describes test automation and impatient people. Treat it as a hypothesis to investigate, never as a count to report.
• Answer-level checks. Ask the agent your buyers' questions and record what it names, what it quotes and what it gets wrong. This is the same discipline as tracking AI referral traffic in GA4, applied one layer further out.

Should you block agents like Muse?

That is now a real commercial decision, and refusing to make it is also a decision.

Amazon gave its reasons publicly: no notice, no consent, an agent that does not identify itself, and concerns about credentials, privacy and security. My own reading is that a marketplace also has more to lose than most businesses from another party owning the buying moment, and Amazon has spent the past year keeping outside agents off its site. Almost nobody reading this has Amazon's position. If you are a brand, a retailer, a hotel, a clinic or a software company, an agent that arrives with a task and a payment method usually represents high purchase intent. It can still fail, pause for an approval, switch supplier or abandon the task, so it is a strong lead rather than a guaranteed sale.

There is a real distinction between a training crawler, a search crawler and a user-directed agent, and the policy should differ for each. I wrote about the mechanics of that split in the piece on AI agent access management, and about how crawler policy quietly becomes distribution policy in the TIME experiment.

A workable default for most businesses:

What would I do in the first thirty days?

A short programme, in this order, because each step makes the next one cheaper.

1. Run the accessibility-tree diff on your twenty highest-value pages. Compare what a person sees against what the computed tree exposes. Every fact that appears in one and not the other is a fact an agent has to guess at.
2. Fix the accessible names. Icon buttons, size pickers, variant swatches, add-to-basket controls, booking widgets. It pays twice, once for agents and once for the people who have always needed it.
3. Rewrite product titles and variant names as constraint answers. Material, fit, size, compatibility, dietary flag, delivery window. Feed copy is now retrieval copy.
4. Audit the catalogue channel settings. Know which agentic surfaces your products flow to, know what happens to subscriptions and bundles, and know the seven-day lag before any change lands.
5. Make transactional email machine-legible. Real text, one reference number, unambiguous dates, no key fact locked inside an image.
6. Write the bot policy down. One page, four visitor classes, a named owner. Review it quarterly, because the tokens and the rules move faster than your redesign cycle.
7. Baseline the answers. Ask Muse, and the assistants your buyers already use, the twenty questions that precede a purchase in your category. Record who gets named. That record is the only honest starting line you will get.

Then keep score monthly. Structured data still matters, and schema that matches the visible copy is the cheapest correctness signal you own, but schema describing a price that never makes it into the accessible page is a label on an empty box.

What this actually changes

For years the job was to persuade a person who had already landed on your page. The page did the work: the photography, the hierarchy, the social proof, the nudge.

An agent reads differently. It arrives with a constraint, takes a structured view of the page, matches what it can name, and moves on. It does not feel anything about your hero image. It cannot be charmed. It can only be answered.

That is not the end of brand. Brand is what makes a person name you in the prompt in the first place, and the prompt is where the whole chain starts. But between the prompt and the purchase there is now a reader that works best with plain, named, verifiable facts.

Give it those, and you are in the running. Leave them sitting in pixels and unlabelled markup, and you are relying on it to guess.

Sources

1. Meta, Introducing Muse: the world's first personal AI agent built for everyone, 8 September 2026.
2. Meta AI Research, How we built safety into Muse.
3. GeekWire, Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping, 21 September 2026.
4. TechCrunch, Meta's Muse is outpacing ChatGPT's early mobile launch, 21 September 2026.
5. CNBC, Meta's Muse AI agent downloads are surging, 21 September 2026.
6. Bloomberg, Meta's new Muse AI app tops charts, 21 September 2026.
7. Shopify Help Center, Shopify agentic storefronts.
8. Shopify, Meta is now an AI channel in your admin, 8 September 2026.
9. Adobe, AI traffic grows but retail sites lag in AI search visibility.
10. Adobe, AI traffic to travel sites surges, for the category-level machine-readability scores.
11. Digital Commerce 360, AI's key conversion metric is improving, 23 April 2026.

Checked against the sources above on 22 September 2026. Muse is moving weekly, so treat every figure here as dated rather than settled.