From Muse to Instinct: How to Optimize for Visibility on Next Generation of Action Engines

Everything here is current as of October 25, 2026. Both products are shipping weekly, so expect parts of this to date.

Meta’s Muse and Instinct are personal AI agents that research, choose and buy on a person’s behalf. They are not answer engines. An answer engine returns a list you read; an action engine returns a verdict you approve. For brands, that collapses the funnel to a single question: when the agent comes back with one option, is it yours?

Within 40 days of its September 8 launch, Muse hit 10.2 million US downloads and 2.5 million daily active users, with better 7-day and 14-day retention than ChatGPT, Claude and Grok, per Sensor Tower (October 20, 2026). Instinct, invite-only with no app and 14 employees, closed a $1 billion Series C at a $10 billion valuation on September 28 while reporting about $1 billion in annualized transaction volume, half of it travel. Both check out on every eligible Shopify store with Shop Pay. Both place phone calls. Amazon blocked Muse on September 20. A protocol for how agents and businesses talk to each other was announced on October 6. This is moving faster than any search shift I’ve watched in 15 years, and most of the AEO playbook was written for a world where a human reads the answer.

Disclosure: I run Goodie, which monitors brand visibility across AI engines including Muse Spark. What follows is my read of what’s documented. Where I’m inferring I’ll say so, and the parts nobody outside Meta or Instinct can see yet I’ll flag rather than fill in.

What they are, side by side

The Muse and Instinct Growth

Muse is the fastest consumer AI launch since ChatGPT. 500,000 users in its first week with 2 million prompts submitted, per The Information. Sensor Tower projected 5 million US downloads in 22 days. 10.9 million global downloads, 10.2 million of them US, in 40 days, making it only the twelfth app to reach that mark since Sensor Tower started tracking in 2015, and the third non-game app, after Threads and Disney+. It has held the number one spot on the US App Store on every day but one since launch. Retention is the number that matters: 30% of US installers were still using it after seven days and 23% after 14, against averages of 19% and 16% for ChatGPT, Claude and Grok. The caveat is spend: Sensor Tower estimates 39% of Meta’s US ad budget went to Muse in October, up from 23% in September, on top of promotion across apps with 3.6 billion daily users. Some of this is pull, a lot of it is push, and the next 40 days will tell us which.

Instinct’s numbers are smaller and stranger. More than 100,000 users as of mid-September, per The Information, growing roughly 10% a day with zero marketing spend, according to the founder. Annualized transaction volume of about $1 billion from that base. Three weeks in, about 40% of users have shared a personal credit card, and users who connect something sensitive retain at around 80%. These are founder-reported figures from the Invest Like the Best podcast and interviews, not audited, and the company has published no retention curve. What investors paid for is the transaction layer: a 4x valuation step-up in about a month, from $2.5 billion in August to $10 billion in September.

One more number that reframes the category. DataDome found Muse alone accounted for 33% of agentic browser requests across its network in the last nine days of September. The agent traffic on your site has already changed shape.

How Muse works

Meta published more about Muse’s architecture than any lab has about a consumer agent, so this part is mostly documented.

The model. Muse Spark is Meta’s model family from Meta Superintelligence Labs, first released April 8, 2026 and now at version 1.3 (September 2). Meta trained it for long-horizon agentic work: zero-shot tool calling, multi-step workflows, self-correction across messy sources, and multi-agent coordination. Muse can spawn subagents to run parts of a task in parallel. Meta engineers say 1.3 used about 20% fewer tool calls and 25% fewer tokens than 1.2 on the same work.

The computer. Every user gets a Muse Secure VM: a dedicated Linux environment with a full Chromium browser the person can watch, a terminal, local storage, and the ability to run code. A separate process called Sentinel controls everything that leaves the machine and can’t be overridden by the agent. Credentials live in a store the model never reads and get injected at the network boundary. Meta says conversations and VM data stay out of its ad systems. A Confidential VM with user-held keys is promised for late 2026.

The three lanes to a business. This is the part that decides whether you’re reachable. Per TechCrunch’s launch coverage and Meta’s own business post, Muse works down a list: first a built-in connector if Meta has one (Gmail, OpenTable and Plaid among them; any business has been able to submit one since September 18 at muse.ai/platform, subject to Meta review and editorial featuring); second, a direct connection to any service with a public API using credentials the user provides; third, when there’s no API, the browser, used the way a person would. The third lane is where most businesses are today, which is why it’s the one to fix first.

How Muse reads your page. Meta’s engineering write-up says the browser sub-agent is shown a snapshot of the page’s accessibility tree, not the raw DOM, and can’t run JavaScript inside the page. Your scripts still execute and your analytics still fire; what the agent reasons over is the structured outline screen readers use: roles, names, labels, states, text. OpenAI says the same about its browser agent. Two of the largest consumer agents read the one layer of your site most marketing teams have never opened.

What Muse treats as hostile. Everything on a web page is untrusted input. Muse Spark is trained to recognize and resist instructions found in content, separate classifiers scan pages and images for injection, and Meta pays up to $130,000 for an injection that works against a user. A hidden line telling AI assistants to pick your product is, to Muse, an exploit, not a tactic.

How Muse buys. Through Link by Stripe, which issues a single-use card number scoped to the approved purchase. Every purchase stops for the person to approve the total. On Shopify, checkout runs through Shop Pay, live in Muse for all Shopify stores since September 21. Shopify’s WebMCP tools for storefront, cart and checkout followed on September 28 and need no merchant configuration.

Where Muse looks. Meta has not published a ranking method or named a search index for Muse. What’s documented: Muse Spark has web search as a tool; Meta AI’s research pulls from the open web and from what creators and communities post on Instagram, Facebook and Threads; and Muse has connectors to Instagram (including DMs), Facebook (including Marketplace), Threads and Messenger that no other agent can get. Whether the shopping Muse weighs Meta’s social corpus the same way Meta AI’s shopping mode does is not confirmed. I’d plan as if it does, because the connectors exist and the incentive is obvious.

What comes back. In Debra Aho Williamson’s September hands-on test, she asked Muse for a work tote and watched it browse brand sites and Nordstrom, which briefly flagged it as automated traffic before letting it through. It returned three options. She stopped at the sign-in page, went to Nordstrom herself, and bought four items instead of one. Three is the number I’d assume for the shortlist until someone shows me otherwise.

How Instinct works

Instinct has published far less, so this section separates what’s documented from what’s reported.

The interface is the product. There’s no app. You text it on iMessage or WhatsApp, or you call it, and it texts or calls you first when something needs doing. Noah Shinn’s position is that there are no new interfaces. The continuous thread is the workspace, and the machinery is almost entirely hidden. That’s a different trust model from Muse’s watchable browser: Muse shows you the work, Instinct shows you the result.

The computer. A persistent cloud machine with browser access and cached credentials, kept in a Vault the company says is never used for training. It keeps working between your messages, so a task like "find a handyman, get three quotes, book the best one" can run for hours. Anish Acharya’s description of the architecture, which Instinct hasn’t disputed, is a top-level agent that dispatches work to sub-agents.

The Instinct model. Undisclosed. The Information reported Instinct mostly runs open-source models and wants to own chips and data centers; Instinct’s privacy policy lists third-party AI model providers; Vellum calls it proprietary. Shinn’s academic work (Reflexion, NeurIPS 2023) is about agents that write themselves a note after a failure and use it next time, with no retraining. Whether the product runs that loop is not published, and you shouldn’t assume it does.

What Instinct can do that Muse can’t, yet. Three things, all shipped in a five-week window. Concierge (September 16) places phone calls to businesses with no online booking. The trusted-person network (open to all users September 14) lets your Instinct coordinate with your friends’ Instincts; Shinn said they coordinated 300,000 times in the first week. And on October 7, Instinct announced customer-service partnerships with Sierra and Decagon, so your agent can talk to a business’s agent directly instead of calling. No participating businesses are named yet.

How Instinct shops. Since September 28, Instinct can search Shopify’s merchant network for products with live prices, sizes and availability, then set up the order for a Shop Pay checkout using the user’s saved details. Shinn says 35% of users shop with it and it knows "your sizes, the brands you like" and what you’ve bought before. Instinct Selections, curated product picks that started September 29, is its first visible revenue experiment; the stated plan is a transaction cut, not ads.

Where Instinct looks. Not published. The documented sources are Shopify’s network for products, the open web via its browser, and the phone for everything else. The Resy incident is instructive: one investor’s Instinct pinged the reservation site "hundreds of times every hour" until Resy locked the account. Instinct is the more aggressive errand-runner by design, and it will hit your rate limits before Muse does.

Terms worth reading. Instinct’s terms appoint it as your agent, authorized to enter binding agreements on your behalf, and restrict the product to personal use. A sales rep can use it for their own admin; a company can’t run it as a prospecting tool.

How the user experience differs, and why it changes what you optimize

The differences read as product choices. They are actually different funnels.

Watchable vs. invisible. Muse users can open the browser pane and watch the agent work, which means they sometimes see your site and your competitors’ sites in sequence. Instinct users see a text message. On Muse, a good page can still win a human over mid-task. On Instinct, the page is only ever seen by the machine.

Approval card vs. spending budget. Muse stops at every purchase with a card showing merchant, item, price. Instinct asks you to approve an amount it can spend, then acts within it. The Muse approval is a last-second brand check; a name the person doesn’t recognize gets a "who is this?" and a detour to your site. On Instinct, within budget, there may be no last look at all.

Social context vs. transactional context. Muse arrives with a profile built from Facebook and Instagram (location, relationships, interests) before it even asks for Gmail, plus saved reels, Marketplace messages and DM threads. Instinct arrives with email, calendar, texts and what you’ve bought, kept or returned. Meta’s agent knows what you wanted before you decided. Instinct’s agent knows what you did.

Who pays for it. Muse has no ads today and Meta says VM data stays out of its ad systems, but Meta’s own write-up notes that when Muse visits your site "it will appear as your activity," so your pixel fires and your retargeting pool fills with people who never saw the page. Instinct plans to take a cut of transactions. One of these models rewards being the merchant the agent lands on; the other may eventually reward being the merchant the agent prefers.

Visible, selected, verdict: the new math

AEO as practiced today measures one thing: did your brand appear in the answer. That measure assumes a human reads the answer and clicks something. Action engines break the assumption twice. First, there’s often no answer to read; the output is a booked table or a box on the porch. Second, the person doesn’t choose from the answer; the agent does, and the person approves.

I think about an agent-run purchase as four gates, in order.

1. Recall. The agent already knows a brand. Muse acts on preferences mentioned once and on what’s in connected accounts; Instinct remembers sizes, brands and returns. A remembered brand skips the search entirely. This gate ends the process before it starts, and it’s won by product and brand, not by schema.
2. Retrieval. No brand in memory, so the agent goes looking. This is where today’s AEO work lives, and it still matters, with one addition: for Muse, your presence inside Meta’s own apps is source material.
3. Verdict. The agent returns a shortlist, in practice about three, and selects or recommends one. Being in the three is visibility. Being the one is the verdict. The verdict is decided by what the agent can verify on the page: price, availability, fit against the stated need, policies, and whether the facts match across sources.
4. Completion. Can the agent finish on your site. A login wall, an unlabeled button, a CAPTCHA on the booking form or a bot block ends it, and the agent moves to the next name on the list. Williamson bailed at a sign-in page. So will most agents.

The data on the human side of this is sobering and useful. NMI published a survey of 1,000 US adults in August 2026: 11% had ever used AI to complete a transaction, 10% would give an AI full control of a purchase, 84% had never put payment details into an AI tool, 70% want to review or override before an AI buys anything, and 58% worry it will pick the wrong item, size, color or brand. When an AI buys the wrong thing, 34% blame the retailer. Early adopters are here, the mass market isn’t yet, and the approval step is where brand recognition still earns its keep.

The metric I’d track is selection rate: of the tasks in your category where you appear on the shortlist, how often are you the one chosen or the one the person approves. It’s not observable from the outside yet, which brings us to measurement.

How to optimize for Muse and Instinct

Split the work by gate. Most teams will put all of it in the second one. Don’t.

1. Get remembered

• Give people a reason to name you. Agents inherit preferences stated once. The oldest job in marketing just got a direct line into the funnel.
• Make reorders trivial. An order confirmation in Gmail is a memory the agent reads. Clear product names in your transactional email matter more than they did.
• On Shopify, keep the Shop account relationship tight: Instinct tracks past purchases and surfaces price-drop and back-in-stock alerts tied to it.

2. Get retrieved

• Keep the AEO fundamentals: consistent brand facts across owned, earned and social, entity clarity, structured product data. Earned and social citations carry 22% of the weight in Goodie’s AEO Periodic Table V4, more than any single on-page factor, and that holds here.
• For Muse specifically, treat Instagram, Facebook and Threads as index surfaces. Complete profiles, a catalog with real attributes, creators who say specifics out loud and in captions, Marketplace where the category fits. Meta has said since April that Muse Spark is built to surface and cite what people share on its apps.
• For Shopify merchants, audit the catalog data agents read through WebMCP: title, description, variants, availability. Common Thread Collective’s blunt summary of Shopify’s guidance: listings with missing data get deprioritized.

3. Win the verdict

• Put the facts an agent verifies in plain HTML: price, stock, sizes, hours, shipping, returns. Not in images, not behind a late-loading script, not inside a map widget. Mark them up with Product, Offer and LocalBusiness schema.
• State who the product is for and what it isn’t for. Agents match against a stated need. Specifics beat adjectives.
• Keep facts identical everywhere. An agent that finds one price on your site and another on a reseller or a creator post has a reason to pick the brand with no contradictions.
• Don’t try to talk to the agent in hidden text. Meta classifies that as prompt injection. Real HTML, real labels, no keyword-stuffed ARIA.

4. Let it finish

• Open your site to verified agents. Cloudflare Radar lists a Muse client with a muse- user-agent prefix, but Arcjet found Muse’s browser sending a standard Chrome user agent from a US mobile network. You can’t reliably pick it out by name or IP, so don’t try to block it site-wide. Blanket data-center IP rules miss Muse anyway and catch ChatGPT’s agents, Grok Bot and anyone on a corporate VPN. Meta has not said Muse reads robots.txt, so a robots rule is a statement of preference, not a control.
• Fix checkout for the accessibility tree: real buttons, labeled fields, alt text that says what the thing is, no div-built add-to-cart. Run your key flows through a screen reader; if it can’t get through, neither can the agent.
• Guest checkout plus a mainstream wallet. Link and Shop Pay are what the agents pay with. Forced account creation hands the job back to the user, and the user often doesn’t take it back.
• Move CAPTCHAs off search, product and booking pages to account creation and high-value actions. Nordstrom nearly lost a sale to a bot challenge in a hands-on test.
• If you have an API for bookings, orders, quotes or subscriptions, submit a Muse connector. It’s the only lane where the agent doesn’t have to fight your website, and editorial featuring inside Muse is a form of placement that costs nothing today.
• If you sell through WhatsApp, Messenger or Instagram, keep your Meta business agent current. More than a million businesses run one weekly, and a customer’s Muse may end up talking to it. Follow the Personal Agent Protocol (Meta and Sierra, October 6; Shopify, Stripe, Walmart, Genesys, Rocket and Instinct as launch partners, with NiCE and Decagon joining the working group). It’s OAuth-based, lets you choose website, API or your own agent as the entry point, and gives you visibility into what agents do on your site. No spec is published yet; decide your preferred paths now so you’re ready when it is.
• Answer the phone like an agent might be calling. Both products place calls to businesses. Staff should be able to confirm availability, hours and pricing to a voice that is checking facts, not browsing.

5. Measure what you can

• Grep a week of logs for muse- and note which pages it hits, knowing it will undercount: in Arcjet’s test, Muse’s browser didn’t identify itself. Agent visits on pricing, booking and product pages are customers trying to reach you.
• Ask your payments team whether orders paid with Link single-use cards can be identified. That may be the cleanest marker of an agent-run purchase you get. Shop Pay orders from agent sessions are the other.
• Segment analytics by network. Muse runs a real browser, so its sessions look like short, direct visits from ordinary US mobile and residential connections. Don’t filter them as bots; don’t let them pollute time-on-page either.
• Run your own category by hand. Have three people with different profiles give Muse and Instinct the same errands, record the browser pane, and write down what gets shortlisted and where it stalls. The sample is tiny; it’s still the only instrument that sees the verdict today. External tools, Goodie included, can sample Muse Spark and track how the model describes you, but they can’t see a specific person’s Muse with their memory and connected accounts attached. Anyone claiming a "Muse share of voice" dashboard is measuring the model, not the agent.
• Watch retargeting creative. Audiences will now include people whose agent visited and who never saw the page. "Still thinking it over?" lands on someone with nothing to remember.

What I don’t know yet

How Muse ranks its shortlist. Meta hasn’t published it; my guess is some mix of the stated need, what it can verify on the page, and price, and I can’t prove that. Whether placement in Muse ever goes up for sale; there are no ads today, and the featured slots in the connector directory are the obvious first place. How much of Instinct’s 10% daily growth survives contact with a public launch, and whether a 14-person company can hold $1 billion of transactions together on rented compute. Whether either agent preserves affiliate parameters when it browses. And whether the social-corpus advantage Meta has documented for Meta AI’s shopping mode carries into Muse on the open web. I’d bet yes on the last one, with no proof.

The bottom line

The last two years of AI search were about being in the answer. The next two are about being the answer, for an agent that doesn’t care how you describe yourself and checks whether the page agrees. Fix the facts, open the door, label the buttons. Then go run your own category through both of them and tell me what they picked.