# Meta’s Muse Lands in Australia Without Coles, Woolworths or Amazon

> Source: <https://rews.cc/a/meta-s-muse-lands-in-australia-without-coles-woolworths-or-a-a05357>
> Published: 2026-11-20 18:00:00+00:00

**Recommendation:** Design shopping agents around product and search browsing, where 85% of observed agent traffic goes, and treat checkout as a small, separately guarded path that requires explicit user approval before spending.
More in the playbook "Designing agent systems": <https://wpnews.pro/playbooks/agent-architecture>

Meta’s personal AI agent Muse went live in Australia on 17 November, ten weeks after its 8 September debut in the US and three weeks after Singapore, with nine local “connector” partners: Canva, Domain, realestate.com.au, Kogan, MECCA, Officeworks, Ticketek, SWEAT and the cashback platform ShopBack. What is missing matters as much as what is there. No Coles, no Woolworths, no airline, no food-delivery app, and no Amazon — which blocked Muse from crawling its store in September, according to [Meta’s own update to its connector list](https://ai.meta.com/muse/). Meta’s Australian launch post gives no Australian-dollar price, saying only that the product is “free for most of what people need, with subscription plans for people who want to do more.”

Muse is not a chatbot you type into. It is an agent: software that holds a goal, calls on external services and takes multi-step action without a human doing each step. Each user’s Muse runs on what Meta calls a Secure VM, a dedicated cloud machine that stores the passwords and payment credentials for whatever services the user links, so the agent itself never sees raw credentials. A connector is the plumbing to one of those services — Gmail, Google Calendar, OpenTable, or a retailer’s catalog and checkout. Before Muse sends an email or spends money, it is supposed to show the user what it is about to do and wait for a yes. That approval step, and the question of what happens when it fails, is the same mechanism this publication examined when [an AI shopping agent’s mistake became the account holder’s problem](https://rews.cc/a/your-ai-shopping-agent-counts-as-an-authorized-user-and-so-d-f0f268): the agent acts as an authorised user, and so do its errors.

How many people are actually using it depends on whose count you read, and the three available counts measure different things. Sensor Tower, tracking app-store installs, put Muse at 10 million downloads in the US within 40 days of launch. Mark Zuckerberg told investors in October that more than 10 million people in the US and Canada had set up a Muse agent — a self-reported figure from the company that built it, not an independent audit. Similarweb, measuring web and app activity, estimated more than 10.24 million daily active users globally. The coincidence of all three landing near 10 million is suggestive but not proof of consistency: downloads, account setups and daily actives are three different funnels, and none of the three firms publishes a methodology note reconciling them.

The clearest independent read on how much of the agent web Muse now accounts for comes from HUMAN Security, a bot-detection firm that inspects live traffic on the sites and apps it protects rather than inferring activity from app-store numbers. Its [October 2026 State of Agentic Traffic report](https://www.humansecurity.com/learn/blog/october-2026-agentic-traffic/) found total agent requests on those properties ran roughly eight times August’s volume and 2.6 times September’s. Muse alone accounted for 59% of October’s agent requests, up 19 percentage points from September, with request volume nearly quadrupling month over month — 3.7 times the volume of ChatGPT Agent, the next-largest agent. ChatGPT Agent’s own request volume rose 44% over the same month, but its share of the total still fell, from 28.5% to 15.9%, because Muse grew faster. Perplexity’s Comet, which held 41% of agent requests in August, fell to 3.6% by October. Meta and OpenAI combined now account for close to 90% of the agent traffic HUMAN sees.

HUMAN’s report also undercuts the premise that agents are already doing much buying: 85% of October’s agent traffic went to product and search pages, up from 81% in September, while checkout and payment pages took just 1.9%. Agents are mostly browsing, not spending — for now.

## How Muse picks a winner

The mechanics of how Muse decides which merchant to show a shopper matter more than the download count, because that ranking is what the whole connector business model is built on, and Meta has published none of it. The AI-visibility firm Profound says it extracted an internal instruction file from Muse’s system on 25 September and [published its contents](https://ppc.land/metas-shopping-agent-must-fill-its-top-5-picks-with-well-known-merchants/): the file calls it “imperative” that the agent’s top five product results come from well-known merchants, a rule repeated in three separate places — the ranking step, the display instructions and the formatting rules — without ever defining what counts as well-known. The file also reportedly tells Muse to discard listings priced outside a category’s normal range, to check no more than three open-web stores per query, and to restrict in-chat checkout to catalog products carrying an “agentic checkout” flag. Meta has not confirmed the file exists, has not commented on it, and has published nothing of its own on how Muse ranks merchants. Profound also sells software that helps brands optimise for exactly this kind of agent, which gives it a commercial interest in the story being told.

Profound’s separate 50-prompt test of the live agent found a typical shopping query pulled roughly 60 candidate products from about 30 merchants in Meta’s own catalog, while Muse’s browser visited three outside sites and surfaced around four more products from two merchants. The catalog supplies about 95% of the raw candidates but wins only about a third of the final slots; the browser’s much smaller set of live-web results wins the rest. Profound’s economist, Jennifer Zou, said “much more weight is placed on candidates returned by the web search” than the catalog volume alone would predict.

**How a Muse shopping query gets its six picks** Based on Profound's 50-prompt test and leaked instruction file, not confirmed by Meta

| From | To | How | 
|---|---|---|
| Shopper's query | Meta catalog *(~60 candidates, ~30 merchants, ~95% of pool)* |  | 
| Shopper's query | Live web browse *(≤3 sites checked, ~4 products, 2 merchants)* |  | 
| Meta catalog *(~60 candidates, ~30 merchants, ~95% of pool)* | Ranking rule *(top 5 must be "well-known" merchants)* | bulk of candidates | 
| Live web browse *(≤3 sites checked, ~4 products, 2 merchants)* | Ranking rule *(top 5 must be "well-known" merchants)* | outsized weight | 
| Ranking rule *(top 5 must be "well-known" merchants)* | Final 6 picks shown *(usually 4 from browser, 2 from catalog)* |  | 

If the leaked rule holds, it inverts the route smaller brands have used for a decade to beat incumbents — build a cheaper or better product, get found through search or social, take share from the name everyone already knows. A rule that fills most of the top five with recognised merchants makes that route harder before a shopper sees a price. An industry has already formed around the gap: Azoma, one of several platforms selling what the trade calls agentic commerce optimisation, pitches brands on getting cited by Muse, ChatGPT, Amazon’s Rufus and Walmart’s Sparky at once. One AEO framework cited in that space weights earned and social citations — mentions on other sites and platforms the brand doesn’t control — at 22% of visibility, more than any single factor on a brand’s own site, and notes that for Muse specifically, a brand’s presence inside Meta’s own apps counts as source material the agent can draw on. None of these weightings come from Meta; they are vendors’ reverse-engineering of a system whose actual rules remain unpublished.

The Australian rollout sharpens that asymmetry. ShopBack, a connector partner, already runs on a cut of the transactions it refers — the same economics examined in [a look at agents wanting a fee on every purchase](https://rews.cc/a/ai-shopping-agents-want-a-fee-on-every-purchase-you-just-don-3e544c): Muse now gives that fee-taking layer a direct line into a shopper’s agent rather than a browser extension. Payments inside Muse run through Stripe’s Link system, with Meta saying the agent is the first covered by Stripe’s purchase warranties and returns if it buys the wrong thing. Will Easton, Meta’s managing director for Australia and New Zealand, tied the timing to the run-up to Black Friday on 27 November: “With the festive season coming up, the timing couldn’t be better,” he said, adding that people keep control “including choosing what Muse can access and approving purchases before they happen.”

With the festive season coming up, the timing couldn’t be better.

Easton’s framing — gift shopping, approval before purchase — describes the interface, not the ranking underneath it. No Australian-specific breakdown of Muse’s local traffic or merchant selection exists yet; HUMAN Security’s own report measures a global protected-site sample, not Australia alone, and the firm has said its next edition will cover the Black Friday and Cyber Monday weekend specifically. That is the number to watch: whether Muse’s 59% share of agent requests and its unexplained preference for well-known merchants show up in Australian retailers’ agent-referred sales before the big local chains that still aren’t connected — Coles, Woolworths, Amazon — decide whether to let the agent in at all.
