Meta's Muse agent is instructed to rank well-known merchants first, cut listings priced outside a category's normal range and check no more than three stores on the open web for each request, according to an instruction file that AI marketing platform Profound says it pulled from the agent on September 25 and shared with State of Brand, which published its analysis on October 8, 2026.

In Short

Meta built a shopping helper called Muse, and a set of hidden rules tells it to show products from shops people already know before anyone else. Small or new brands still get shown, but they start lower on the list, and only products in Meta's own catalog can be bought without leaving the chat. If you run a newer brand, being known by name and being in Meta's catalog now matters more inside this agent than having the lowest price.

What the file says

The State of Brand article, published by the Outlever-run brand newsroom on October 8, rests on a single document: a set of shopping instructions that Profound says its research team found in Muse's system files. Profound shared the full text with the publication. Profound sells software that tracks how AI agents treat brands, a commercial interest State of Brand discloses in its own piece, and Profound itself cautions that the file may have changed since late September.

Taken at face value, the document describes a ranking order that runs in a fixed sequence. Muse's first instruction is to save the shopper money. When it reviews results, it discards anything that does not match the request, does not look high quality, or sits outside the normal price range for that type of product. That last filter cuts in both directions. A listing priced far above the category is removed, and so is one priced far below it - which, as State of Brand observes, is often exactly where a newer brand prices to get noticed.

Then comes the ranking. According to State of Brand, the instruction file returns to brand recognition three separate times. Well-known sellers are named in the ranking step of the discovery workflow. The instructions for displaying results call it "imperative" that the top five products come from well-known merchants or websites and match what the shopper asked for. The formatting rules repeat the point a third time.

Two further preferences reinforce the pattern. When a shopper names a brand or retailer, Muse's catalog search places that brand's own store at the top and fills in other sellers below it. By default, the agent also favors brands and retailers that sell direct over resellers.

None of these choices is unusual in itself. A human merchandiser at a department store would make most of the same calls. The difference is scale and timing: Muse applies the rules for every shopper, on every request, before any option has been seen. For a brand still building recognition, the cost of being less known used to be measured in clicks. Inside Muse, State of Brand argues, it is measured in rank.

The catalog does most of the work

Profound published separate research on Muse and Instinct, a Meta personal agent that operates over text messages, on October 2. That study ran 50 prompts from one retail category through a single Muse account. Profound describes the sample as small and exploratory, and the numbers carry that caveat.

On a typical prompt, Meta's catalog returned about 60 products from 30 merchants. Muse's browser, working the open web in parallel, visited three sites and found four products at two merchants. The six products Muse ended up showing usually split four from the browser and two from the catalog.

So the catalog supplies the volume and the web supplies the picks. In the median case, the catalog provided about 95% of the candidates, according to Jennifer Zou, an economist at Profound with a PhD from Harvard, who suggested the imbalance probably reflects that the catalog is cheaper to search. Yet the final selection tilts the other way. "Much more weight is placed on candidates returned by the web search," Zou said.

The instruction file goes a long way toward explaining that gap, according to State of Brand. The browser search is narrow by design, and what survives it has already passed a merchant-recognition screen.

Three stores and an early exit

Muse is told to run a browser search alongside the catalog search on every request, "especially for home goods," unless the shopper asked only for Facebook Marketplace listings.

The search itself is small. By default, the browser agent looks at no more than three merchant sites and returns no more than ten products. If two sites supply enough for a "high-quality and seller-diverse" list, the agent is told to stop early.

Everything the browser brings back must clear a set of checks. The link has to point to an actual product page rather than a search results page. The page has to show the item is in stock and can be added to a cart. The agent has to be able to verify a real product image, and placeholders and logos are discarded. Muse also prefers the country version of a site that matches the shopper's location, and it reopens every link itself before displaying anything.

What the file does not say is how the three stores are chosen in the first place. A brand that misses that cut is never compared at all - a point State of Brand makes plainly, and one Profound's research cannot answer from a single account.

Meta keeps checkout for its own catalog

The purchase path splits along the same line as the search. Products in Meta's catalog that carry an "agentic checkout" flag can be bought inside Muse through a Shopify-based purchase flow. Anything Muse finds on the open web goes through browser checkout instead, with the agent clicking through the merchant's own site, and the product card for that item has in-chat checkout switched off.

Carts follow the same rule. Muse will build a cart for a shopper to return to, but only from flagged catalog products. "Products without that capability have no cart, and neither does browser checkout," the file says.

That is a notable turn for Meta. In June 2025, the company began moving Facebook and Instagram Shops from on-platform checkout to website checkout, ending payment processing and order management for shops and sending buyers to merchant sites. Little more than a year later, its newest consumer product puts checkout back inside a Meta surface - provided the product sits in Meta's catalog.

The catalog is home ground for Meta. It powers shopping on Instagram and Facebook, and the instruction file rates its coverage as good for fashion, home decor and beauty, and only okay for everything else. Muse can also pull Facebook Marketplace listings. If a shopper shares an Instagram post with tagged products, Muse goes straight to those items and is told to skip search entirely.

The practical split, as State of Brand frames it: a brand on the open web can still be recommended, but being bought or carted inside Muse requires a listing in Meta's catalog.

How the Shopify-based flow is wired has not been confirmed by Meta. When Muse went live in the United States on September 8, an independent code analysis found three checkout modes - browser, Shopify and Stripe Link, with the Shopify path most likely running on Shop Pay over UCP, the open protocol that lets agents build carts and pay without a custom integration for each merchant. Meta was among the companies that joined the UCP Tech Council on April 24, 2026, alongside Amazon, Microsoft, Salesforce and Stripe.

What Muse knows about the shopper

Before it shops, Muse reads a stored profile of the shopper's preferences, checks a general file about the user and searches its memory for sizes and tastes. For gifts, it works from what the shopper says about the recipient plus a notes page it keeps on that person.

Some details must be settled before any search runs: the wearer's gender and size for clothing, for instance, or the exact device a part needs to fit. Muse asks about them one at a time. The file tells the agent never to infer someone's gender from their name and never to fill in a default value. A shoe size saved for one brand does not carry over to another.

How much all of this changes results remains unmeasured. Profound's pitch to State of Brand leaned heavily on personalization, the publication notes, but the firm's test used a single account, so it cannot say how far two shoppers' results diverge for the same request. Zou said Profound is building a panel of synthetic user profiles to find out.

An agent that does not announce itself

Even when Muse lands on a brand's site, the brand probably will not know. According to Profound, Muse and Instinct do not identify themselves when they browse. They route through residential internet connections, appear as ordinary visitors, and their answers do not cite the sites they read.

That behavior has already produced a confrontation. Amazon blocked Muse from shopping on Amazon.com, saying in part that the agent did not identify itself as a bot, according to State of Brand. Amazon's position is not new. Its Agent Policy, which took effect on March 4, 2026, requires AI agents to "clearly identify themselves as automated systems at all times." The block landed during the week to September 27, less than three weeks after Muse's release.

Profound attempted to estimate the traffic anyway. In its tests, some Muse sessions came immediately after requests labeled "meta-webindexer" from IP addresses Meta does not publicly claim. Meta-WebIndexer is the crawler Meta uses to build the search index behind Meta AI. In Muse's first 20 days, those requests grew 1.6 times as much on US and Canadian retail sites as on sites elsewhere. The comparison covers about 280 sites sorted by domain suffix, and Profound calls it only a proxy.

The result, in State of Brand's words, is an agent judging product pages on behalf of real customers with no clear record of the visit in a merchant's analytics. Practitioners who rely on declared user agent strings to separate bots from people have little to filter on if the visitor looks like a household browser.

How many shoppers are affected

Scale is the reason a 50-prompt study and a single leaked file attract attention. Muse has been downloaded more than 2.5 million times since its September 8 release, Profound reported, citing CNBC.

That figure sits well below other published estimates, and the gap likely reflects different measurement windows and markets. Sensor Tower data covered by PPC Land on October 4 put Muse at more than 5 million installs across the US and Canada within 22 days, with Muse ranking as the top app overall on the US App Store and Google Play for 12 days. Meta has published no install or usage figures of its own, so every number in circulation is a third-party estimate.

The audience may also widen beyond consumers. On September 28, Meta started selling the same agent technology to businesses, according to State of Brand. That date matches the creation of Meta Enterprise Platform, led by former MongoDB chief executive CJ Desai, which named Muse among its first four products - though Meta has published no price, release date or market list for the enterprise versions.

Whatever rules sit inside Muse, then, now shape what a large number of shoppers see first.

Why this matters for marketers

For roughly a decade, the challenger brand's route to market ran through discovery. A better product or a lower price, found through search or Instagram, could take share from incumbent names. State of Brand's reading of the file is that Muse works against nearly every step of that route: it promotes recognition, trims price outliers, narrows the open-web search to three stores and keeps in-chat purchase for its own catalog.

The pattern is familiar from AI search more broadly. PPC Land has tracked how answer engines concentrate visibility among established sources - a GetCited study published in September found that 86% of AI "best product" answers cited at least one source with a disclosed commercial interest. Muse's instruction file is different in kind. Rather than an emergent tendency inferred from citation counts, it is a written rule that names merchant recognition as a ranking criterion.

It also lands as rival platforms push their own agentic checkout into merchant stacks. Google, for example, switched on native checkout in AI Mode and Gemini for Shopify merchants from September 18, with an opt-out rather than an opt-in. Each platform is drawing the line between "recommended" and "buyable in place" in a slightly different spot, and catalog integration increasingly decides which side a product lands on.

Meta's advertising model sits in the background. Meta has said Muse conversations and virtual-machine data stay out of its ad systems, and MBI Deep Dives argued on October 1 that Muse may never need to carry ads at all. A ranking that rewards brand familiarity, however, rewards exactly what brand advertising on Facebook and Instagram is sold to build. That link is an inference, not something the file states.

What Profound and State of Brand say brands can control

Zou's view splits by company size. "If you're a well-known merchant/retailer, it's probably fine to rely on discovery via browser search," she said. "If you're a smaller brand, Shopify/Meta catalog integration is a more worthwhile investment."

State of Brand accepts that reasoning as far as it goes but notes its limit. For a smaller brand, the catalog is the way into the candidate pool and the only route to being bought or carted inside Muse. It does not change the ranking rules. Once a product is in the pool, well-known sellers still go to the top.

The publication's own list of controllable factors is largely operational, mapped directly to the file's checks: checkout enabled in Meta's catalog; product pages with one item per page, a real image and visible stock status; pricing within the category's normal range; explicit sizing, fit and compatibility data, which Muse must resolve before it searches; and tagged products on Instagram posts, since a shared tagged post skips search entirely.

None of those steps moves a challenger to the top of the list, and State of Brand is direct about what does. Muse places well-known sellers first and places a brand's own store first when a shopper asks for it by name. Getting people to ask for a brand by name is the oldest job in brand marketing. Inside Muse, it is also the most direct way around the rules Meta wrote.

What remains unknown

Several questions are open. Meta has not commented on the file, and it is not clear whether the instructions Profound captured on September 25 are still in force. Nothing in the document defines "well-known," leaving open whether the agent relies on Meta's own data, web signals or the underlying model's general knowledge. The criteria for selecting the three open-web stores are absent. And with a single account and 50 prompts, Profound's study cannot show how much personalization reorders results between shoppers - the question its synthetic-profile panel is meant to address.

Timeline

Summary

Who: Meta, whose Muse agent is governed by the instruction file; Profound, the AI marketing platform that says it extracted the file and studied Muse's behavior; State of Brand, which published the analysis; and merchants, particularly smaller and challenger brands competing for placement.

What: An instruction file attributed to Muse tells the agent to rank well-known merchants first, calling it "imperative" that the top five products come from well-known merchants or websites, to discard listings priced outside a category's normal range, to check no more than three open-web stores per request, and to restrict in-chat checkout and carts to Meta catalog products carrying an "agentic checkout" flag.

When: Profound says it pulled the file on September 25, 2026, published related research on October 2, 2026, and State of Brand published its analysis on October 8, 2026.

Where: Muse, available in the United States and Canada via iOS, Android, muse.ai and WhatsApp, drawing on Meta's Instagram and Facebook catalog, Facebook Marketplace and the open web.

Why: Muse has been downloaded millions of times since its September 8 release, so its ranking rules shape which products a large number of shoppers see first. The rules favor recognized brands and Meta-catalog merchants, while the agent's unidentified browsing leaves merchants with little visibility into how their pages are evaluated.