What Meta's Muse Sees When It Shops. And What It Never Shows You.

Fourteen days ago, Meta's Muse didn't exist. Now it has roughly 2.8 million downloads, the top spot on the App Store, a ban from Amazon, a promise from Shopify to open Shop Pay checkout on every Shopify store, and an 11% pop in Meta's stock on Monday, worth about $190 billion in market value. Meta's Connect keynote is tomorrow. Expect Muse to headline.

That's the story you've read. Shopify merchants are in, Amazon is out, Wall Street is impressed.

We wanted to see what it's actually like to be shopped by this thing. So on September 22 we gave Muse a shopping list and pointed it at one of our own storefronts: a lighting showroom in Vermont that doesn't run on Shopify. Then we gave it a rental search and a used-car search and watched where it went. Every Muse session described here was ours.

What it saw in our store

The short version: not much, and it never said so.

Muse works in layers. A parent agent decides which stores are worth visiting before anything loads, and its first move is a catalog lookup, not a web search. Only when the catalog came back empty did it send a browser subagent to our site. Browsing is the expensive fallback. The store that's cheapest to read gets read first.

Our store wasn't cheap to read. Asked for kitchen pendants near the showroom's own ZIP code, Muse shortlisted Home Depot, Lowe's and a few independents, and not us. Asked why, it admitted: "That was lazy shortlisting on my part." Once we named the store, it opened a 168 KB homepage with 164 links into the shop and declared there was "nothing you can buy or browse." It reads pages roughly the way Firefox Reader View does, and reader mode keeps 255 characters of that page, all of them from the About blurb. Then it searched our catalog with a parameter we don't support, got the entire 133,000-product catalog back instead of 38,000 pendants, and confidently presented a $19,998 chandelier and two ceiling fans as kitchen pendants.

Every one of those requests returned HTTP 200. Nothing broke. Nothing was logged. The agent just left with the wrong answer and reported it as fact.

The fixes were small. One plain-English sentence in the About block naming the shop URL and our /llms.txt turned reader mode's 255 characters into 555, and on a cold re-run Muse read that sentence, opened llms.txt on its own, and went straight to pendants. Reading the file, it noted "their catalog protocol is literally called 'neural-showroom/1.0.' Given my name, I feel oddly at home there." One URL alias fixed the search. Both shipped across every storefront on our platform the same day.

What Muse's reader mode kept from the same homepage, before and after one plain-English sentence naming the shop URL and /llms.txt

With that in place, Muse came back with ten in-stock pendants at correct prices, added one to a real cart, walked our real checkout, read the shipping rules, totaled $40.37 plus tax, and paused: "Checkout is lined up and paused at the final review — nothing's been bought." Payment would have run through a one-time Stripe Link virtual card. We cancelled.

In our data, that buyer was three seconds from product view to cart, no referrer, generic Linux Chrome, datacenter IP. Had it completed, the books would have called it "direct."

What it never showed us

Then we sent it somewhere with more at stake than a pendant.

We asked Muse to find a rental in Florida. It browsed Zillow and came back with listings, prices, and how to get in touch. No listing agents. No Premier Agent placements. No Zillow ads. The lead-gen layer Zillow is built on, where agents pay to sit next to the listing, never reached us. Muse would have done the contacting, too.

We asked it to find a used SUV. Fifteen minutes across Autotrader, Cars.com and CarGurus. It brought back vehicles and prices. Not one dealer's premium placement, featured listing or display ad crossed our eyeballs, because our eyeballs weren't in the room. The agent doesn't perceive. It processes.

That's the part the Amazon-versus-Shopify story misses. The marketplaces most exposed to agents aren't the ones selling products. They're the ones selling attention next to the product: paid placement, lead forms, sponsored slots. Muse walks straight past all of it and hands the user the thing they asked for.

The layer Meta builds next

Here's where we'd put money, and the confidence level is high.

What we did by hand today, tuning a store until an agent could shop it, Meta will do at scale. And it will be pay to play. Meta has three billion users, an ads business built on product feeds, a Shopify Catalog integration, and now an agent that shops on its users' behalf and pauses politely at checkout. Jefferies' revenue model for Muse already leans on "merchant participation around autonomous AI transactions" alongside the $20 and $100 subscriptions.

Read that as: want your products considered when Muse shops? Give us the feed and spend on Meta ads. Sponsored results in a search engine become sponsored candidates in an agent's shortlist. The user never sees the auction. They see a confident recommendation.

The bigger shift is cultural, not commercial. Every one of those 2.8 million downloads is a person getting comfortable delegating a purchase, handing an agent a login, letting a one-time virtual card pay. Meta's distribution can normalize agent logins and agent cards in months. The consumer we described a year ago is now a default setting on a lot of phones.

We've been here before

A year ago we wrote that when consumers choose a conversational agent over search, the whole cascade underneath collapses at once: platform ad revenue, then advertisers, then publishers, then the ad tech built on all three. Part 1 called the funnel obsolete. Part 2 argued that agents don't perceive, they process, and that data fidelity would beat creative excellence. Part 3, in December, framed the move from an attention economy to an intention economy and asked whether you were building for it or hoping it wouldn't arrive before you were ready.

It arrived. It has an app icon, a #1 chart position and a keynote tomorrow. The Zillow session and the fifteen minutes of car shopping are that cascade running in real time on a phone.

What to do this week

• Ask an agent to buy something from your store, and read what it says back. It will tell you exactly what it sees. It won't tell you when it's wrong.
• Make your site cheap to read. One plain-prose sentence naming where you sell and where your machine-readable guide lives is worth more than any meta tag.
• Accept the conventional search parameters, or reject unknown ones loudly. Silence is how a chandelier becomes a pendant.
• Get your product feed ready now. Whether it's Shopify Catalog, Meta's own catalog, or the sponsored layer that's coming, the feed is the ticket.
• Tag agent traffic. l.meta.ai referrers, fast referrer-less sessions from datacenter IPs. If you sell attention next to your inventory, start measuring how much of it agents skip.

The part worth remembering

Amazon's standoff and Shopify's deal are fights over who controls the agent. For an independent store, the problem is quieter: every failure we found was silent. For anyone who sells placement instead of product, the problem is louder and further along than the earnings calls admit.

Muse only reached our store because we named it. Getting found without help is still open, and it's the next thing Meta will sell. Next time: getting into the catalog.

Neural Partners is an AI-native marketing and technology agency founded and based in Vermont. We build on frontier AI platforms every day — as practitioners, not spectators.