Episode 192: The PE Guy and Cameo CEO Steven Galanis Close Out Marketecture Live Chicago

Full analysis

Meta shipped Muse, a personal AI agent app that buys things for you. It plugs into Shopify, Walmart, PayPal, Instacart, and Expedia out of the gate. Amazon blocked it. That last detail matters more than the download numbers everyone is quoting.

What's actually being decided here: nothing you commit to today is hard to undo. Muse is free, it's a few days old, and the interesting claims (parity with ChatGPT's launch, $200 billion in market cap restored) are unverified podcast talk. But the structural question Ari Paparo raised is real and doesn't have a deadline: if an AI agent decides what you buy, who controls the shelf, and what happens to advertising when a bot does the shopping? That's worth thinking through now, before the buying habits harden.

The Skeptic. Eric Franchi says Muse is "on par with same time period of downloads for ChatGPT, higher in DAUs." Come on. Meta has 3 billion existing users and put Muse on the Connect stage with the full promo machine behind it. ChatGPT launched cold with a tweet. Comparing the two launch curves is meaningless. The 2.5 million figure from TechCrunch is real; the "beating OpenAI" framing is not. And the money numbers Franchi tossed out, a $5 billion Scale deal and $200 billion in market cap, are exactly the kind of round figures that sound precise and check out to nobody. Treat the traction as real. Treat the superlatives as marketing.

The Builder. Muse is genuinely different under the hood, and John Gruber flagged why: every user gets their own persistent Linux VM in Meta's cloud, a full private computer that keeps running between sessions. That's not a chatbot with a plugin. It's a machine that can hold state, log into your accounts, and actually complete a checkout. For anyone building agent features, that's the reference design now. The lesson is that "agent that does the thing" needs real compute per user, not a stateless API call. If you're wiring an agent into your product, Muse just reset what customers will expect: it should finish the task, not hand back a link.

The Compute Pragmatist. A persistent Linux VM per user is expensive. Meta says Muse stays free, you buy extra tokens (the units of text the model processes, which you pay for by volume), and "if you're using it for business, you'll burn through them." That pricing tells you the real cost lives in the compute. The model is almost beside the point. Meta can eat that because the data is the payoff: they now see what you buy, not just what you scroll. Nobody else building agents has 3 billion users to subsidize the VM bill against ad and commerce upside. This is a rich-lab feature. A startup running the same architecture at scale would go broke before it found a business model.

The Enterprise Buyer. The eMarketer finding from Stephanie Padernacht is the practical takeaway, and it has nothing to do with Muse. Google Gemini pulls brand facts mostly from news and editorial sites. OpenAI's ChatGPT pulls mostly from retail sites and reviews. Same question, different sources, so the same brand needs two different visibility playbooks depending on which AI the customer uses. Worse, the places people actually spend time, YouTube and Netflix, feed almost none of these answers. If you sell anything and you're planning "how do we show up in AI answers," this is the map. Two engines, two strategies, and your biggest reach channels don't count.

The tension that matters: does Muse help Meta's ad business or eat it? Paparo laid out both forks cleanly. If Muse ignores your Instagram data and just shops well, then the better it gets, the less anyone needs an ad, and it competes with Meta's own core business. Or Meta rigs it so a product is only eligible to be bought through Muse if the brand advertises, which turns the agent into a new ad slot. Meta will pick door two. They always monetize the surface. The Amazon block is the giveaway: Amazon won't let Meta's agent see its catalog because whoever owns the agent owns the demand, and Amazon isn't handing that to a competitor.

What this hinges on: whether an AI agent that shops for you becomes a real ad surface or a real ad killer. The whole open web has spent two years worried about AI answers eating referral traffic. Agentic commerce is the next version of that fight, one layer down, at the checkout instead of the search box. Everything else in this episode, the distressed M&A (Infillion buying Foursquare, Taboola buying Dianomi), the new commerce networks from Citi and McDonald's, Scope3 renaming itself Apostia on $1.3 million of transactions, is small and mostly noise. Muse is the story because it moves the buying decision itself inside a Meta-controlled box. Prediction: By Meta's Q4 2026 earnings call in late January 2027, Meta will have launched a paid placement product that lets brands influence what Muse recommends or buys, tied to advertiser spend.

Confidence: Medium — Meta always monetizes a consumer surface; timing is the only real risk.

Why: Meta gave Muse away free while quietly metering tokens, which means the compute per user is expensive and the payoff has to come from somewhere else. Meta's entire business, 98% of revenue, is advertising, and every consumer surface it has ever built (Feed, Stories, Reels, Marketplace) eventually became an ad slot. Paparo named the exact mechanism on the podcast: a product becomes "eligible to be bought by Muse if you advertise." The Amazon block confirms the value at stake, because Amazon will not let a rival's agent decide purchases inside its catalog. The opposite outcome, Meta running a neutral agent that ignores advertiser money and quietly cannibalizes its own ad business, would break every pattern in the company's history and leave the compute bill unpaid.

Revisit by 2027-01-31: We're right if Meta has announced or launched any advertiser-paid mechanism to influence Muse's product recommendations or purchases by the Q4 2026 earnings call. We're wrong if Muse's recommendations remain free of paid placement and Meta has shipped no such product by then.