gaut@0xgaut4h ago

📝 Muse is your next customer. How does it decide what to buy? Personal AI agents are here, and they’re shopping for us. Muse has been downloaded more than 5 million times in just 22 days. That may feel—and is—a tiny number for now, but my gut says we’re just reaching an inflection point. AI is no longer just searching for us. It’s choosing for us. That’s a good reason to figure out how these agents think today, before they become the majority of our sites’ traffic. When we delegate shopping to agents, they decide which products we see and which sellers they consider. That begs the question(s): Who determines how Muse shops? What influences its decisions? What makes it choose one product over another? I don’t have all the answers, but I do have some leads. Unlike chatgpt.com, personal agents are less constrained by response speed and a model’s training data, so they can do more work to find an answer. By the time Muse starts your search, it’s already made a bunch of decisions: it has gathered your preferences and information from previous conversations, along with instructions on where to search, which stores to look at, which products to compare, and how to choose. Those instructions determine which products win the shopper’s attention and, ultimately, their purchase. How does the Muse search sausage get made? Luckily for us, someone has already dug into Muse’s literal instructions (Muse’s brain, if you will). Bare in mind Meta can change these at any time, and they may be outdated by now. When you send a prompt, Muse first takes a look at what it remembers about you: your shoe size, your preferences, and relevant context that might improve the search. Then, Muse goes down two rabbit holes: Meta’s product catalog, which contains products and merchants for Facebook and Instagram, and a browser it can use to search vendors outside that catalog. The catalog is neatly organized by product name, photos, prices, and availability. In the browser, Muse can look at up to three merchant sites and ten products, unless the user gives it more specific instructions. Once it finds products, it’s asked to filter them. It removes bad matches, decides if a product is trustworthy, and tries to make sure the item is in stock. Once that’s finished, it ranks each result by relevance, fit, and seller trust. Who sells the product is explicitly part of the process. If you ask for a deal, Muse has an extra round of work to understand typical price, find other offers, and check for coupons. So what actually happens when you search? Since I’m a curious guy, I had Codex fire off 25 shopping prompts to my Muse across categories like kettles, lamps, speakers, towels, and backpacks. I got 125 recommendations back, along with a few anecdotal ideas about its behavior: - Seller reputation matters: In one example, Muse picked a $25 kettle over a $19 one, citing seller reputation and trust. Both were on sale at Walmart, but the $25 was manufacturer sold, while the cheaper one came from a third party. - Price isn’t king: The lowest-priced option came in first 9 times out of 25. Muse prioritizes something else. - Muse ignores products that don’t meet the user’s needs: Muse cares a lot about not recommending something clearly out of bounds. - The same question can have different winners: If you run the same prompt multiple times, Muse changes its mind unpredictably. - Finding a product didn’t mean I could buy it: For example, Muse recommended a speaker from the Coast Guard Exchange, which restricts purchases to eligible shoppers. - If you name the product, Muse will oblige: It puts a lot of weight on users naming the brand / product directly, which makes sense because it shows clear intent. - Muse can give up: One backpack search returned five results, all from the same REI website, after Muse gave up on searching alternatives it couldn’t access. What does this mean for your brand and products? We need to prepare for a world where personal agents are the new middlemen by catering to their preferences. Your brand needs to stand out by appearing in ways the agent prioritizes and by organizing your information in ways it expects (like frameworks). For Muse, the specific answer depends on whether you’re an established brand or a smaller business. For smaller businesses, having your products listed in Shopify and Meta’s catalog may help. In either case, you want to make sure agents can find your products. Another edge lies in your product descriptions: arm Muse with as many facts as possible to help your product win the recommendation. If you’re selling a backpack, you want expansive descriptions that go above and beyond in describing the type of person who should buy it (descriptions that might sound a bit ridiculous to a human reader). You can essentially work backward from a user’s prompt, like “what’s the best backpack for commuting in the rain on a bike?” and ask yourself: what would Muse want to see in a description to pick that product? Is this the same as SEO/AEO and AI traffic optimization? Yes, and also no. Muse and other personal agents rely on many of the same tools you’ve already optimized, like your website and blog content. Much of what makes your existing page good for search—clear descriptions, accurate specs, and availability—also helps the agents choose it. But agents are increasingly starting to favor experiences purpose-built for them. They like to find information neatly organized in an API, a connector, an MCP. It’s cheaper for them (fewer tokens) and faster too. For Muse, the pre-built catalogs and checkout integrations make it favor those options. Dozens of new frameworks have popped up to address this issue. Stripe, OpenAI, Anthropic, Google, and about half of San Francisco companies are all in a race to figure it out. It’s a topic I’ll dive into in another article. If you read this far, thank you! My name is Gaut and I'm spending time with F500 / large brands to navigate growing and understand their AI and now agent traffic. If this is something you’d want to chat about, please don’t hesitate to reach out! https://x.com/i/article/2108955141386682368