I Never Saw the Website. I Trusted the Agent.

The refrigerator#

Last week my refrigerator's ice maker started throwing an IF E error. I spent a little over an hour taking the freezer apart with ChatGPT walking me through it, until I could feel the drag in one of the fan motors with my own hands. A year ago I would have hired a repair tech, because I wouldn't have known which part was bad or how to get to it.

Once we knew the part, I dictated one message to Muse, meaning I spoke to my computer and let voice-to-text type it out. It covered the fridge model, the error code, what I found, that I wanted a genuine LG part, and that I wanted it fast. Muse confirmed the part number fit my model on three parts sites, RepairClinic, Encompass and Parts Dr. Then it compared the sellers and came back with options. I replied "Parts Dr, fastest," told it which card to use, and the part shipped that afternoon.

I have never been to the Parts Dr website. I couldn't tell you what it looks like. I trusted the agent, and that is a bigger shift for brands than most people selling on Amazon have priced in.

It worked like a personal shopper, not a search box#

The fridge was the simplest case: one detailed message and a quick pick. The other purchases were conversations.

When the scroll wheel on my mouse broke, I told Muse I wanted a precise mouse for business work, that I'd take wireless only if there was no noticeable lag, and that I wasn't after a gaming mouse. It came back with four options: the Logitech MX Master 3S, the MX Anywhere 3S, the Razer Pro Click V2 and a wired Logitech. I asked for pros and cons, then for a recommendation. It picked the MX Master 3S for two reasons. Its scroll wheel fixed the exact thing that had broken, and Logitech's Mac software is the mature way to map a button to the dictation app I use all day, while Razer's Mac software is still a preview.

No keyword search would have connected my mouse to my dictation app. The agent did, because it knew I dictate all day, and the product with the better software support won the sale.

The other orders went the same way. For a Thunderbolt hub I gave it my port requirements, at least two USB-A ports and as many USB-C as possible, and it found the CalDigit hub that matched and confirmed it was in stock on CalDigit's own store. For a portable SSD it found that Samsung's own store was out of stock, bought it on sale at Best Buy, and walked me through the delivery options so the drive arrived after I got back from a trip. For a keyboard it held the checkout while I traveled, restarted it when I was home, and when the first checkout attempt failed, retried without creating a duplicate charge.

That is how I'd want a good assistant to handle a considered purchase: understand what I care about and narrow the field, then explain the trade-offs and handle the logistics.

Five orders, and where they went#

Over the past month, Muse handled five purchases for me:

What I bought	Where Muse bought it	Total
Logitech keyboard	Logitech.com	$107.99
Samsung portable SSD	Best Buy	$292.25
CalDigit Thunderbolt hub	CalDigit's own store	$249.99
LG refrigerator fan assembly	Parts Dr	$279.42
Logitech mouse	Best Buy	$86.59

That comes to $1,016, about $203 per order. My Amazon orders are mostly small replenishment buys, so this is a different kind of spending: considered purchases where I wanted help deciding.

Amazon wasn't allowed to compete for any of them. Amazon has blocked Muse since September 20. Its message reads: "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed" (Forbes). Amazon's stated reasons are that Meta gave no notice, that Muse doesn't identify itself as an agent, and that it appears to store customers' login details (GeekWire).

My average lines up with what Meta is seeing. On last week's earnings call, Mark Zuckerberg said: "The most common use case for Muse is shopping. With more than 25% of people who are using Muse engaging in this with an average purchase size of over $200" (transcript). He also said Meta plans to "get paid a fee based on the transactions that we help facilitate." Muse passed 10.2 million US downloads in its first 40 days, compared with 3.7 million US downloads for ChatGPT in its own first 40 days, and it has held No. 1 on the US App Store nearly every day since mid-September (TechCrunch).

A sixth item, where Amazon won#

There was a sixth item Muse didn't finish. It took a $30 Razer mousepad through checkout on Razer.com and stopped at my approval screen. Razer wanted $10 for shipping, so I declined, opened Amazon myself and bought the same pad without the shipping charge.

Muse was willing to send me to Razer with the $10 charge, and I was the one who compared it with Amazon. Shipping costs matter to shoppers: extra costs like shipping are the top reason people abandon a checkout, other than just browsing, cited by 40% of US online shoppers who abandoned one (Baymard Institute). Amazon's advantage here is delivery. It delivered more than 13 billion items same-day or next-day to Prime members worldwide in 2025 (Digital Commerce 360). Amazon got this sale only because I checked it myself. Muse couldn't put Amazon in front of me.

Why I stopped caring which store it was#

For years Amazon has been where people start: in 2024, 56% of US online shoppers began product searches on Amazon, compared with 42% on search engines (Jungle Scout via Practical Ecommerce). Fast, free delivery kept them there, and brands learned to rank for short keyword searches.

An agent changes what I need from the store. I describe what I want in full sentences, and the agent does the narrowing. Amazon knows this is where search is going: since May, the search bar in its app recognizes questions and hands them to Alexa for Shopping (About Amazon).

The payment step is where my trust moved. I approve one total on one screen before any money moves. In my orders Muse paid through Stripe Link (Shopify stores can also use Shop Pay), and at stores that don't take Link, Stripe issues "a single-use virtual card scoped to the approved purchase" (Stripe). Purchases through Link also carry Link's purchase protections, which Meta says cover damaged or lost items, price drops and no-fee returns on eligible purchases (Meta). Between the approval step and that protection, the store's own reputation mattered much less to me. That's how I ended up buying from a site I've never seen.

How Muse found stores and narrowed the options#

Muse shops two ways. Meta gives it a catalog through its partners, including Walmart, Best Buy, Gap, Sephora, Wayfair and the whole Shopify catalog (TechCrunch). It also runs its own browser on a cloud computer and visits sites directly, the way a person would (Meta).

Most of the shortlist comes from the browser. Profound, an AI search analytics company, tested Muse across 50 shopping prompts in one retail category and found that in a typical six-product shortlist, four picks came from live browsing and two from Meta's catalog (PPC Land). Profound also reported that Muse's instructions tell it to favor well-known merchants, prefer direct sellers over resellers, and put a brand's own store first when the shopper names the brand (PPC Land). Meta hasn't confirmed those instructions, but my orders fit them: the keyboard and the hub both came from the brand's own store, Logitech.com and CalDigit's, and when Samsung's own store was out of stock Muse went to Best Buy.

Parts Dr shows the browsing route. It's an appliance parts retailer in Sioux Falls, South Dakota, founded in 2010. It isn't a Shopify store and isn't a Meta partner. Its product page for my part tells a machine exactly what it is. The page code carries structured product data with the brand and manufacturer part number, the price and stock status, plus the customer reviews and ratings. The page also lists which fridge models the part fits, and the site promises same-day shipping on in-stock orders placed before 3 p.m. Central. That's everything an agent needed to confirm the fit and compare the offer.

Parts Dr also runs Cloudflare bot protection. A plain automated request to its site gets a "Quick Security Check" page. Muse got through anyway. Security firm Cequence found that Muse generally looks like an ordinary Chrome browser and doesn't declare itself an agent (Cyber Insider), which is the same behavior Amazon objects to. Meta, Sierra, Walmart, Shopify and Stripe are working on a standard for agents to identify themselves, the Personal Agent Protocol, but it's still a draft (Quartz; The Next Web).

How agents find your products, off Amazon and on it#

So the practical question for a brand is how an agent finds your product, understands why it fits, and gets through checkout. The answer is different on your own site and on Amazon.

Off Amazon#

Be in the catalogs agents read. The shortest route is Shopify. Its Agentic Storefronts give merchants "out-of-the-box access to major AI channels" from the Shopify admin, covering ChatGPT, Microsoft Copilot, AI Mode in Google Search and the Gemini app, through Shopify Catalog (Shopify). In September Meta was added as a channel, with products shared by default (Shopify Changelog). Brands that aren't on Shopify can join Shopify Catalog through its Agentic plan.

The AI platforms run their own programs too. OpenAI publishes a product feed spec for ChatGPT (title, description, price, availability and a product URL), currently open to approved partners (OpenAI spec; OpenAI onboarding). Google is opening checkout inside AI Mode and Gemini to select merchants in the US, Canada and Australia through an early-access program in Merchant Center (Google). Last week Walmart announced it is building an Agent Gateway to give agents a direct connection to its store, alongside its existing integrations with Muse, ChatGPT and Gemini (Digital Commerce 360). Meta has no merchant sign-up for Muse shopping beyond its developer program for connectors, so Shopify Catalog is the documented way in.

Make your own pages readable to the browsing agents. Put structured product data on every product page: brand, part or model numbers, price, stock, and your shipping and return details. Google says that markup, together with a Merchant Center feed, helps it "correctly understand and verify your data" (Google). Publish the facts an agent needs to argue for you, like compatibility, specs, software support and delivery dates. That's what won my mouse and my fridge part.

Check that an agent can actually get through. Bot protection turns agents away, sometimes on purpose and often by accident. Walmart, a Muse partner, had Muse checkouts failing because of its human-verification step (TechCrunch), and since September 15, Cloudflare has blocked agent traffic by default on ad-supported pages of newly added sites (Cloudflare). Decide which agents you want to let in, then test a real purchase through one to make sure guest checkout works.

On Amazon#

On Amazon, the agent that matters is Alexa for Shopping, which is what Rufus became when Amazon merged it with Alexa+ in May. Amazon says it is trained on "Amazon's extensive product catalog, customer reviews, community Q&As, and information from across the web" (About Amazon), so that's where the effort goes.

Fill in every attribute in the structured fields, especially compatibility and sizing. An Amazon staff member put it plainly in the Seller Central forums: "the more comprehensive your compatibility information is in the structured fields, the better" (Seller Forums).

Write bullets that state facts. An Amazon staff member's bullet guidance says to avoid subjective claims (Seller Forums), and in the listings we run, bullets that say what's in the box and the situations it's used in hold up far better than adjective bullets like "Premium Quality You'll Love." An agent weighing trade-offs, like Muse did with my mouse, needs facts it can compare.

Treat reviews and customer questions as inputs to the agent. If shoppers keep asking the same question, answer it in the listing itself.

Know that your ads are already in the agent. Amazon Ads says "existing sponsored ads campaigns are automatically eligible to serve in Alexa for Shopping," including as sponsored prompts (Amazon Ads). Amazon doesn't appear to report those placements separately, so ask whoever runs your ads what your reports show.

Decide on Buy for Me on purpose. Buy for Me is a feature inside Alexa for Shopping that buys from brand websites for Amazon shoppers. Amazon reads product information from brand sites by default, and its brand FAQ says to email shopdirect@amazon.com to opt out (Amazon). If you stay in, Amazon now accepts product feeds, which give you more control than letting it read your site (About Amazon).

How I think Amazon responds over the next 12 to 18 months#

These are my predictions, with the evidence I'm leaning on.

1. Amazon builds up its own agent instead of letting Meta's in. Amazon says Alexa for Shopping, counting its time as Rufus, reached more than 350 million customers in 12 months (About Amazon), and Andy Jassy has said "a lot of customers are ultimately going to choose to use a great shopping agent from that retailer" (Retail Brew). Its agent also shops beyond Amazon: the Buy for Me feature inside Alexa for Shopping can buy tens of millions of products from other brands' websites without the shopper leaving the Amazon app (About Amazon). Whoever owns the agent owns the customer relationship and the ad space, and Amazon won't hand that to Meta.

2. If Amazon lets outside agents in, it will set the rules. Jassy said in October 2025 that Amazon expects "over time, to partner with third-party agents" (Modern Retail). I expect any deal to require agents that identify themselves, shoppers signed into Amazon, and checkout through Amazon. Amazon already requires automated tools that connect to seller accounts to "clearly identify themselves as automated systems" (Amazon Seller Forums). When it blocked Muse, Amazon said agents that buy for customers "should operate openly and respect service provider decisions about whether or not to participate" (Retail Dive). Critics point out that Buy for Me reads brand websites and buys from them unless the brand opts out, the opposite of that standard (TechTimes). Amazon's side is that Buy for Me identifies itself and offers an opt-out, which Muse doesn't. I expect Amazon to hold both positions: open when its own agent is shopping other sites, closed when someone else's agent shops Amazon. In August an appeals court also lifted an order keeping Perplexity's agent off Amazon, calling the agent "a tool, not a person" under federal anti-hacking law (Ninth Circuit). That makes a hacking-law case against agents harder to win, though Amazon's terms of service still apply.

3. Who owns the checkout matters more than which AI model is underneath. Amazon has invested $18 billion in Anthropic, the company behind Claude, according to its second-quarter 2026 filing, and Anthropic says Claude helps power Alexa+ (Amazon 10-Q; Anthropic). The same filing shows a $50 billion commitment to OpenAI this year. If Claude or ChatGPT can one day buy from Amazon, I expect it to run through Amazon's own checkout and ad system, because that's where the money is.

4. Paid placement inside the agent becomes a major ad channel. It's already started with the sponsored ads in Alexa for Shopping. I expect more of a brand's ad budget to follow, and Meta's plan to charge a fee on sales points the same direction off Amazon.

5. Agents start comparing delivered cost across stores. Muse timed my SSD delivery around a trip, but it went ahead with Razer's $10 shipping on my mousepad and I did the price comparison myself. As agents get better at comparing the full delivered cost and arrival date, I expect that to become part of how they choose. If shipping makes you lose that comparison, the shopper may never see why.

What I'm watching this holiday season#

Adobe expects AI-referred traffic to US retail sites to be up 130% for the season and 159% on Thanksgiving Day itself (Digital Commerce 360). Visits referred by AI already converted 60% better than other traffic in July, with 53% more revenue per visit (Digital Commerce 360), from what Adobe calls a modest base of users (Adobe). If you don't track AI referrals as their own channel in analytics yet, set that up before the holidays so you have a baseline.

Back in January we wrote When AI Becomes the Buyer about agents starting to make purchase decisions. We'll report what we see across the brands we work with this season, and whether Amazon makes a move. If you want that when it's out, subscribe below. I send one email when a new piece goes up.

---

Texin.ai runs Amazon for brands and helps them show up wherever their buyers search, including AI search and agents. Talk to us.