Alexa, who's winning agentic commerce?

Welcome to Product Cocktail, where the takes are as polarizing as a shot of Fernet—but the insights come together like a perfectly crafted daiquiri.

Rufus (now Alexa for Shopping) has been used by 350 million customers in the last 12 months (as of August 2026). Alexa for Shopping customers are 60%+ more likely to convert than those who don't use it.

Meanwhile, ChatGPT's Instant Checkout—a product I spent six LinkedIn posts glazing up last year—is dead.

Across two posts, I analyzed Instant Checkout from a customer and merchant perspective, pointing out several issues that gave me pause about OpenAI's offering.

The seam: a bolted-on checkout experience and post-purchase black hole.

OpenAI admitted their own failures here when they did an about-face on Instant Checkout in March, shifting to "allowing merchants to use their own checkout experiences while we focus our efforts on product discovery." Turns out that eCommerce is hard. (No shit.)

The trust signal gap.

A Semrush survey conducted in December 2025 indicates that while many customers use AI to research products (77% at least once a month), significantly less (41%) have purchased after using AI during their research.

To wit, after Walmart made 200,000 products available directly in ChatGPT, they observed 3x lower conversion for Instant Checkout eligible items compared to ones that require clicking out to Walmart.com.

The evolution of product positioning.

ChatGPT didn't understand your product. Per Forrester analyst Emily Pfeiffer, "OpenAI could scrape some retailers' websites to obtain data on products surfaced in ChatGPT... crawling and scraping is inadequate to get the full breadth of product data that you need to do a good job of commerce."

Meanwhile, days before OpenAI dropped their retread-to-product-discovery PR, Google shipped Universal Commerce Protocol (competitor to OpenAI/Stripe's Agentic Commerce Protocol) updates that load real-time product data to prevent out-of-stocks and pricing errors.

I concluded the original series with optimism couched in realism:

The problem: I correctly identified the gaps in the initial product, but I had too much optimism that they could be resolved by OpenAI.

Demand
I read the capability and assumed the appetite. I focused too much on the capability problem that OpenAI and Stripe were solving, while ignoring the customer demand problem.

It all sounded very impressive: 700M weekly ChatGPT users being connected to 2.2M Shopify and Etsy merchants and 500M Walmart SKUs. I misinterpreted the 50M ChatGPT shopping queries as willingness-to-buy, as opposed to willingness-to-ask.

In fairness, the payments aspect was handled expertly by Stripe's crack team of PMs and engineers: payment rails, Stripe Link and its 200M users, Agentic Commerce Protocol (ACP). Beautiful, shiny, and useless without customer adoption. Chalk that up to drinking too much Kool-Aid at Stripe Sessions, I guess.

Supply

I flagged the limited catalog from a place of optimism, not skepticism.

OpenAI promised millions of merchants on Shopify alone. When the floodgates opened in late January 2026, thirty trickled through. (According to Forrester. OpenAI and Shopify never disclosed a number.) There was never an order of magnitude or GMV boost, and the product died eight weeks later.

Turns out that "OpenAI underestimated how difficult the enablement of transactions was going to be" (Gartner).

This is the same mistake, made on both sides of the marketplace. Buyers were counted in users, wallets, and query volume. Sellers were counted in eligibility. Just because a thousand people walk into your store every day doesn't mean you're going to sell anything.

Instant Checkout launched in late September. By March, OpenAI's agentic commerce product team was straight up not having a good time:

March 4: Daniel Danker—Walmart's EVP of AI, Product, and Design—signaled the end of Instant Checkout at a Morgan Stanley conference. Same day, The Information drops a scoop "OpenAI Scales Back Shopping Plans for ChatGPT."

March 18: Walmart announces a move to put their own agentic commerce bot, Sparky, inside of ChatGPT and Gemini.

March 24: OpenAI announces a pivot to focus on product discovery.

They didn't exit commerce, they handed checkout back to the merchants. ACP survived as "the infrastructure that connects users to merchants across the full shopping journey." I'm sure Stripe is thrilled to be building a glorified Product Experience Management (PXM) protocol.

The speed at which they made this decision can likely be credited to the competitive pressure they faced at the time (and their shitshow of a product strategy), but the lack of initial adoption signals (on either the demand or supply side) undoubtedly made that call easier.

In late October, I called Walmart's partnership with OpenAI the "first major retailer validating the standard" and used Shopify and Etsy as proof this was an ecosystem rather than a glorified tech demo.

Turns out everyone was renting-to-own.

Walmart built Sparky and moved it into ChatGPT, bringing Walmart authentication, baskets, and personalization with it. As Danker put it, "when Sparky travels, it's the Walmart store meeting you where you are, instead of a completely broken experience." Daaaamn, Danker.

Shopify pulled the Instant Checkout support, shifting to an outbound linking strategy, (likely to avoid OpenAI's fees and broken checkout experience).

Etsy still appears to be playing nice. During the Instant Checkout kerfuffle in March, they indicated they were launching a ChatGPT app to regain more control over the look and feel and give their merchants access to shopper data earlier in the journey. It launched as a beta in May.

I misread who held the power in these partnerships. Each of these retailers already had distribution and the customer relationships, ChatGPT was just another channel—and one that turned out not to justify the cost. They just needed the ramp for their tech teams to deliver something that worked for their customers.

Amazon made this calculation earlier, betting on Rufus and starting to block LLM scrapers well before OpenAI announced Instant Checkout.

Now that I've really established the self-own by republishing that line twice, I'll do what I should have done in the first place: apply the same standard of analysis to Amazon as I did to OpenAI at the time.

To be clear, this holds where Amazon is the store. They owned the conditions that made it work on its own turf.

When I wrote the Instant Checkout analysis, I should have applied a more careful lens to the 'zon. It was an oversimplification to think that the Everything Store couldn't be everything, everywhere, all at once... including an agentic commerce player.

The primary reason that I didn't give Amazon its due is because I hate Alexa. It's not a good product by 2025+ AI standards (to be fair, as I learned after sinking hundreds of dollars into new devices, Google Home isn't either). Sure, it gained a lot of consumer traction, but it never became more than a glorified kitchen timer/annoying weather-person and it was an abject commercial failure. In short, I don't think it puts Amazon in the best light as a real consumer AI player.

But Amazon has a ridiculously large Product and Engineering organization, and their product search algorithm chops are unparalleled. The (apparent) inability to build voice doesn't translate into an inability to turn ecommerce data into agentic commerce. Interface competence and commerce capability are different assets.

Structurally, I applied an architectural analysis for the challenger, while applying reputational heuristics for the incumbent. This was a lazy approach and I'll do better next time. (Alexa still blows.)

So... where does that leave Agentic Commerce, broadly?

In my quest to perform ritual sacrifice on myself through this newsletter edition, I came across some fascinating data from Walmart's Danker I'll leave you with. (Aside from his extremely savage quote above that undoubtedly keeps the hapless OpenAI PM-Engineering team who built Instant Checkout up at night.)

Walmart's top-selling categories on Instant Checkout? Vitamins and protein supplements—particularly customers new to GLP-1s that are receiving advice to increase nutrient intake. Automotive, beauty, home management, hardware, and tools account for over half of Instant Checkout Orders. (Notably, certain products like TVs were omitted from the set of eligible SKUs.)

To me, that list reads as conversational research landing on a single considered purchase. Not a grocery list. Not a school shopping list. A question with with a product as the answer.

That's what OpenAI kept when they killed Instant Checkout: product discovery. A comparison of images, pricing, features and reviews to point customers to a product.

That's the same motion that Alexa Shopping is built on. They already owned the transaction, they just moved upstream to own the conversation too.

Amazon claims nearly $12 billion in incremental annualized sales from owning that conversation.

The second half of my Amazon shrug off was around their true agentic commerce play: Buy for Me. Their Shop Direct program exposes 100M+ products from 400K merchants off site, with tens of millions eligible for "Buy For Me" in which Amazon completes the purchase on their behalf. It's a potentially risky bet: The failure modes that impacted Instant Checkout likely seep into the "Amazon Buy for Me" or "Shop Direct" experiences.

This suggests to me that Amazon is really concerned with having the definitive right answer in that product discovery conversation, even if it means linking out to a competitor.

I ended my "Market Response" piece in October with a provocative, slightly inside baseball line:

Amazon did all three: they blocked the crawlers, built the assistant, then invested up to $50B in OpenAI.

OpenAI had to build a store around their assistant—catalog, trust, payments, fulfillment, returns, and the purchase intent to drive conversion. Amazon only had to build the assistant.

Send me questions, feedback, and cocktail recipes:
[email protected].

_{Icons made by Icongeek26 from www.flaticon.com.}