June 01, 2026

Google Ships the Cart, Amazon Rents Out the Agent, Microsoft Opens the Checkout

May 19 - June 1, 2026. The fortnight the agent got a cart, a wallet, and a price tag. Google gave the consumer one cart across Search, Gemini, YouTube, and Gmail. Amazon put its shopping agent on AWS for rent. Microsoft turned Copilot into a checkout.

Last week we covered the catalog becoming the ad, voice becoming checkout, and BNPL joining the agent. This fortnight the surfaces themselves moved. On May 19, Google I/O introduced Gemini 3.5 Flash, the Gemini Spark personal agent, and a Universal Cart that follows the shopper across every Google property. On May 20, Google Marketing Live turned that vision into merchant instructions, shipping Conversational Attributes in Merchant Center and expanding the Universal Commerce Protocol into new countries and new verticals. On May 20, Microsoft opened Copilot Checkout and Brand Agents. And on May 27, Amazon did the thing nobody expected: it put the engine behind Alexa for Shopping on AWS, so any retailer can build its own version of Amazon’s agent in about 60 days. The agent stopped being a feature the platforms own and started being a product the platforms sell.

Here’s what happened.

Google I/O: the agent gets one cart and a wallet

On May 19, at Google I/O 2026, Google introduced Universal Cart, a single shopping cart that works across merchants and across Google surfaces. A shopper can add an item to the same cart while searching, chatting with Gemini, watching YouTube, or reading Gmail. Once the item is in the cart, the system works in the background: tracking price drops, flagging when a product is back in stock, surfacing price history, and even catching compatibility problems (the example Google gave was conflicting PC components). Checkout completes through Google Pay or on the merchant’s own site, with the brand staying merchant of record. Universal Cart is rolling out in U.S. Search now and reaches the Gemini app this summer, with YouTube and Gmail to follow. The launch partner list is the signal: Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair, plus Shopify retailers including Fenty and Steve Madden.

The cart sits underneath the bigger I/O announcement. Per TechCrunch’s keynote coverage, the two headline products were Gemini 3.5 Flash and Gemini Spark, a persistent personal agent that runs on top of it. Spark runs on virtual machines through Google Cloud, operates 24/7, and works on long-running tasks in the background. The Agent Payments Protocol (AP2), which lets an agent make purchases under spending limits and brand or product restrictions the shopper sets in advance, is coming to Gemini Spark later this year. Google reported more than 1 billion daily shopping queries and a Shopping Graph of more than 60 billion product listings as the substrate all of this runs on.

Why it matters for merchants: For two years the agentic shopping story has been “the agent finds your product.” Universal Cart adds the part merchants underrate: the agent holds the basket. Once a consumer keeps one persistent cart that follows them from Gmail to YouTube to Search, the cart becomes the relationship, and the merchant who is not eligible to be added to that cart is not in the consideration set at the moment of purchase, no matter how good the product page is. The price-tracking and back-in-stock automation also quietly moves the abandoned-cart playbook off the merchant’s email list and onto Google’s surface, which means the re-engagement moment you used to own (the “still thinking about it?” email) is now Google’s to trigger. The concrete read: confirm your products are eligible to be added to Universal Cart through your platform or feed, and assume the cart, not the storefront, is where the next purchase decision gets made. The Spark and AP2 timeline (later this year) is the window to get the catalog ready before the autonomous-buying layer turns on.

Update: Google extended the same pattern into travel in August, adding a hotel checkout to AI Mode that completes in Google Pay while the booking partner stays merchant of record. The merchant-of-record choice described here is holding across categories, which makes it a deliberate architecture rather than a retail-specific concession.

Google Marketing Live: the schema that tells you which fields the agent reads

The day after I/O, on May 20, Google Marketing Live turned the consumer vision into merchant homework. Per Search Engine Land, Google launched Conversational Attributes in Merchant Center: a new set of product-data fields built for the way people ask AI assistants questions rather than the way they type keywords into a search box. The new attributes include Question and answer, Document link, Related product, Item group title, Variant option, and Popularity rank. They complement the existing feed and feed directly into AI Mode, Gemini, and Google’s Business Agent. Conversational Attributes is rolling out globally. Google also shipped AI Performance Insights, a share-of-voice style report showing how a retailer’s products surface across AI-powered shopping experiences, launching in a narrower set of markets (the U.S., Canada, Australia, New Zealand, and India).

Google paired the schema with a Universal Commerce Protocol expansion. Per Google’s Marketing Live shopping post, UCP-powered checkout is rolling out to Canada and Australia in the coming months and later to the U.K., UCP is coming to YouTube Shopping ads and Demand Gen campaigns in the U.S., and Google is bringing the protocol to entirely new categories starting with hotel booking and local food delivery. (We mapped the protocol split in our UCP vs ACP post and the council expansion in UCP doubles.)

New conversational attribute	What it answers for the agent
Question and answer	Sizing, fit, ingredients, use cases asked in natural language
Document link	Manuals, spec sheets, care guides the agent can cite
Related product	Compatible accessories and substitutes for cross-sell
Item group title	Groups variants so the agent reasons about the family, not the SKU
Variant option	Color, size, material the agent filters on
Popularity rank	Signals which variant to surface first

Why it matters for merchants: This is the most concrete catalog instruction Google has ever published. For months the advice has been “enrich your product data” without a spec sheet. Conversational Attributes is the spec sheet. Google is telling you, field by field, which data the agents read when they answer a shopper’s question, and the merchant who fills in Question and answer, Related product, and Document link is eligible to win conversational queries the merchant who ships a keyword-era feed cannot. AI Performance Insights closes the loop by making AI share-of-voice measurable for the first time, which means “are we showing up in AI Mode” stops being a guess, at least in the five markets where it launches. The European note is in the rollout map: UCP checkout is going to Canada, Australia, and the U.K., not the EU, and AI Performance Insights skips the EU too. European merchants get the Conversational Attributes schema (it is global) but neither the agentic checkout nor the measurement layer yet, so the move for EU sellers this quarter is to do the catalog work now and have it ready for when the rest crosses the Channel. (See our coverage of the 99% catalog-readiness gap for why most feeds are not ready for this.)

Amazon puts Alexa for Shopping on AWS for rent

On May 27, Amazon launched the AWS Agentic Shopping Assistant, a service that lets retailers build their own AI shopping assistant on the same technology that powers Alexa for Shopping (the assistant formerly named Rufus). Retailers tie the assistant to their own storefront, product catalog, customer data, and branding, and Amazon supplies the architecture, starter code, and learnings to get it live in as little as 60 days. Kate Spade, owned by Tapestry, is the first public customer and built a gifting assistant. Amazon says other retailers are already testing the service. More than 300 million customers used Rufus in 2025, and per GeekWire’s reporting, Amazon says the assistant drove nearly $12 billion in incremental sales last year.

The strategic move is the reversal. Two weeks ago Amazon retired the Rufus brand and put Alexa for Shopping in its own search bar, a consumer play. Now it is selling the same engine to the retailers it competes with, an infrastructure play. The company that spent 2025 blocking third-party agents and litigating against Perplexity is now both running its own agent and renting the technology to everyone else’s storefront.

Amazon’s agent, three postures	Move
On Amazon.com (consumer)	Alexa for Shopping in the main search bar (May 13)
Off Amazon, into the open web (consumer)	Buy for Me and Shop Direct, 400,000+ merchants
Sold to other retailers (infrastructure)	AWS Agentic Shopping Assistant, live in ~60 days

Why it matters for merchants: This is the build-versus-buy decision arriving early. A mid-market retailer that wanted an on-site shopping agent was, until May 27, looking at a custom build or a startup vendor. Now there is an AWS product with Amazon’s consumer-tested model behind it and a 60-day timeline. The upside is obvious: a proven agent on your own storefront, your data, your brand. The cost is strategic dependency, because the retailer that runs its storefront agent on Amazon’s stack is handing Amazon a read on its catalog, its customers, and its conversion data, the same way running ads on a competitor’s network does. The first question is not “should we build an agent” but “do we want that agent to be Amazon’s.” For merchants who already sell on Amazon, note the asymmetry: Amazon now monetizes you three ways at once, through marketplace fees, through Sponsored Prompts inside Alexa, and now through AWS if you license the assistant for your own site. Read the pricing before the 60-day clock starts.

Update: The following week, Amazon put a number on the pitch, saying conversational sessions on the assistant convert at 3.5 times the rate of keyword search. See the conversion figure in context.

Microsoft consolidates Copilot Checkout and Brand Agents

On May 20, Microsoft published its agentic commerce update, consolidating Brand Agents, Copilot Checkout, and its AI-visibility tooling into one resource for merchants. Copilot Checkout, which Microsoft launched as a channel earlier this year, lets a shopper describe a need, get a recommendation, choose a product, and complete the purchase inside the conversation, collapsing the funnel into a single thread. Brand Agents are AI-powered brand representatives that merchants can run in pilots for on-site and search engagement. Microsoft’s framing, from Roger Dunn, was blunt: “your product data is your shelf placement.” The post points merchants at structured, machine-readable product data, real-time inventory and pricing sync, and clean feeds to both the Microsoft and Google Merchant Centers, and it notes Microsoft works across both ACP and UCP rather than betting on one.

Why it matters for merchants: Microsoft sits alongside ChatGPT, Amazon, and Google as a surface where the transaction completes inside the conversation (Copilot Checkout shipped back in January, ahead of Google’s cart), and the pattern across all of them is now unmistakable: every assistant is becoming a place where the purchase closes, not just where the research happens. The merchant takeaway is the one Dunn stated. Across all four surfaces, the determinant of whether you show up is the same structured feed, and the fact that Microsoft explicitly accepts both ACP and UCP means the merchant who builds clean, protocol-agnostic product data once gets Copilot, ChatGPT, and Google off the same work. This May update is the productized, merchant-facing version of the checkout channel. The action is not Microsoft-specific. It is to stop maintaining one feed per channel and start maintaining one enriched, machine-readable catalog that every assistant can read.

Where the AI commerce stack stands this fortnight

Layer	Update this fortnight
Consumer agent + cart	Google ships Universal Cart and Gemini Spark across Search, Gemini, YouTube, Gmail
Merchant catalog schema	Google launches Conversational Attributes and AI Performance Insights in Merchant Center
Protocol geography	UCP checkout expands to Canada and Australia, U.K. later; new hotel and food verticals
Agent-as-a-product	Amazon puts Alexa for Shopping engine on AWS, retailers build in ~60 days
Conversational checkout	Microsoft opens Copilot Checkout and Brand Agents, ACP and UCP both supported

What merchants should do this fortnight

1. Fill in the Conversational Attributes Google just published. This is the most concrete catalog instruction any platform has shipped. Question and answer, Related product, Document link, Item group title, Variant option, Popularity rank are now fields in Merchant Center, and they are the data the agents read when a shopper asks a question instead of typing a keyword. Start with your top SKUs, write real answers to the real questions customers ask, and check AI Performance Insights to see whether your products surface in AI Mode and Gemini.

2. Confirm you are eligible for Universal Cart. When the cart follows the shopper across Search, Gemini, YouTube, and Gmail this summer, being addable to it is the new shelf. Check with your platform or feed provider whether your catalog is eligible, and treat the cart, not your storefront, as the place the purchase decision now happens.

3. Make the build-versus-buy call on an on-site agent deliberately. Amazon’s AWS Agentic Shopping Assistant makes a storefront agent a 60-day project instead of a custom build. Before you start the clock, decide whether you want that agent to run on Amazon’s stack, with the catalog and conversion visibility that hands Amazon, or on a more neutral one. The convenience is real; so is the dependency.

4. Maintain one catalog, not one feed per channel. Google, Microsoft, Amazon, and OpenAI are now all reading structured product data, and Microsoft accepts both ACP and UCP. The merchant who keeps a single enriched, machine-readable catalog (complete attributes, real-time price and inventory, conversational answers) gets every surface off one body of work. The merchant who maintains a separate feed per channel pays four times and falls behind on all of them. For European merchants, do the catalog work now even though UCP checkout is still routing to Canada, Australia, and the U.K. before the EU. The data is the part you control, and it is ready-to-ship the day the checkout layer arrives. (See our coverage of Europe’s regulatory lag for the timeline.)