I/O 2026 and What Google Wants the Cart to Become

Sundar Pichai opened Google I/O 2026 with the line "Welcome to the agentic Gemini era." Two decades watching technology pivots will give you a healthy reflex about phrases that are doing a lot of work in a small space, and this one earns its eyebrow. But read the keynote alongside the shopping post that went up the same afternoon, and the actual commerce story is more concrete than the framing. It has a shape, launch partners, and a UK launch date that is conspicuously absent.

The thing Google introduced at I/O 2026 is the Universal Cart, an intelligent shopping cart that lives across Search, the Gemini app, YouTube and Gmail. The cart is the new front door. Everything else in the keynote (Gemini 3.5 Flash, the new search box, information agents, generative UI inside Search, Spark as a 24/7 personal agent) exists in part because the cart needs surfaces to sit on. For UK retailers reading the coverage from outside the Shoreline bubble, the practical question this week is integrate or wait, and what happens either way.

Scale and shape of Google AI in 2026

Before the cart specifically, a sense of the scale Google is now operating at, because the scale is doing a lot of the argument by itself.

Google says it now processes more than 3.2 quadrillion tokens a month across its surfaces. Two years ago that figure was 9.7 trillion. One year ago, roughly 480 trillion. A seven-times year-on-year jump on a number that was already vast.

The user-side figures match the back-end. AI Overviews now reaches 2.5 billion monthly active users; AI Mode, which Google launched at I/O 2025, is at one billion monthly users a year later, with queries roughly doubling each quarter since launch. The Gemini app went from 400 million MAU last year to 900 million this year. These are Google's own numbers, presented at its own keynote, so calibrate. But the order of magnitude is the order of magnitude, and the implication is that "AI search" is now the same shape of surface as plain search was when we first learned to optimise for it.

The cart is the new front door

Last year's I/O announced AI Mode shopping features, agentic checkout via Google Pay, and personal-photo virtual try-on. All of those were Search-shaped: shopping smarts grafted onto a Search results page. The Universal Cart is a different artefact. It is an object you carry across Google's surfaces. You can add to it from a Search result, from a Gemini chat, from a YouTube creator's recommendation, from a Gmail confirmation that surfaces a back-in-stock item. It runs on Gemini models, sits on Google Wallet, and proactively does the work the shopper would otherwise do herself: price tracking, deal surfacing, compatibility checks. Launch partners include Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair and select Shopify brands.

Underneath it is the Universal Commerce Protocol that Google co-developed with Shopify earlier in the year. UCP-powered checkout is rolling out to Canada and Australia in the coming months. The UK is listed as "later." No date.

If you remember the original Buy on Google, or AMP's brief turn as the shopping shop window, the cadence will be familiar. Google has tried to be the front door before. What's different this time is that the surfaces it's promising to span are surfaces it actually owns and that consumers already use every day.

From the warehouse floor

Marcus Webb

The headline figure from I/O is the Shopping Graph's 60 billion product listings. The methodology beneath it is more interesting. To be a launch partner on Universal Cart in the way Vidhya Srinivasan described from the keynote stage, a retailer has to expose its inventory, its prices, its loyalty perks, and its checkout state to Google's surfaces in something close to real time. The vendor's commercial team will tell you that "real time" means within a few minutes. The vendor's product team, more quietly, will tell you what actually happens when a busy weekend hits the warehouse and the inventory sync lags forty minutes.

In 2017 I watched a mid-cap UK grocer commission a perfectly competent demand forecasting model. The model was right. The buying team overruled it on every Christmas SKU because, as the head of buying put it, that wasn't how it felt on the ground. Six months later the project was quietly retired and the head of buying was promoted. The pattern matters here because Universal Cart is a perfectly competent agent. The retailers it will be most painful for are the ones whose data layer cannot do the job the agent is asking of it. Their Merchant Center feed updates nightly, not by the minute. Pricing is set by a planning team who change it less often than the model assumes. The stock numbers are accurate to the warehouse but not to the picker on Tuesday afternoon who has just dropped a pallet.

None of this is a reason to refuse to integrate. It is a reason to be honest about which of those operational properties your business actually has. Pages 47 to 52 of any real warehouse audit will tell you a different story to the slide deck.

From the consumer side

Sarah Chen

The woman next to me on the BTS Sukhumvit line at 4pm last Friday was switching between three apps as she rode. Lazada for the kitchen stand she's been comparing for a week, LINE for the building's group chat about lift maintenance, Shopee Live for a clothes seller she watches but rarely buys from. Three apps, three accounts, three notification streams, none of which know about each other. The Universal Cart announcement is a pitch directly at her, even though she is not the announced market.

Whether she would let Google's agent shop on her behalf is a question with data behind it. A September 2025 Bain survey, cited by eMarketer, asked US shoppers who they trust for end-to-end agentic commerce: retailer-owned AI tools came in at 25 percent, Google and other big-tech companies at 16 percent, an AI platform like ChatGPT at 7 percent. We've covered the broader trust gap before; none of those numbers are encouraging if you are the platform.

Asia answers the question slightly differently. Alibaba's Qwen agent on Taobao has reached around 300 million monthly active users, per TNW's I/O coverage, by treating the shopping agent as a feature of the platform people already shop on, rather than as a destination. The lesson, which I have used since I got the Shopee Live story wrong in 2023, is that consumer behaviour in this region is poorly predicted by extending US precedents. It is also poorly predicted by extending Asian precedents back to the UK.

eMarketer is also forecasting, separately, that 95 percent of US AI-platform-driven ecommerce sales this year will still take place on retailers' own sites. The platform aggregates discovery while the shopper completes the purchase on the brand's own site.

What this means from a UK desk

From a Manchester engineering desk, the practical takeaway from I/O 2026 is narrower than the keynote framing suggests. The UK launch of UCP checkout is later than Canada and Australia, and Google did not commit to a date. By the time it arrives, retailers in the launch markets will have spent six to nine months exposing their feeds, pricing, and checkout state to Google's surfaces, and Google will have published a great deal about which patterns work and which do not.

The right preparation is to audit your product data, your inventory sync cadence, and your zero-click discovery posture, and ask whether they would survive being read by an agent on someone else's terms. Worth being honest about that, before the press release lands.