Gemini Learns to Shop, But Only From One Store So Far

Google wants you to believe your AI assistant can buy things for you. Right now, in India, it can buy a handful of things, for a handful of people, from exactly one retailer. Both of those statements are true at the same time, and the gap between them is the whole story.

The news itself is small on paper. Google is testing Flipkart shopping inside Gemini and AI Mode in Search in India. You ask about a product, and instead of bouncing you into a search results page and then a browser tab and then a login screen, the flow drops you into Flipkart-branded checkout. The test is limited to select products and select users, with a broader rollout planned for later in October.

That’s it. That’s the announcement. And I think it’s more interesting than the size of the test suggests.

Why one retailer matters more than ten

I review AI tools for a living, which mostly means watching demos that work beautifully once and then fall apart when a real person with real preferences touches them. Shopping agents have been the worst offenders in that category. The pitch has always been the same: tell the assistant what you want, it finds it, it buys it, you’re done. The reality has usually been a chatbot that reads product listings out loud and then hands you a link.

Checkout is where these things break. Not search. Not recommendations. Checkout. Payment methods, address selection, order confirmation, returns, the actual transfer of money from your account to someone else’s. That’s the part every agent demo quietly skips, and it’s the part Google is testing here.

Doing it with one retailer, in one country, with a subset of products, is the correct way to attempt this. A single integration means Google can find out what actually goes wrong when an AI hands a user off to a payment flow. Ten integrations at launch would mean ten different sets of failures arriving simultaneously, and no clear signal about which were Google’s fault and which were the retailer’s.

The standard underneath

The more consequential piece is the protocol layer. Google has framed this work as an open standard meant to let AI agents interact with retailers across the shopping journey, checkout included. Google has said the technology would let shoppers buy eligible items through this path.

That framing tells you what the Flipkart test actually is. It’s not a Flipkart partnership dressed up It’s a test of whether a general-purpose agent-to-retailer interface holds up in production, using Flipkart as the first real participant. Flipkart being Walmart-owned makes it a reasonable pick for that. It’s a large operation with the engineering resources to build against a new standard and the volume to stress-test it.

If the standard works, the second retailer is easier than the first, and the tenth is easier than the second. If it doesn’t, Google has one broken integration in one market instead of a mess spread across its shopping product.

What I’d watch before getting excited

A few things would tell me whether this graduates from a demo into something I’d actually recommend:

• Who owns the failure. When an order goes wrong, is that a Gemini support problem or a Flipkart support problem? Agentic commerce creates a new seam, and seams are where customer service goes to die.
• What “select products” means after October. If the expanded rollout is still limited to simple, low-variation items, that suggests the hard cases such as sizing, variants, and bundles aren’t solved.
• Whether the AI’s pick is the best pick. An assistant that surfaces one product and routes you to buy it has enormous influence over what you purchase. The incentive structure there deserves scrutiny.
• Retailer appetite. Merchants gave up a lot of control when marketplaces took over discovery. Handing the customer relationship to an AI layer is a bigger ask.

My read

India is a smart place to run this. High mobile commerce volume, a population comfortable with app-based payments, and a market where Google already has deep Search penetration. If the flow can survive there, it can probably survive most places.

But I’m not going to pretend a limited test with one partner is proof of anything yet. What it is, is the first version of this idea I’ve seen that starts with the boring, difficult part instead of the flashy part. Most agent shopping products start with conversation and never reach the transaction. This one started with the transaction.

That’s a better order of operations. Whether Google follows through past October is the actual question, and the answer will be visible in how many retailers show up next.

🕒 Published: September 28, 2026