Google’s own framing of this shopping push is the part I keep rereading: the company has described its agent work as an open standard meant to let AI agents interact with retailers across the shopping journey, checkout included. Not “next to” checkout. Not “up to” checkout. Including it.
And yet, in the test Google just started running in India, buying a product through Gemini or AI Mode in Search drops you into Flipkart’s checkout. Not a Google-hosted one. That single design decision tells you more about the state of agentic commerce than any keynote slide.
What is actually shipping
The facts are narrow, so let me keep them narrow. Google is testing purchases from Walmart-owned Flipkart directly through Gemini and through AI Mode in Search, in India. The test covers select products and select users. Checkout happens on Flipkart’s system. A wider rollout is planned for later in October.
That’s it. No published conversion numbers, no order volumes, no named pilot merchants beyond Flipkart itself. Everything else floating around right now is inference, mine included, and I’ll flag it when I’m inferring.
Why the checkout handoff is the interesting bit
If you’ve built a bot that tries to complete a transaction on someone else’s behalf, you already know where the pain lives. It isn’t intent parsing. It isn’t product matching. It’s the last twenty percent: cart state, tax and shipping calculation, payment tokens, fraud checks, inventory locks, returns policy, and whatever regional payment rules apply. In India that last one is not a footnote. Payment flows there carry real regulatory and authentication structure.
Owning that stack as a third party means owning liability you don’t want and reconciliation you can’t win. Handing off to Flipkart’s checkout means Google gets to be the discovery and decision layer while the retailer keeps the parts that are legally and operationally expensive.
For those of us building smaller agents, this is a permission slip. You do not need to rebuild commerce to build a useful shopping agent. You need to be excellent at the handoff.
Five things the handoff has to get right
Here’s the architecture checklist I’d apply to any agent that ends in “and then the user buys it.” These are my recommendations, drawn from the general shape of this pattern rather than from anything Google has published about its implementation.
- Deterministic product identity. Your agent’s idea of “the blue 512GB one” has to resolve to a single retailer SKU before you hand off. Fuzzy matching that lands the user on a near-miss product page is worse than no agent at all, because they’ll blame you and buy it anyway.
- State that survives the jump. Cart contents, quantity, variant, and any applied offer need to arrive intact on the retailer side. If the handoff loses the variant, the user re-picks it, and your agent just added a step instead of removing one.
- Idempotency at the boundary. Users retap. Networks drop. Two identical purchase intents fired a second apart must not become two orders. Assume this will happen during your first week of real traffic.
- Honest failure modes. Out of stock, price changed, offer expired, delivery unavailable at that pincode. These aren’t edge cases in commerce, they’re Tuesday. An agent that says “I found it for â‚ąX” and then hands off to a page showing a different number has broken trust permanently.
- A clean audit trail. Log what the agent recommended, what it resolved that to, and what it handed off. When a support ticket arrives, “the model decided” is not an answer anyone accepts.
The structural question nobody has answered yet
Retailers have spent two decades optimizing their own product pages, upsells, bundles, and loyalty prompts. An AI layer that answers the question and then routes straight to checkout bypasses most of that. Flipkart keeps the transaction and the customer data. It gives up a chunk of merchandising real estate.
Whether that trade holds up depends on volume, and we have no volume data. My read is that retailers will accept the trade while AI-driven traffic is small and re-examine it if it gets large. That’s speculation, clearly labeled.
What I’d build this week
If you’re shipping bots, the practical takeaway is to treat “intent to purchase” as a first-class, structured output of your agent, separate from your conversational layer. Emit a clean object: resolved SKU, quantity, variant, expected price, retailer, timestamp. Then let a thin adapter turn that into whatever the retailer accepts, a deep link today, a proper agent-facing API when one exists.
Build the adapter as the swappable part. The standards for how agents talk to retailers are still forming, and the teams who separated intent from transport will port over in an afternoon. Everyone else will be rewriting their checkout logic in October.
đź•’ Published: