Google’s own framing of this is the part I keep rereading: the company has described its agent shopping work as an open standard meant to let AI agents interact with retailers across the whole shopping journey, checkout included. Not browse. Not compare. Checkout. That last word is doing an enormous amount of work, and anyone who has built a bot that has to touch money knows exactly why.
Which brings us to India, where Google is testing shopping from Walmart-owned Flipkart directly inside Gemini and AI Mode in Search. The test is small: select products, select users, Flipkart-branded checkout appearing inside the flow, with a broader rollout planned for later in October. That’s the whole verified picture. No adoption numbers, no conversion stats, no named executive doing a victory lap. And honestly, the modest size of it is what makes it interesting to me as a builder.
Why checkout is the wall every bot hits
I’ve built plenty of assistants that can answer questions about products. That part is close to solved. Retrieval, ranking, a decent prompt, some structured product data, and you have something that feels smart. Then a user says “okay, buy it,” and the whole thing falls apart.
Here’s what actually breaks:
- Identity. The agent knows who the user is in your system. It does not know who they are in the retailer’s system. Linking those two accounts safely is a real design problem, not a config flag.
- Payment authority. Somebody has to be accountable for a charge. Agents are bad at accountability by default. You need an explicit, auditable moment where a human said yes to a specific amount.
- Inventory and price truth. The price your model saw thirty seconds ago may not be the price at the register. Stale state is a support ticket waiting to happen.
- Order lifecycle. Buying is the easy half. Cancellations, partial shipments, refunds, returns, and disputes all need a path back through the agent or they need a path out of it.
Look at the choice Google appears to be making in this test. Checkout is Flipkart-branded and the flow sends the user toward Flipkart’s own checkout. The agent does the discovery and the handoff. The merchant keeps the money, the trust boundary, and the legal relationship with the buyer. That’s a conservative architecture and it’s the right one to start with.
The handoff pattern is worth copying
If you’re designing an agent that touches commerce, this is the pattern I’d build first, regardless of what model or framework you’re using.
Keep the agent on the discovery side of the line
Let it search, filter, compare, and explain. Let it assemble intent: this product, this variant, this quantity. Then hand that intent to the merchant as a structured payload and let the merchant’s surface handle authentication and payment. Your agent’s job ends at “here is what the user wants, verified.”
Make the intent object explicit
Don’t let purchase intent live in a conversation transcript. Give it a shape: product identifier, variant, quantity, price the user was shown, timestamp, session. Then you can validate it, log it, replay it, and diff it against what the merchant returns. When something goes wrong, you have an artifact to argue with instead of a chat log to interpret.
Treat confirmation as a first-class step
One clear screen, one clear amount, one clear action. Not a model deciding it has enough signal to proceed. I’d rather lose a bit of conversational smoothness than debug an unwanted order.
What the open standard angle actually means for us
The interesting long-term piece isn’t that Gemini can buy a phone case. It’s that Google is positioning agent-to-retailer interaction as a standard rather than a private integration. If that holds, small builders eventually get to plug into the same rails as the big assistants instead of writing a bespoke scraper for every store. That’s the version of this I’m rooting for, because right now the alternative is a handful of platforms with private deals and everyone else screen-scraping.
I’d temper expectations though. A limited test with select products and users is a test, not a launch. We don’t have published results, and a broader rollout planned for later in October is a plan, not a guarantee. I’ve watched enough commerce integrations slip to hold that loosely.
What I’d do this week
If commerce is anywhere on your roadmap, stop building a buy button and start building a clean handoff. Define your purchase intent schema. Decide where your trust boundary sits and write it down. Instrument the moment of user confirmation like it’s the most important event in your system, because it is. Then when standardized agent checkout becomes broadly available, you’re swapping a transport layer instead of rebuilding your architecture.
The models got good at talking about products a while ago. The hard, unglamorous work is everything that happens after the user says yes, and that’s the part India’s test is quietly probing.
🕒 Published: