For twenty years online shopping has meant a person browsing, comparing and clicking "buy". In 2026 a second pattern is appearing: the person tells an AI assistant what they need, and the assistant finds, compares and increasingly buys it. OpenAI's Agentic Commerce Protocol, built with Stripe, lets merchants sell inside ChatGPT. Google's Agent Payments Protocol and the card networks' agent frameworks define how an agent proves it has the user's permission to pay. The term for all of this is agentic commerce.

Adoption is early, more so in India, where UPI and cash-on-delivery shape everything. But the preparation work overlaps almost completely with good e-commerce practice, so there is no reason to wait. This is a practical list.

How an AI shopping agent actually finds products

An agent does three things a human does not. It reads your product data as structured text rather than looking at the page. It compares your offer against dozens of competitors in seconds. And it will only recommend what it can verify: price, availability, delivery time, returns. A store with beautiful photography and vague specifications is invisible to it.

Six things to fix, in order

1. Product structured data on every product page

Schema.org Product and Offer markup with price, currency, availability, GTIN or MPN where you have one, shipping details and return policy. This is the same markup Google Shopping uses, so it pays twice. Most Indian stores we audit either lack it or have a plugin emitting incorrect prices.

2. A clean, complete product feed

Google Merchant Center and the emerging agent protocols consume feeds, not pages. Every product needs a unique identifier, a precise title, a factual description, real stock status and consistent categories. If your catalogue has 800 products and 300 of them say "in stock" because nobody updated them, agents will learn not to trust you.

3. Machine-readable policies

Returns, refunds, delivery areas, delivery times, warranty. An agent asked "which store delivers to Thrissur by Friday with free returns" needs those facts in structured form. Write the policy pages in plain language, mark them up, and make sure llms.txt points to them.

4. A checkout that can be driven by an API

The agent protocols work by calling a merchant endpoint to create a cart, quote shipping and complete the order with a delegated payment token. A custom Laravel store exposes this with a few routes. A heavily customised WordPress WooCommerce site with checkout plugins layered on top often cannot. This is the step that separates stores built as platforms from stores built as brochures.

5. Payment tokens, not stored cards

Agent payments rely on tokenised credentials with scope and limits, which is also where RBI's card tokenisation rules already pushed Indian merchants. Razorpay, PayU and the international gateways are adding agent-aware flows; make sure your integration is on a current SDK version so you can switch them on.

6. Reviews and proof that machines can read

Aggregate rating markup, verified-purchase review feeds and a Google Business Profile with consistent details. Agents weigh trust signals explicitly. We covered the mechanics in our GEO and AEO guide.

What to ignore for now

  • Building your own shopping agent. Let the platforms do it; be the merchant they find.
  • Paying for "AI visibility" tools that cannot show you a referral. Measure first, as described in our AI search tracking guide.
  • Rewriting product copy for robots. Factual, specific copy serves humans and agents equally.

The India angle

Two things are different here. ONDC already forces sellers to publish structured catalogues and accept orders from any buyer app, which is agentic commerce in everything but name; sellers on ONDC are ahead. And UPI's growing credit and mandate features give agents a payment rail that does not depend on cards. A store that handles both will be ready whichever protocol wins. Our earlier post on UPI and ONDC covers the groundwork.

Frequently asked questions

Can customers buy from my store inside ChatGPT today?

Only if your store is on a platform that has integrated the Agentic Commerce Protocol, which in late 2026 is still a short list weighted towards the US. The preparation steps above are what get you onto that list when it opens.

Will this hurt my own website traffic?

Some discovery will move into assistants. The order and the customer are still yours if your data is the one the agent trusts. Stores that are not machine-readable lose the sale to one that is.

Does this apply to a B2B or service business?

Yes. Agents already compare service providers on price, location, availability and reviews. The structured-data and policy steps apply directly; the checkout step becomes an enquiry or booking API.

We build e-commerce platforms with these capabilities from the start. See retail and e-commerce solutions or request an audit of your existing store.