The case for this index
The buyers are becoming software.
AI assistants have started doing the shopping — finding products, filling carts, checking out. In the past year every major payments company shipped a protocol that lets an agent pay, and NPCI is reportedly building those rails into UPI itself. So the question is no longer whether agents will buy. It's whether your store will let them.
Every payment rail went agentic in one year
These aren't research papers — they are shipping protocols from the companies that already move most of the world's online money. Each one answers the same question: how does a merchant get paid when the buyer is an AI agent acting on a human's behalf?
Agentic Commerce Protocol, built with Stripe. Lets an agent complete a purchase without leaving ChatGPT: the merchant receives a normal order, the buyer never opens the store's website. Already live with US merchants.
Agent Payments Protocol — a common "mandate" format that records what the buyer authorised their agent to spend, so banks and merchants can trust a purchase no human clicked. Backed by dozens of payments companies.
Trusted Agent Protocol — gives merchants a standard way to tell a legitimate shopping agent apart from a hostile bot at the front door, instead of blocking both.
Unified Agent Protocol — AI agents verified and authorised inside UPI itself. The moment this ships, agent payments stop being a foreign concept and become part of the rails every Indian D2C store already runs on.
Agents are already shopping
People already ask ChatGPT, Claude and Perplexity to find and buy things. Until the new rails are universal, the agent visits your store the same way a person does — through the storefront, in a browser: it searches, opens a product page, picks a size, adds to cart, and starts checkout. That browser journey is exactly what this benchmark runs, 183 times so far across 50 Indian D2C storefronts, with a hard stop before payment.
This isn't a simulation of some future shopper. It is a measurement of the shopper that exists today — the one your analytics can't tell apart from a lost customer.
Payment is nearly solved. The storefront isn't.
The protocols above standardise the easy part: proving who the agent is and moving the money. Every one of them quietly assumes the agent already made it to checkout. But no protocol can pick a size, close a popup, survive a forced-login wall, or get past an OTP form. That part happens on your store — and it is exactly where agents fail today. The specific ways they fail are catalogued in the failure taxonomy: OTP walls, CAPTCHAs, bot-blocks, broken flows, popups that never close.
What the data says
of ranked stores stopped the agent before it ever saw a payment screen — only 5 of 43 let it through. The rails are coming; the stores are not ready. That gap is what this index measures.
Good markup doesn't fix this: stores with perfect structured data still scored zero on the live journey, and static readability explained less than 3% of what the agent actually achieved. The only way to know whether an agent can buy from a store is to send one — which is why the live journey is 70% of the score.
Fair questions
Won't stores just become agent-native, making this irrelevant?
Eventually, maybe. But agent-native storefronts arrive one integration at a time, and small D2C brands adopt last — so for stores like these, the browser journey is how agents will shop for years. More importantly, most of the walls this benchmark finds are choices, not accidents: OTP gates, forced logins, bot-blocks. A store that blocks agents today will be just as slow to welcome them on the new rails.
What happens to the index when that shift arrives?
It shifts too — the scoring rubric is versioned. Version 1, this leaderboard, measures what is true in 2026: agents driving the same storefront humans use. Later versions will score the new surfaces instead — does the store publish product feeds and agent endpoints, and can an agent complete a purchase through them? The question underneath never changes: is this store ready for a buyer that is software?