Methodology

1. 1Every store gets the same treatment: a category-appropriate shopping mission (e.g. “buy a men's slim-fit shirt under ₹1500”), executed by the same agent — at least three times, with more runs where an early read was inconclusive.
2. 2The agent operates a real Chromium browser through a DOM-distillation loop: it sees the page's interactive elements and a screenshot, and acts one step at a time — no site-specific code, no selectors, no shortcuts.
3. 3The journey is measured across six stages: discovery → product page → cart → checkout → delivery form → payment surface.
4. 4A store's result is the median of its runs. A stage outcome or failure finding is published only when at least 2 of 3 runs agree; otherwise the store is flagged inconsistent.
5. 5Runs where the agent itself was at fault (model stall, tool error, misread mission) are voided — they never count against a store, and are queued for re-runs.
6. 6If every run on a store voided this way, there is no store-attributable evidence to rank on. That store is excluded from the leaderboard (listed separately as inconclusive, not scored) rather than published with a number that reflects agent error, not store behavior.

• The agent never pays. The journey terminates the moment a payment surface (UPI field, card form, COD option) is reached. No orders are placed.
• Bot-blocks are findings, not obstacles to defeat. If a store blocks automated browsers, that is recorded and the run ends. The agent never evades detection.
• The synthetic shopper uses an owner-controlled email and a non-receiving phone number. OTP walls are terminal findings — no OTP is ever received or entered.
• Runs are throttled (≥60s between visits to the same store) and capped, so the load on any store is a small fraction of one human shopping session.

ARI score = 70% agent journey + 30% static machine-legibility. The agent-journey half is weighted by stage:

Checkout access carries the largest weight because forced login and OTP walls are the single biggest blocker to agentic commerce. Graded dimensions (PDP legibility, checkout access, form friction) award partial credit; the rest are pass-through. Static legibility covers structured product data (JSON-LD), robots.txt stance toward AI agents, llms.txt, sitemap, and OpenGraph tags.

Every finding is backed by run logs and evidence screenshots. If you run one of these stores and believe a result is wrong, write to nirat.adi+ari@gmail.com — reproducible corrections get the store re-run and the scorecard updated.