Picture the next version of a repeat purchase. A customer tells an assistant to reorder the face wash they bought in June, under ₹800, delivered before Friday. The assistant checks three brands, compares price and delivery, and places one order. Nobody loads a product page. Nobody sees your hero image, your reviews carousel, or the offer bar you spent a sprint testing.
That is the shift worth preparing for. Not "AI is coming to retail", but something narrower and more concrete: for a growing slice of demand, the buyer is software, and software does not shop the way people do.
The commerce world has spent 2025 and 2026 building the plumbing for this. The Agentic Commerce Protocol came out of OpenAI and Stripe and powers checkout inside ChatGPT and compatible agents. Universal Commerce Protocol was announced by Google and Shopify, covering discovery through to post-purchase. The sponsors differ. What they ask of a merchant does not differ much at all.
The storefront becomes an interface
A useful way to hold this: your website has two audiences now, and only one of them has eyes.
For the human, the site is still a shop. Design, story, proof and reassurance all do their job.
For the agent, the site is an API. It wants product data it can parse, a stock number it can trust, a delivery promise it can compare, and a checkout it can complete without a human clicking anything. As one 2026 guide puts it, protocol problems have replaced user experience problems - instead of testing button colours, merchants are defining endpoints and feeds.
That reframing is the whole article. Everything below is what the second audience needs.
The six surfaces an agent reads
A product feed with real identifiers. Agents work from structured catalogues, not scraped pages. For ACP, merchants push a compressed feed file to an endpoint with daily updates: title, description, price with an ISO currency code, availability, images, and flags for whether an item is eligible for search and for checkout. Titles cap at 150 characters, descriptions at 5,000. Schema.org Product markup and GS1 GTINs are the common denominators across implementations. If your catalogue has invented SKU formats and no barcodes, you are invisible to the comparison step.
Stock that is true at the moment of the call. A human sees "only 2 left" and discounts it. An agent treats it as fact and books against it. If your availability is a nightly batch, every order placed between syncs is a coin toss.
A delivery promise that is computed, not decorative. "Delivered in 3-5 days" written into a theme file is marketing copy. An agent comparing you against two competitors needs a figure tied to pincode, courier serviceability and current cut-off time. Brands that invested in live carrier connectivity and normalised tracking are the ones whose operations are legible to agents today.
Policies as data. Returns window, replacement rules, COD availability, cancellation terms. If these live only in a styled page written for humans, an agent either skips you or guesses. Either outcome is yours to lose.
A checkout an agent can drive. Programmatic add-to-cart, programmatic shipping selection, and the ability to accept agent attestation headers from the payment processor. This is the piece most brands cannot retrofit in a week, which is why it belongs on a roadmap rather than a to-do list.
Post-purchase that answers machines. Order status, tracking and returns initiation that an agent can query on the customer's behalf. If your only status channel is a WhatsApp number staffed by a person, the agent cannot help, and the customer comes back to you by hand.
What stops working
This is the uncomfortable half, and the reason the shift is strategic rather than technical.
Most conversion rate work of the last decade targets a human mid-session: urgency timers, scarcity badges, exit intent popups, bundle upsells placed at a specific scroll depth, trust badges near the pay button. On the agent path, none of those fire. There is no session, no scroll, no hesitation to interrupt.
What survives is anything that still reads as structured fact. Price survives. A genuine delivery advantage survives. A real returns policy survives. Ratings survive if they are exposed as data rather than as a widget. Brand preference survives, because the customer named you in the instruction - which makes the brand work you already do more valuable, not less.
What does not survive is persuasion aimed at a pair of eyes. If a meaningful share of your conversion lift comes from on-page pressure tactics, that lift does not transfer. Worth knowing before you plan next year's CRO roadmap.
We wrote separately about how agents decide which brand to pick in agentic commerce and 'buy for me' agents, and about the discovery side in optimising for AI shopping agents. This piece is the layer underneath both: what has to be true of your systems for either to matter.
The readiness test, in four questions
Run these against your own store this week. They are deliberately answerable without a vendor.
Can a machine get your catalogue without scraping? If the honest answer is "they can read our website", you have discovery exposure. A feed with identifiers, prices and availability is the entry ticket.
Is your published stock number true right now? Not "mostly". Pick five fast-moving SKUs, compare the number on the storefront against the warehouse system at the same minute. The gap you find is the rate at which agents will book orders you cannot fill.
Is your delivery estimate computed per pincode? If it is a static string, an agent comparing you on speed is comparing a claim with a competitor's calculation.
Could an order be completed without a human touching a screen? Not whether you want that today - whether the path exists at all.
Score it honestly. Most Indian D2C brands we see clear the first question, struggle with the second, and have not started the fourth.
What to actually do first
Start with the feed and the stock number, in that order, and ignore the protocol conversation entirely until both are clean.
The reason is unglamorous: both are the same work your own operations already need. A correct catalogue with identifiers improves your marketplace listings and your ad feeds. Accurate availability reduces cancellations, which protects marketplace seller metrics and saves a support conversation you are paying for. If agent volume in India stays small for another year, the work still pays.
The order looks like this. Clean the catalogue so every SKU has a stable identifier, a parseable title and a correct price. Make availability reflect reality on a cadence measured in minutes, not nights. Replace the static shipping promise with a computed one. Expose policies as data. Only then look at agent-facing checkout, which is where platform and payment choices start to matter.
If your stock and order state live in three systems that disagree, none of this is reachable, and the fix is architectural rather than cosmetic - that question is covered in when a Shopify brand actually needs an OMS.
The India-specific wrinkle
Two things make this harder here than in the markets the protocols were designed around.
The first is cash on delivery. A meaningful share of Indian D2C orders still arrive unpaid, and published 2026 benchmarks put COD return-to-origin rates at 30-40 percent against 20-30 percent for prepaid. An agent that can place COD orders on a customer's behalf, with no human weighing whether they actually want the item, pushes volume into exactly the payment mode that fails most often. Deciding whether agent-placed orders are prepaid-only is a commercial decision, and it is better made deliberately than discovered in a month-end RTO report.
The second is address quality. Indian addresses carry landmarks, informal locality names and inconsistent pincode usage, and address problems already drive a large share of failed deliveries. A human buyer corrects their address when the form looks wrong. An agent passes through whatever it holds. If your checkout relies on a shopper noticing a bad autofill, that safety net disappears.
Neither is a reason to stay out. Both are reasons to decide your rules before the volume arrives rather than after.
The part worth sitting with
Agent-led buying rewards operational honesty in a way human shopping never did. A shopper forgives a stale stock count, a vague delivery window, a returns policy buried in a link. They shrug and buy anyway, because they already decided they liked you.
Software does not forgive any of it. It reads what you published and acts on it. Published a number you cannot honour, and you get the order and the cancellation and the damaged metric.
That is the real preparation: not a protocol integration, but being a brand whose published facts are true. The brands that get there will find the agent conversation easy. The ones that do not will find that the first thing agents learn about them is that their data cannot be trusted.
If you want the gap measured rather than guessed - what your catalogue, stock accuracy and delivery promise actually look like to a machine - book an operations review. We test the four questions above against your real systems and come back with the order of work, including the cases where the answer is that your current setup is fine for now.
