MSME & Global Commerce · emerging evidence

The Open Catalog Advantage: Why an MSME's ONDC Listing May Outlast Its Amazon Listing as AI Agents Start Shopping

Last reviewed 2026-08-09. Written by Chandranshu Kumar, Founder, Raveneye Global. · 12 min read

An Amazon listing and an ONDC listing look similar to a shopper, but they sit on opposite architectures, and that difference grows in weight as AI shopping agents start doing the finding and the choosing. The question of ONDC AI agent discovery turns on that architecture. An Amazon listing lives inside one company's closed system: the platform controls the catalog format, the search that surfaces it, and whether any outside agent is allowed to read it at all. An ONDC listing is published to an open protocol, Beckn, whose specifications are public on GitHub, and it is discovered by multicast, meaning one buyer search is broadcast to every seller whose catalog matches, across competing buyer apps. Reading the protocol specifications, that open, multicast, portable catalog is structurally closer to what an autonomous shopping agent needs than a listing locked behind a proprietary interface. This piece makes that structural case from public documents, sets it against the new agent-payment protocols such as Google's AP2, and states the limits plainly: this is an architectural argument, not a measured forecast, and ONDC is still early and India-only.

The listing looks the same, the architecture does not

To a person browsing on a phone, a product on Amazon and the same product on an ONDC buyer app are hard to tell apart. Both show a price, a picture, a description, and a way to buy. The similarity is on the surface. Underneath, one is an entry in a single company's private database, reachable only through that company's own app and its own rules, and the other is a record published to a shared, public protocol that any compliant application is allowed to query.

That distinction was easy to ignore while the reader was always a human tapping a screen. It stops being easy to ignore once the reader is a machine. AI shopping agents do not browse storefronts the way people do. They read catalogs, compare structured records, and act through interfaces. The question that decides whether a small business is reachable by such an agent is no longer how attractive its listing looks. It is whether the agent can read the catalog at all, and whether it is allowed to.

This article argues, from public protocol documents rather than from any test we ran, that an MSME's ONDC listing is structurally better positioned for that machine-read future than its listing inside a closed marketplace. The word to hold onto is structurally. The claim is about how the two systems are built, not a prediction of specific market share, and it comes with limits that the final sections state plainly.

What ONDC actually is: a protocol, not a marketplace

The most common misreading of ONDC is that it is a government-run version of Amazon, another destination where sellers post products. It is not a destination at all. In its own words, the Open Network for Digital Commerce states that it "is not a central mediator or intermediary, neither does it interface with buyers and sellers," and that it "is merely a communication protocol that allows Buyer Network Participants and Seller Network Participants to interact with each other." There is no ONDC app to shop on. There is a set of rules that many apps agree to speak.

Those rules are an extension of the Beckn protocol, an open specification for decentralized commerce transactions. The ONDC-Official organization on GitHub publishes the protocol layer over Beckn, the developer documentation, domain specifications for retail, logistics, financial services and mobility, and reference implementations in several languages. The grammar a seller's catalog must follow is not a trade secret held by one firm. It is written down in public and anyone can implement it.

ONDC was set up in India as a non-profit at the end of 2021, with an advisory council that includes Nandan Nilekani, the architect of India's earlier public digital infrastructure. Its framing is deliberately infrastructural rather than commercial: a common rail that competing buyer apps and seller apps ride, in the same spirit as a payments network that no single bank owns.

Multicast discovery, in plain terms

The mechanism that matters most for the agent argument is how discovery works. When a buyer app sends a search into the network, it goes to a gateway. If the search does not name a specific seller, the gateway broadcasts, or multicasts, that request to every seller application whose catalog matches the context of the request. ONDC describes the gateway's job as ensuring "discoverability of all sellers in the ONDC Network by multicasting the search request received from buyer applications to all the seller applications."

Read that carefully, because it inverts the platform model. On a closed marketplace, a seller's visibility is a function of that platform's ranking, advertising, and internal logic, and it exists only inside that one app. On ONDC, a seller's catalog is a candidate for every matching search across every buyer app on the network at once. The listing is not tied to a single front door. It is answerable to many, through one shared query grammar.

What an AI shopping agent needs from a catalog

Set the two architectures aside for a moment and describe the reader. An autonomous shopping agent, whether it lives in a chat assistant, a browser, or a payment app, needs four things from the supply side to act on a buyer's behalf. It needs a catalog it can parse without scraping a human page. It needs a consistent grammar so that a product from one seller can be compared to a product from another. It needs the ability to move from discovery to an order through a defined sequence of steps. And it needs to be permitted to do all of this without depending on a single gatekeeper's goodwill.

A protocol network is, almost by definition, built to supply exactly those four things. Beckn defines the message shapes for search, selection, initialization, and confirmation, so the path from finding a product to placing an order is a standard set of calls rather than a bespoke integration per store. Because the specification is public and shared, a product record from one seller carries the same structure as a record from another, which is precisely the comparability an agent depends on. And because discovery is multicast across the network rather than owned by one app, no single company sits between the agent and the catalog with the power to switch access off.

None of this was designed with 2026 shopping agents in mind. The point is that a system built to let many independent applications transact against many independent catalogs, through one open grammar, happens to describe the interface an agent wants. The peer-reviewed research on Beckn, published in Information and Software Technology in 2026, reaches a compatible conclusion on different grounds: it argues that decentralized transaction networks are "more inclusive for small players" and "more resilient against adversarial events" than platform-based models. Inclusive and resilient are also the properties that make a catalog dependable for an automated buyer.

Why a closed listing is reachable only at the platform's discretion

Now the contrast. A listing inside a closed marketplace is a strong asset today because that marketplace brings its own enormous stream of human buyers. But its reachability by an outside agent is a decision the platform makes, not a property the seller owns. The catalog format is proprietary. The search that surfaces it is proprietary. The interface an external agent would use to read it exists only if, and for as long as, the platform chooses to publish it, on the platform's commercial terms.

This is not a hypothetical concern. It is the same dependence that sellers already feel through fees and rule changes. Across categories, Amazon's referral fees commonly run from 8 to 15 percent, and third-party analyses estimate that many sellers pay something closer to a quarter to a third of revenue once fulfillment and other charges are added. Amazon publicly committed to hold United States referral and fulfillment fees steady through 2025 and to add no new fee types, then announced fee changes taking effect in January 2026. The specific figures matter less than the pattern: the terms of participation, including who can access the catalog and how, are set by the platform and can be revised by the platform.

Carry that pattern into an agent-mediated market. If autonomous agents become a meaningful channel, a closed platform has every incentive to decide which agents may read its catalog, on what commercial terms, and with what priority, exactly as it decides advertising placement today. A seller whose only machine-readable presence lives inside that system inherits that dependence. The listing can be excellent and still be unreachable by a given agent, because reachability was never the seller's to grant.

The agent-payment protocols point the same way

The clearest signal that commerce is moving toward open, machine-readable grammars is that the largest technology companies are now racing to define them for the payment step. In September 2025, Google announced the Agent Payments Protocol, or AP2, described as an open protocol developed with more than sixty payment and technology partners, including Mastercard, PayPal, American Express and Coinbase, to let AI agents initiate and settle payments across platforms. AP2 is positioned as an extension of the Agent2Agent protocol and the Model Context Protocol, and it records the buyer's instruction and their approval of a specific cart as cryptographically signed mandates, so there is an auditable record of what the human actually authorized.

AP2 is deliberately payment-method-agnostic, starting with cards and adding real-time push rails such as India's UPI and Brazil's Pix and digital assets on its roadmap. Around the same time, OpenAI and Stripe released the Agentic Commerce Protocol under an open Apache 2.0 license and switched on Instant Checkout inside ChatGPT for Etsy and Shopify merchants. Two of the biggest names in software concluded, independently, that agent-led commerce needs an open, shared standard rather than a private one.

Here is the connection to ONDC. These new protocols standardize the payment and the checkout handshake, the moment money moves. They largely assume the discovery and ordering grammar as a separate concern. ONDC, through Beckn, already is an open, shared grammar for the discovery and ordering half of the transaction, live at national scale in one large market. An open payment protocol and an open commerce protocol are complementary layers of the same idea, that machines transacting on a person's behalf need public grammars, not proprietary ones. A seller already fluent in the open discovery grammar is positioned for the world the payment protocols are being built for.

The structural case, and its limits

The argument so far is an architectural one, and it deserves architectural humility. ONDC is real and growing. Its own network figures report more than 616 cities live, over 300 network participants, more than 7.6 lakh sellers and service providers, and orders in the range of 16 million in a month, up from roughly 10 million in mid-2024. Those are serious numbers for a four-year-old network. They are also a fraction of the volume that flows through the largest closed marketplaces, and the network is confined to India.

Several caveats follow. Agent-mediated shopping is early, and no one, including us, can yet measure how much commerce will actually route through autonomous agents or on what timeline. We ran no test, benchmark, or measurement of agent behavior against ONDC or against any marketplace for this article; every claim here is read from public protocol documents and public network statistics. An open specification makes a catalog readable in principle, but real interoperability still depends on data quality, catalog completeness, and adoption, and ONDC has documented growing pains on all three. And a closed platform could choose to publish an excellent open agent interface of its own, which would narrow the gap this article describes.

What survives all of that is the structural point, and it is worth stating precisely. The value of an ONDC listing to a future agent rests on a property the seller holds: the catalog is published to a public grammar that any compliant agent may read. The value of a closed listing to a future agent rests on a decision the platform holds: whether to expose that catalog to that agent, and on what terms. Over a long enough horizon, an asset whose reachability the seller owns tends to outlast an asset whose reachability someone else grants. That is the sense in which an MSME's ONDC listing may outlast its marketplace listing once autonomous agents take over the finding and choosing. It is a claim about who controls access, not a promise about next quarter's sales.

What this means for an Indian MSME right now

The practical implication is not to abandon the marketplace that pays the bills today. It is to stop treating a single closed listing as the whole of a business's machine-readable presence, and to build presence in the places whose reachability the business actually controls. For most Indian MSMEs that means three things at once: keeping the marketplace listing where the human buyers are now, establishing a real presence on the open network so the catalog lives on a public grammar, and making sure the business's own identity and information are structured and consistent everywhere a machine might read them.

That last point is where the open-network question meets the broader machine-readiness question. Being on ONDC handles the protocol-level discovery of a business's products. It does not, on its own, handle whether an answer engine can describe the business correctly, whether the map surfaces the business for a nearby search, or whether its reputation is legible enough for a machine to trust. Those are separate surfaces, and a business can be strong on one and invisible on another. The useful discipline is to measure across all of them rather than assume that presence in one place implies presence in the rest.

The framing to end on is the one this whole cluster keeps returning to. The reader of a business is increasingly a machine, and machines reward legibility, consistency, and open access over polish. An open catalog is one instance of that principle, an important one, but only one. The question worth asking about a given business is not whether its marketplace listing looks good. It is which of the surfaces a machine now uses can actually read that business, and which cannot yet.

The evidence

Key findings, with their sources

ONDC describes itself as "not a central mediator or intermediary" but "merely a communication protocol" that lets buyer and seller network participants interact, with no ONDC shopping app of its own.

established ONDC official website, ondc.org.
ONDC gateways ensure discoverability of all sellers by multicasting a buyer app's search request to all matching seller applications on the network, rather than confining a listing to one storefront.

established ONDC official website, ondc.org.
The ONDC-Official GitHub organization openly publishes the protocol layer over Beckn, developer documentation, domain specifications, and reference implementations in multiple languages.

established ONDC-Official, github.com/ONDC-Official.
ONDC reports more than 616 cities live, over 300 network participants, more than 7.6 lakh (760,000+) sellers and service providers, and orders around 16 million in a month.

established ONDC official website network figures, ondc.org (2025).
ONDC monthly transactions reached roughly 10 million in mid-2024, up from about 1 million in early 2023, showing a young network growing by roughly a million transactions a month.

contested Reported ONDC network figures, 2024 (secondary reporting).
Peer-reviewed research on the Beckn protocol argues decentralized transaction networks are "more inclusive for small players" and "more resilient against adversarial events" than platform-based models.

established Dittrich et al., "Beyond Platforms: Growing Distributed Transaction Networks for Digital Commerce", Information and Software Technology, 2026, arXiv:2504.18602.
Google announced the Agent Payments Protocol (AP2) in September 2025 as an open protocol with 60+ partners, extending A2A and MCP, recording buyer intent and cart approval as cryptographically signed mandates.

established Google Cloud Blog, "Announcing Agent Payments Protocol (AP2)", September 2025, cloud.google.com.
AP2 is payment-method-agnostic, starting with cards and listing real-time push rails such as UPI and Pix and digital assets on its roadmap.

established Google Cloud Blog and AP2 documentation, ap2-protocol.org, 2025.
OpenAI and Stripe released the Agentic Commerce Protocol under an open Apache 2.0 license on September 29, 2025, and enabled Instant Checkout in ChatGPT for Etsy and Shopify merchants.

established Stripe Newsroom, "Stripe powers Instant Checkout in ChatGPT", September 2025, stripe.com; OpenAI, "Buy it in ChatGPT", openai.com.
Amazon referral fees commonly run 8 to 15 percent by category, with third-party estimates of 25 to 30 percent of revenue in total fees; Amazon held US fees steady through 2025 and announced changes for January 2026.

contested Industry analyses of Amazon seller fees, 2025-2026, plus Amazon fee announcements.

Calibration

What is proven, what is promising, what is unproven

Evidence tier	Tactics	What the evidence says
established	ONDC as an open protocol rather than a marketplace; multicast discovery across buyer apps; open specifications on GitHub; the arrival of open agent-payment and agentic-commerce protocols (AP2, ACP); the decentralized-network research finding.	ondc.org; github.com/ONDC-Official; Dittrich et al. (Information and Software Technology, 2026); Google Cloud Blog (AP2, 2025); Stripe and OpenAI (ACP, 2025).
emerging	The reading that an open, multicast, portable catalog is structurally better positioned for autonomous shopping agents than a proprietary listing, and that open discovery and open payment protocols are complementary layers of one shift.	Original analysis of the public protocol specifications and the composition of Beckn-style discovery with AP2-style payment; agent-mediated commerce is still early and its volume is unmeasured.
contested	The specific prediction that an ONDC listing will "outlast" a marketplace listing in practice, and any exact figure for marketplace fees or ONDC transaction growth.	A structural argument about who controls access, not a measured forecast; fee and growth figures come from secondary and industry sources and vary by category and date.

Reference

Glossary

ONDC: The Open Network for Digital Commerce, an India-based non-profit that runs an open protocol, not a marketplace, letting independent buyer apps and seller apps transact through a shared grammar rather than inside one company's system.
Beckn protocol: The open specification for decentralized commerce transactions that ONDC extends. It defines standard message shapes for search, selection, order initialization, and confirmation, so any compliant app can transact against any compliant catalog.
Multicast discovery: The ONDC mechanism where a buyer app's search is broadcast by a gateway to every seller whose catalog matches, so a listing is discoverable across many buyer apps at once rather than tied to a single storefront.
Agentic commerce: Shopping in which an autonomous AI agent does the finding, comparing, and buying on a person's behalf, which shifts the important audience for a listing from a human reader to a machine reader.
AP2 (Agent Payments Protocol): Google's open protocol, announced in 2025, that lets AI agents authorize and settle payments across platforms using cryptographically signed mandates recording what the human actually approved.
Machine-readable catalog: A product catalog structured to a public, documented grammar so that software, as well as a human browsing a page, can parse it, compare it to others, and act on it without scraping.

Straight answers

Frequently asked questions

Is ONDC just an Indian government version of Amazon?

No. ONDC is not a marketplace and has no shopping app of its own. It calls itself "merely a communication protocol" that lets independent buyer apps and seller apps interact. Amazon is a single company that owns its catalog, its search, and its interface. ONDC is a shared set of open rules that many competing apps agree to speak, closer to a payments rail than to a store.

Why would an open catalog matter for AI shopping agents?

An agent needs to read a catalog without scraping a human page, compare products in a consistent grammar, move from discovery to an order through defined steps, and do so without depending on one gatekeeper. An open protocol network like ONDC is built to supply exactly those properties. A closed listing is readable by an outside agent only if the platform chooses to expose it, on the platform's terms.

Does this mean an MSME should leave Amazon?

No, and this article does not claim that. The marketplace is where large numbers of human buyers are today, and that value is real now. The point is not to abandon it but to stop treating a single closed listing as the whole of a business's machine-readable presence, and to also build presence on surfaces whose reachability the business actually controls, including the open network and its own structured business information.

What is AP2 and how does it relate to ONDC?

AP2 is Google's open Agent Payments Protocol, announced in 2025, that standardizes how AI agents authorize and settle payments. It handles the money step and largely assumes discovery as a separate concern. ONDC, through Beckn, is already an open grammar for the discovery and ordering step. They are complementary layers of the same shift toward open, machine-readable commerce grammars rather than proprietary ones.

Is the claim that an ONDC listing will outlast a marketplace listing proven?

No. It is a structural argument, not a measured forecast. It rests on one durable point: an ONDC listing's reachability by an agent is a property the seller holds, because the grammar is public, while a closed listing's reachability is a decision the platform holds. Agent-mediated commerce is still early, its volume is unmeasured, and no test was run for this article. It is directional, not settled.

Provenance

Sources

ONDC official website (self-description, multicast discovery, and network figures), ondc.org (established)ondc.org
ONDC-Official GitHub organization (open protocol specifications, developer docs, reference implementations), github.com/ONDC-Official (established)github.com
Dittrich et al., "Beyond Platforms: Growing Distributed Transaction Networks for Digital Commerce", Information and Software Technology, 2026, arXiv:2504.18602 (peer-reviewed, established)arxiv.org
Google Cloud Blog, "Announcing Agent Payments Protocol (AP2)", September 2025 (established)cloud.google.com
AP2 Agent Payments Protocol documentation, ap2-protocol.org (established)ap2-protocol.org
Stripe Newsroom, "Stripe powers Instant Checkout in ChatGPT and releases the Agentic Commerce Protocol", September 2025 (established)stripe.com
OpenAI, "Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol", September 2025 (established)openai.com
Business Standard, "Nandan Nilekani pitches ONDC to start-ups to democratise e-commerce", 2023 (established)business-standard.com
Reported ONDC monthly transaction figures (roughly 10 million by mid-2024), StockGro, 2024 (contested, secondary reporting)stockgro.club
Industry analyses of Amazon seller fees, 2025-2026 (contested, industry estimates)litcommerce.com

Every figure above is attributed to a real, dated source and tagged with its evidence tier. Where a claim could not be verified to a primary source, it is not stated as fact.