The Export Answer Gap: When a Foreign Buyer Asks AI for an Indian Manufacturer, Who Gets Named?

Half the export base, and a buyer who no longer browses first

The scale of what is at stake is not in dispute. India's Economic Survey 2025-26 puts the MSME sector at more than 7.47 crore enterprises, contributing 31.1% of GDP, 35.4% of manufacturing output, and 48.58% of the country's exports. A separate government-data report cited by KNN India shows the MSME share of merchandise exports rising from 45.74% in 2023-24 to 48.55% in 2024-25 in value terms. However the two figures are rounded, the direction and the order of magnitude agree: close to half of what India sells abroad is made by a small or medium enterprise, not a large corporation.

What has changed is not the size of that export base but how the buyer on the other end finds it. A May 2026 Gartner survey of 645 B2B buyers, conducted in August and September 2025, found that buyers now use an average of seven information sources during a purchase, and 45% reported using generative AI, primarily to gather information on vendors and products. The same survey found 69% still turn to a human sales contact to validate what the AI told them, so the pattern is not full automation. It is a new first step inserted ahead of the conversation a buyer used to start by browsing a directory or a trade-show floor.

That new first step is where this piece is aimed. If a foreign buyer opens ChatGPT, Perplexity, or Google's AI Mode and asks a sourcing question, something has to answer, and that something has to name a small number of entities out of a very large field. The open question is which kind of entity gets named: the individual exporting MSME, or the aggregator platform that lists thousands of them.

What India built for the human buyer

India's export-promotion machinery is substantial, and it was designed for a buyer, or a trade officer, who reads a page. The Directorate General of Foreign Trade runs the Trade Connect ePlatform, a digital gateway that gives MSME exporters access to tariff data, certification requirements, verified buyer directories, and paperless issuance of certificates of origin, alongside multilingual courses on export procedure. The Federation of Indian Export Organisations, the apex export-promotion body set up by the Ministry of Commerce in 1965, serves the interests of roughly 200,000 exporters directly and indirectly.

In November 2025 the Union Cabinet approved the Export Promotion Mission, a flagship scheme with a total outlay of ₹25,060 crore running from FY 2025-26 through FY 2030-31, structured around two sub-schemes: Niryat Protsahan for financial enablers such as trade finance, and Niryat Disha for non-financial, market-access, and ecosystem support, explicitly targeted at MSMEs, first-time exporters, and labour-intensive sectors.

This is real infrastructure, and it addresses a real problem: helping a small manufacturer find financing, certification, and a verified buyer. None of it, however, was built to answer a different question that now sits upstream of all of it: when a foreign buyer types a sourcing question into an AI system rather than a search box, does any of this infrastructure make the individual exporter legible to the system doing the answering? A buyer directory built for a human reading a webpage and a citation graph built for a model generating an answer are not the same artifact, even when they describe the same exporter.

What AI shopping and sourcing systems already run on

The infrastructure that generative engines actually ground their commerce answers in is now documented, and it is not built around the individual small exporter. Google has said its Shopping Graph, the data layer behind AI Mode's shopping features, now holds more than 50 billion product listings, with more than 2 billion of them refreshed every hour, fed by structured product feeds submitted through Google Merchant Center. A listing earns a place in that graph by submitting clean, continuously updated structured data at scale, which is exactly the kind of feed a large platform maintains as a matter of course and an individual manufacturer, without dedicated staff for it, usually does not.

OpenAI's commerce build-out tells a similar story, and its recent history is instructive about where the priority sits. OpenAI and PayPal announced an Agentic Commerce Protocol partnership in October 2025, and OpenAI rolled "Buy it in ChatGPT" out to all United States ChatGPT users, including the free tier, on February 16, 2026. By March 2026, OpenAI had pulled back from letting purchases complete directly inside a chat and pivoted toward routing transactions through established merchant apps such as Instacart, Target, and Booking.com. Even the company building agentic commerce from scratch chose to plug into large, already-structured retail platforms rather than build discovery for the long tail of individual suppliers first.

None of this infrastructure, as documented, currently has a cross-border B2B sourcing use case in view; it is consumer retail, and it is United States-first. That matters for how the export-answer gap should be read: the machinery an AI system leans on to answer a shopping question was not built with an Indian export manufacturer in mind at all, in either direction.

The aggregator's structural advantage

Set the AI shopping infrastructure next to the marketplaces that already dominate B2B sourcing, and the mechanism behind an aggregator default becomes visible without needing to allege intent. IndiaMART, India's largest online B2B marketplace, reported 8.7 million supplier storefronts as of its Q3 FY26 results in January 2026, a 6% year-on-year increase, with 221,000 paying suppliers at quarter-end. Alibaba.com describes its platform as connecting sellers with more than 40 million registered buyers across more than 200 countries, a figure the company publishes itself and that should be read as a marketing claim rather than an audited count.

A marketplace like either of these maintains exactly the properties that make a source easy for an answer engine to trust and cite: structured, consistently formatted product pages at enormous scale, frequent updates, a single well-known domain with years of accumulated links and mentions, and, for Alibaba specifically, a brand a language model has almost certainly encountered many times over in its training data. An individual exporting MSME, even a highly capable one, typically has one website, updated irregularly, with none of the schema markup, third-party corroboration, or update cadence that a model's underlying retrieval and ranking systems are built to reward.

What has not actually been measured yet

The limit of the evidence here is real. No publicly available study has run a structured, multi-engine test of buyer-intent export queries, such as "who manufactures stainless steel kitchenware in India for export" or "verified Indian supplier of cotton yarn," against a defined set of product categories that map to India's actual MSME export mix, and then recorded whether the named entity was an individual exporting firm, a marketplace listing, a trade directory, or a general news article. Absent that study, any specific percentage claiming to describe "how often AI names the aggregator over the MSME" would be invented.

What can be said is that the surrounding evidence points in a consistent direction. The infrastructure AI commerce systems ground their answers in rewards structured scale (the Shopping Graph, the marketplaces), the infrastructure India has built for exporters rewards human navigation (Trade Connect, FIEO's directories, the Export Promotion Mission), and no bridge between the two has yet been documented or measured. That gap between what exists and what has been tested is itself the finding worth reporting.

A workable method for closing that gap would sample real HS-code product categories weighted by MSME export share, run a fixed set of buyer-intent query templates across ChatGPT, Perplexity, Gemini, and Google AI Mode at a defined cadence, classify each named source as manufacturer-owned, marketplace, directory, or other, and track how that mix moves as individual exporters improve their own structured data and third-party corroboration. It is also worth noting that India's Open Network for Digital Commerce has stated an intention to extend into cross-border trade as one of several planned expansions, alongside travel and hospitality, but has not yet made that live; there is currently no domestically built discovery layer purpose-made for export buyers either.

What the evidence supports, and what it does not

Two things are true at once, and neither should be flattened into the other. MSMEs are not a marginal part of India's export story; they generate close to half its value, and the government has committed real money, ₹25,060 crore over six years, to keep growing that share. At the same time, the growing share of buyers who now ask an AI system before they ask a person means the channel that decides who gets a first conversation is shifting toward infrastructure that was not built around individual exporters at all.

The responsible conclusion is not that Indian MSMEs are invisible to AI buyers, and not that the problem is solved because export support programs exist. It is that the question has not yet been measured, the structural evidence available says the mechanism favors scaled, structured aggregators over individual firm websites by default, and any exporter that wants a different outcome for itself has a concrete, testable thing to work on: whether its own identity, catalog, and credentials are structured and corroborated in the forms these systems can actually parse and cite, independent of whether a marketplace listing exists for it too.