The Category League Table: Which MSME Verticals Actually Get Named by AI, and Which Don't

The table that does not exist yet

No public, continuously updated dataset shows, category by category, how often India's AI answer engines name an actual micro, small, or medium enterprise rather than defaulting to generic buying advice or a link to a marketplace. What exists instead is aggregate citation data that is mostly global and domain-level, plus a small number of non-peer-reviewed industry studies testing a handful of categories in a handful of Western cities. Nothing purpose-built for India's MSME base, which the government's own real-time Udyam dashboard put at 9,24,39,224 registered enterprises as of 9 August 2026, exists in public form.

That gap matters because the answer to "should an MSME worry about AI answers" is not one number. A property lawyer and an accountant in the same city are not competing against the same aggregator gravity, and neither is competing on the same terms as a supplier already listed among IndiaMART's roughly 88 lakh registered suppliers. Treating AI visibility as one undifferentiated risk, or one undifferentiated opportunity, understates the variation the closest available evidence already shows.

This piece does two things a research-grade reading of an open question should do. It reads what has actually been published, most of it global rather than Indian, and states plainly where that evidence does and does not reach. And it proposes the method a genuine Indian category league table would need, rather than presenting an invented one.

What published citation data already shows: the aggregator head start

Ahrefs' Brand Radar analysis, built from roughly 76.7 million Google AI Overviews and close to a million prompts each in ChatGPT and Perplexity, found that Wikipedia is the single most cited domain across all three systems, drawing 16.3% of citations in ChatGPT, 12.5% in Perplexity, and 8.4% in AI Overviews. YouTube leads in Perplexity at 16.1% and AI Overviews at 9.5%, while Reddit and Quora rank high only in AI Overviews. The consistent finding across all three engines is that independent brand and small-business websites do not appear in the top ten cited domains for any of them; the citation layer is dominated by aggregators, user-generated platforms, and news outlets.

A separate Ahrefs study of 75,000 brands, filtered to domains with a Domain Rating above 40 and keywords with meaningful search volume, measured what actually correlates with being named. The strongest correlate was YouTube mentions, a Spearman coefficient of 0.737 with both ChatGPT and AI Overviews visibility. Broad brand mentions across the web correlated almost as strongly, 0.656 to 0.709 depending on the engine. Classic SEO metrics performed far worse: backlinks and URL Rating showed weak correlations, and Domain Rating only a moderate one. The practical reading is that being named now depends more on being mentioned across a network of independent, trusted sources than on ranking mechanics alone, and that is a structural head start for any aggregator or marketplace whose name already appears in thousands of reviews, directory listings, and social posts.

The peer-reviewed counterweight is the Generative Engine Optimization (GEO) paper published at KDD 2024, which tested specific content interventions rather than domain-level footprint. It found that adding cited statistics, direct quotations, and references to authoritative sources measurably raised a source's visibility inside generated answers across the systems it tested. That result matters for a small firm precisely because it does not depend on scale: a single well-structured, well-corroborated page can earn visibility the same lever an aggregator earns it with, even if the aggregator still starts with a wider footprint.

The category effect, tested outside India

The clearest evidence that naming behavior varies by category, not just by engine, comes from two independent, non-peer-reviewed industry studies run in 2025 and 2026. Neither covers India, and neither should be read as settled science, but both are transparent about method and both point the same direction: category structure predicts outcomes more than any single firm's quality does.

Where India's aggregator layer differs from the studies above

None of the studies above tested India, and India's aggregator layer is not a simple import of the Western pattern. It combines dominant single-category marketplaces with a government-backed open protocol built explicitly to avoid a single platform owning the layer. IndiaMART describes itself as India's largest online B2B marketplace, reporting 23.4 crore buyers, 88 lakh suppliers, and 13.2 crore product and service listings on its own corporate site, a scale that makes it the default aggregator for a large share of India's industrial and wholesale MSME supply base.

Running alongside it is the Open Network for Digital Commerce (ONDC), a protocol rather than a platform, on which 306 network participants were live across 616-plus cities and 26 active domains as of its own published network figures, with 7.64 lakh-plus sellers and service providers and more than 16 million total orders recorded in a single month (May 2025). ONDC was built precisely to prevent any one buyer or seller app from becoming the sole gatekeeper the way a closed marketplace can. Whether that open-protocol design changes how AI answer engines treat an ONDC-listed seller, compared to a seller on a closed marketplace or a seller with no aggregator presence at all, is not something any published study has yet tested. It is a genuinely open question, and a Indian-specific variable the method below needs to account for rather than assume an answer to.

A proposed method for building the missing table

Building a real Indian category league table is not a technical problem so much as an undone one. A credible version would need a category taxonomy anchored to an existing standard, such as Udyam or NIC sector codes, rather than an invented list, so results are comparable to India's own MSME classification. It would need a consistent query set per category, mirroring the four-prompt structure the PolyGrowth study used (generic, qualified, neighborhood-level, and recommendation-framed), run across at minimum the three engines the closest global studies already track: ChatGPT, Gemini, and Perplexity.

Each response would need to be scored into one of a small number of clear categories, a named individual business, generic buying advice with no business named, a link or reference to a marketplace or directory rather than an individual seller, or no usable answer at all, and that scoring would need to repeat on a fixed cadence, because model outputs drift and a single snapshot cannot support a claim about a category's standing over time. The output of that method is a per-category naming rate relative to other categories, not a single blended national score, because the evidence above already shows a blended score would hide the exact variation that matters to an individual business owner deciding where to invest attention.

What the evidence already implies, and what it does not

Two things can be said with reasonable confidence from what is already published, without inventing an Indian-specific number. First, categories that already have a credentialed, structured directory layer, the way property lawyers do in the PolyGrowth data and home services and B2B categories do in the Local Dominator data, are likely to show measurably different naming behavior than categories without one, largely independent of any individual firm's quality. Second, the mechanics that correlate with being named, broad mentions across a network of trusted third parties and structured, cited content, structurally favor whichever entity already has the wider, more corroborated footprint, which in most categories today is the marketplace or directory rather than any single MSME.

What remains genuinely unknown is the exact naming rate for any specific Indian vertical, whether ONDC's open-protocol design shifts the pattern in a way a closed marketplace does not, and whether queries in Hindi or another Indian language change engine behavior relative to the English-language tests summarized here. Those are the questions the proposed method above is built to answer. Until someone runs it for India, the responsible claim is the one this article has made: category structure is a real and evidenced predictor of AI-naming behavior, and no one has yet measured where India's own MSME categories fall on that spectrum.