GEO Isn't SEO for Bharat: Do Aggregators Eat the Answer Before the MSME Gets a Turn?

The question Bharat's MSMEs cannot yet see the answer to

Ask ChatGPT or Perplexity for a steel pipe supplier in Ludhiana or a reliable AC repair service in Chennai, and the assistant has to decide, in a fraction of a second, which sources to trust and which few names to put in the answer. The practical question behind this article is whether that decision structurally favors a small number of large aggregators, IndiaMART for B2B manufacturing and trade, JustDial and Sulekha for local services, over the individual business a buyer actually needs to reach. It is the Bharat-specific version of a debate playing out globally as generative answer engines replace the ranked list with a synthesized recommendation.

The stakes sit on top of a population that has the least capacity to respond to a citation disadvantage it cannot see. Over 7.83 crore enterprises, more than 78 million, had registered on India's Udyam Registration Portal and Udyam Assist Platform as of 28 February 2026, according to the Ministry of MSME's reply in the Rajya Sabha. Yet the same government's own survey partner found that internet usage for actual entrepreneurial activity, placing orders, taking UPI payments, transacting online, ran at only about 26.7 percent among unincorporated enterprises in 2023-24, and that a mere 81 establishments per 1,000 used the internet to make sales at all. That is the base population any citation-concentration effect would land on, whether or not anyone has yet measured the effect directly.

What Generative Engine Optimization actually measures

The relevant research anchor is the 2024 paper that coined the term: Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, and Deshpande's "GEO: Generative Engine Optimization," accepted to KDD 2024. The paper's central move is a reframing. Classic SEO optimizes for rank in a list a human scans. GEO treats the unit of competition as a citation inside a synthesized answer, a fundamentally different target, since an engine can answer a query while naming only a handful of sources out of the thousands that rank for it.

Using a benchmark of diverse queries paired with real web sources (GEO-bench), the authors tested which content-level changes moved a source's visibility inside generated answers, and found that adding citations, statistics, and quotations from authoritative sources could lift visibility by up to 40 percent in the engines they tested, though the effect varied meaningfully by domain, the paper is explicit that no single tactic works uniformly everywhere. The paper was built and tested on general web queries, not India-specific B2B or local-service search, which matters: nothing in it directly measures Indian MSME categories, and applying it here is an extrapolation, not a replication.

What the paper does establish, and what carries over regardless of geography, is the shape of the game: a source wins a place in the answer by being citable, meaning structured, corroborated, and backed by the kind of authority signals an engine can verify quickly. That is a different, and in some ways higher, bar than simply existing online.

The structural case that aggregators start ahead

If citability is the currency, then scale, content velocity, and inbound-link density become direct advantages, and those are exactly the assets a large marketplace accumulates by default and an individual MSME's often-thin website usually lacks. Published data on which domains actually get cited in AI answers, while general-web rather than category-specific, shows how concentrated that advantage already is elsewhere. Ahrefs' Brand Radar analysis of roughly 76.7 million Google AI Overviews and about 950,000 prompts each on ChatGPT and Perplexity for June 2025 found Wikipedia cited in 16.3 percent of the ChatGPT answers sampled and 12.5 percent of Perplexity answers, with YouTube cited in 16.1 percent of Perplexity answers and 9.5 percent of AI Overviews, a handful of very large domains capturing a disproportionate share of citations across unrelated queries.

That pattern is not India-specific and it is not about B2B or local-service commerce, so it cannot be read as proof about IndiaMART, JustDial, or Sulekha. But it is the same structural signature, constant fresh content, heavy internal and external linking, high domain trust, that a national marketplace or local-search aggregator would be expected to carry into its own category if the pattern repeats there. The open question is whether it does.

ONDC was built to counter this, one layer down

India's own policy answer to a related concentration problem already exists, and it is worth reading precisely because it does not settle the question this article is asking. The Open Network for Digital Commerce (ONDC) is explicitly designed, per a Ministry of Commerce & Industry reply in the Lok Sabha, so that "seller-side apps make their full catalogues discoverable to all buyer-side apps, while buyer-side apps disclose key parameters used for sorting or listing search results, enabling sellers to understand and improve their ranking," with the stated goal that sellers are visible "regardless of size, scale or digital sophistication." As of 9 December 2025, more than 1.16 lakh retail sellers were live on the network across over 630 cities and towns, per the same government reply.

The network has since scaled further. Per a Department for Promotion of Industry and Internal Trade written reply in the Rajya Sabha on 7 August 2026, ONDC grew from 0.2 million transactions in FY23 to 218 million in FY26, crossing 500 million cumulative transactions in July 2026, with more than 2 lakh retail merchants and over 10 lakh mobility and logistics service providers active on the network by that point.

That is a genuine, code-level commitment to non-discriminatory discovery, and it directly addresses concentration risk inside apps built on the ONDC protocol. But it is a different layer from the one this article is about. ONDC's equal-visibility rules bind buyer-side and seller-side apps that speak the ONDC protocol. ChatGPT, Perplexity, and Google's AI Overviews are not ONDC buyer apps. Nothing in the public ONDC specification requires or even addresses how a general-purpose AI assistant should weigh an ONDC-registered seller against an IndiaMART listing or a JustDial profile when it synthesizes a natural-language answer to an open question. ONDC solving discovery fairness inside its own network, if it does, is not evidence about what happens in the separate, ungoverned layer where a generative engine decides whom to cite.

What a real test would look like

Because no public dataset currently applies a citation-share method to Indian MSME discovery queries, the next step is to describe the test rather than assert a result for it. A defensible version would follow the same logic as GEO-bench: assemble a representative sample of buyer-style queries across MSME-relevant verticals and cities, the kind an actual buyer or consumer would type, such as a category plus a location, rather than a brand name. Run each query across multiple engines, ChatGPT, Perplexity, Google's AI Overviews, and Gemini, since the Ahrefs data above shows citation behavior already differs meaningfully between them. Code every citation that appears by domain type: national aggregator (IndiaMART, JustDial, Sulekha, and comparable platforms), an ONDC-linked buyer-app storefront, an individual business's own website or Google Business Profile, or another category entirely. Compute the share of citations landing on each type, and repeat the run over time, since answer-engine behavior changes as the underlying models and their retrieval layers are updated.

That is a method, not a finding. Until it, or something like it, is actually run against Indian queries, any specific percentage claiming to show how often aggregators out-cite individual MSMEs in AI answers should be treated as unverified, regardless of how confidently it is stated.

What the evidence supports

Two things can be true at once. The general pattern in published citation data, a small number of large, heavily linked domains capturing outsized citation share, is real and independently documented, and GEO's own findings show that citable, corroborated authority is what moves visibility inside a generated answer, not simply being online. IndiaMART and JustDial, and by its own description Sulekha, carry exactly the kind of scale and content velocity that pattern rewards elsewhere, and the population they aggregate, more than 78 million registered MSMEs, most still building basic transactional internet habits per ICRIER's own 2025 survey, is the population least equipped to counter a citation disadvantage it cannot measure.

What is not established is the number itself. ONDC shows that India's policy apparatus has already built a serious, code-enforced answer to a related fairness problem, but at the marketplace-protocol layer, not the generative-answer layer this article is about, and the two should not be conflated. Until the query-sampling method above is actually run against Indian MSME categories, aggregators structurally out-citing individual businesses in AI answers remains a hypothesis consistent with the surrounding evidence, not a measured fact, and the only way to know where a specific business stands today is to check it directly rather than assume the pattern from adjacent data.