The Vernacular Answer Gap: Why 900 Million Indians Search in Indic Languages While AI Engines Answer in English

Two facts that do not sit comfortably together

The first fact is about demand. India's internet population has crossed 900 million, and it is overwhelmingly a vernacular population. The second fact is about supply. The systems that now read the web and hand back a single synthesized answer are, by their own builders' admission and by independent measurement, at their strongest in English. Put those two facts in the same room and a gap appears between the language a market searches in and the language a machine answers best in.

That gap matters more than it used to, because the interface changed. When search returned ten blue links, a weak Indic-language result still sat on a page a person could scan, and the person could switch to English, reword the query, or scroll. When an answer engine returns one paragraph and names one or two businesses inside it, the quality of that single Indic-language answer decides what the searcher sees. There is less room to recover from a thin or mistranslated response, because the response is the destination, not a list of options.

No public dataset yet measures the specific case at issue, how ChatGPT, Gemini, and Perplexity answer local-business and MSME-relevant queries across India's scheduled languages. That test does not exist in the open literature. What does exist is a body of peer-reviewed and government evidence about the two ends of the gap: how India searches, and how well these models perform in Indic languages generally. This article reads across that evidence and is explicit about where measurement stops and inference begins.

How India actually searches

The demand side is well documented. The IAMAI-Kantar Internet in India 2024 report counted 886 million active internet users in 2024, growing about 8 percent year on year, and projected the base past 900 million in 2025. Rural India, at 488 million users, is now the larger half at 55 percent of the total, and it is growing at roughly double the urban rate. This is not a metro, English-first internet. It is a national one.

Language use follows from that. By the same report, nearly all internet users, about 98 percent, access content in Indic languages, and 57 percent of even urban users prefer content in a regional language. The census numbers explain why. In the 2011 Census, roughly 128 million Indians, about 10 percent, reported any ability in English, and only a fraction of a percent spoke it as a first language, against 528 million native Hindi speakers and hundreds of millions more across Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Odia, Punjabi, and the rest of the 22 languages in the Constitution's Eighth Schedule.

The policy layer has moved to match this. Bhashini, the government's National Language Technology Mission under the Ministry of Electronics and Information Technology, was launched in 2022 to build translation and speech models as public infrastructure across all 22 scheduled languages, and has published more than a thousand models on its platform. The direction of travel is unambiguous: the Indian internet is becoming more vernacular, not less, and the state is investing to make it so.

What the models actually do in Indic languages

The supply side is where the gap is measured. In 2024, a Google Research team published IndicGenBench at the ACL conference, a human-curated benchmark spanning 29 Indic languages across 13 scripts and 4 language families, covering tasks like cross-lingual summarization, translation, and question answering. Its headline finding was blunt: even the strongest model tested showed a significant performance gap in every Indic language compared with English, and the authors concluded that further work is needed before multilingual models can be called inclusive.

A 2025 study from a group at the Pune Institute of Computer Technology and IIT Madras, pointedly titled Better To Ask in English?, put the question to newer models. Using the IndicQuest dataset of 4,000 fact-based question-answer pairs across English and 19 Indic languages, the authors evaluated GPT-4o, Gemma-2, and Llama-3.1 on region-specific questions about Indian history, geography, politics, literature, and economics. Their conclusion is the one that matters for this piece: English consistently outperformed the Indic languages across the models and metrics tested, even on questions rooted in Indian context, and the models hallucinated more often when answering in low-resource Indic languages.

The visible symptom: AI answers in Hindi

The benchmark findings show up in shipping products. When Google brought AI Overviews to Hindi in India in 2024, TechCrunch documented the quality problems directly. Asked in Hindi what foods to eat in summer, one overview returned chiknai wali cheezien, literally sticky or greasy things, where the English answer said oily. Asked about YouTube's leadership, the Hindi answer garbled the tense and stated that Neal Mohan was Google's CEO until a date, while the English version correctly said he is the CEO as of that date. Small rewordings of the same Hindi question produced an answer in one form and nothing in another.

The reporting also named the mechanism, and it is the same one the academic work points to. As TechCrunch put it, results for similar questions in English were much better than the Hindi ones, partly because there are more and better sources available in that language. The engine is only as good as the corroborated material it can read, and for most topics there is far more of that material in English than in any Indic language. The model does not fail because it dislikes Hindi; it fails because the web it reads is thin in Hindi on that topic.

Why this lands hardest on MSMEs

A large enterprise can compensate for a weak Indic-language web. It can publish authoritative content in several languages, earn coverage that gets cited, and maintain the consistent, structured identity that answer engines reward. Most micro, small, and medium enterprises cannot. Their online presence is often a single listing, a social profile, and a handful of reviews, frequently in a mix of languages and rarely structured for a machine to parse. When a customer asks an engine in Marathi or Tamil for a recommendation, the business that gets named is the one the machine can read and corroborate in that language, and that is usually not the small local firm.

This is where the vernacular answer gap becomes a machine-readiness problem rather than a linguistics problem. The question is not whether an MSME's customers speak the language; it is whether the business's identity, offering, and credibility exist in a form that an engine can find and trust when it answers in that language. The corroboration an answer engine looks for, consistent details across sources, third-party mentions, structured data, reviews, is scarcer in Indic-language form for small firms, so even a firm whose customers are entirely vernacular can be invisible to the vernacular answer.

The stakes are rising because commerce itself is going vernacular. India's Open Network for Digital Commerce, ONDC, built on the open Beckn protocol to bring kirana stores and small sellers onto a shared network, had onboarded hundreds of thousands of sellers and was processing more than twelve million transactions a month by early 2025, having crossed 200 million cumulative transactions. As discovery for these sellers moves toward voice and vernacular interfaces, the ability to be surfaced correctly in an Indic-language answer stops being a marketing nicety and becomes a condition of being found at all.

There is a compounding effect worth naming. A small firm that is thinly represented in its customers' language generates little corroborated Indic-language material about itself, which gives the engines less to read, which makes them likelier to omit or misdescribe it, which in turn produces fewer vernacular mentions of the firm. The under-represented stay under-represented. Breaking that loop does not require a large budget so much as deliberate structure: a clear, consistent business identity, accurate details repeated across the places an engine checks, and content in the languages the firm's customers actually use rather than only in English.

What we can claim, and what we cannot

Three things here are well established. India searches overwhelmingly in Indic languages, by the IAMAI-Kantar and census records. Generative models perform measurably worse in Indic languages than in English, by IndicGenBench and the IndicQuest factual-accuracy study. And that weakness is visible in shipped products, by the AI Overviews reporting. None of those three is speculative.

One thing is not yet established. There is no public, audited dataset that measures how ChatGPT, Gemini, and Perplexity specifically answer local-business and MSME-relevant queries across India's scheduled languages, or how often a small firm that should be recommended is omitted or misnamed in a vernacular answer. That is an open question. It could be measured, by running a matched set of buyer-intent queries in English and in each target Indic language across the major engines and scoring the answers for accuracy, completeness, and correct business attribution, but until someone runs that study at scale, the per-engine MSME figure does not exist and no one should quote one.

What the surrounding evidence supports is a directional conclusion, not a precise number: the general Indic-language performance gap almost certainly extends into local and commercial queries, and the effect is worst in the least-resourced languages. That is enough to act on without inventing a statistic. The right response is to measure a specific business against the specific languages and surfaces its customers use, rather than to assume either that vernacular AI already works or that it is hopeless.