22 Scheduled Languages, One Answer Engine: Mapping Where Google's AI Overviews Actually Work in India

Two features, one country, a fast-moving language list

Google runs two distinct AI answer surfaces in Search: AI Overviews, the automatically generated summary box that appears above organic results, and AI Mode, a separate, more conversational search experience the company launched in India in June 2025. Both draw on Gemini models, both decide which sources to cite, and both have expanded their language support in stages rather than all at once. Treating "Google's AI answers" as a single, fixed thing hides the fact that language coverage for each feature has been a moving target for two years, and that the target is still moving.

That matters for a country with 22 languages listed in the Eighth Schedule of the Constitution, spoken by populations ranging from Hindi's 528 million native speakers down to the roughly 25,000 people who report Sanskrit as a mother tongue. A feature that "supports Indian languages" in a press release can mean one language or thirteen, and the difference changes who receives an AI Overviews answer at all.

August 2024: English, and a Hindi toggle

AI Overviews reached India on August 15, 2024, alongside the UK, Japan, Indonesia, Mexico, and Brazil, as part of Google's first expansion of the feature outside the United States. The India-specific addition was a toggle letting users switch the summary between English and Hindi, plus a text-to-speech option. No other scheduled language shipped at launch.

Two weeks later, TechCrunch reporter Ivan Mehta tested that Hindi rollout directly and published what he found. Reordering the words in an identical Hindi question sometimes produced no answer at all. A question about summer foods returned "Chiknai wali cheezien" ("sticky things") in Hindi, where the English version correctly suggested "oily." A question about YouTube's leadership returned a wrong tense on a basic fact, an incorrect claim that a past CEO still held the role, while the English answer was correct. A query about food near a Delhi gurudwara summarized a source article but dropped the detail that snacks were sold outside the place of worship, not inside it. Google told TechCrunch its accuracy rate for AI Overviews was "on par with other features like Featured Snippets," while acknowledging errors happen when its systems misinterpret web content. The point is not that Hindi was uniquely broken; it is that the single Indic language Google chose to ship first still carried the kinds of translation and context failures that erode trust in a local-business answer.

What "the language list" actually looked like, quarter by quarter

Google's own blog posts let the coverage timeline be reconstructed with dates rather than impressions. In October 2024, when the company announced its largest AI Overviews expansion to that point, more than 100 countries, it also published its full global language list: English, Hindi, Indonesian, Japanese, Portuguese, and Spanish. Six languages worldwide, one of them Indic. At that point, Hindi was not one of several Indian languages Google supported; it was the only one.

The next confirmed addition came at Google I/O in May 2025, when the company said AI Overviews had passed 200 countries and 40 languages, explicitly naming Arabic, Chinese, Malay, and Urdu among the new additions. Urdu became the second scheduled Indian language with dedicated AI Overviews coverage, roughly nine months after Hindi.

AI Mode's expansion ran on its own separate clock. It launched in India in English only in June 2025. In September 2025, TechCrunch reported Google added five new AI Mode languages, including Hindi, after more than six months of English-only operation. One month later, TechCrunch reported a further seven Indian languages, Bengali, Kannada, Malayalam, Marathi, Tamil, Telugu, and Urdu, arriving alongside the India launch of Search Live, as part of a global push past 35 new languages.

The map today: 13 covered, 9 not, roughly 51 million people in between

Reading Google's current support documentation directly, rather than the announcement blog posts, gives the clearest present-day picture. Google's support pages for both AI Overviews and AI Mode list the same set of Indian languages as supported: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Sindhi, Tamil, Telugu, and Urdu. That is 13 of the 22 scheduled languages, a large jump from the single language covered in October 2024, and a genuine expansion worth acknowledging rather than dismissing.

The remaining nine are Assamese, Bodo, Dogri, Kashmiri, Konkani, Maithili, Meitei (also called Manipuri), Sanskrit, and Santali. None appears on either support page as of this writing. Using the 2011 Census figures for native speakers of each, Assamese at roughly 15.3 million, Maithili at roughly 13.6 million, Santali at roughly 7.4 million, Kashmiri at roughly 6.8 million, Dogri at roughly 2.6 million, Konkani at roughly 2.3 million, Meitei at roughly 1.8 million, Bodo at roughly 1.5 million, and Sanskrit at roughly 25,000, the nine uncovered languages sum to approximately 51 million native speakers. That sum is this publication's own addition of the individually sourced Census figures; Google does not publish a population-weighted coverage number, and neither does any Census document.

A qualifier belongs here. "Supported" on Google's documentation describes which languages the interface will generate an answer in; it says nothing about whether that answer is accurate, complete, or consistent for a given query, the exact gap the 2024 Hindi testing exposed. Coverage and quality are two different claims, and only the first one is something Google's own pages currently let an outside reader verify.

Why this line falls hardest on MSMEs specifically

India's internet population is not converging on English. The Internet and Mobile Association of India and Kantar's "Internet in India 2024" report put active internet users at 886 million in 2024, with 98% of users accessing content in Indic languages and 57% of urban users, not just rural ones, also preferring regional-language content over English. A small business searching for a supplier, a customer searching for a shop, and a first-time internet user in a district town checking a government scheme are all more likely to phrase that query in a scheduled language than in English.

The Ministry of MSME reported more than 7.83 crore (78.3 million) enterprises registered on the Udyam and Udyam Assist platforms as of February 2026. Those firms are not concentrated in metros where English-language search habits are strongest; a large share operate in the same districts where Assamese, Maithili, Santali, and Kashmiri are the language of daily commerce. For a business in that position, whether its category of query even generates an AI Overview or AI Mode answer in the customer's own language is not a hypothetical UX detail. It is the difference between showing up in the answer and never being read in the first place.

There is also a supply-side constraint sitting underneath Google's feature coverage that a support-page checkmark cannot fix. W3Techs, which tracks the declared content language of the sites in its index, finds Hindi used as the on-page language of fewer than 0.1% of tracked websites, this despite Hindi being the single most AI-answer-covered Indian language and the native tongue of over 40% of the country. An answer engine can only surface and cite what the underlying web actually contains in that language, and for most scheduled languages, that underlying corpus is thin regardless of what the interface supports.

What the model evidence says about the gap underneath the interface

AI4Bharat's MILU benchmark, published in November 2024 and later accepted at NAACL 2025, offers the closest thing to an independent, methodologically documented read on how large language models actually perform in Indic languages, separate from any single company's feature rollout. Testing over 45 models across 11 Indic languages and 42 subjects drawn from real regional and national exams, the study found the best-performing model tested, GPT-4o, reached only 72% average accuracy. Every other model scored lower. The paper's authors report two findings relevant here: models perform measurably better in higher-resource languages than lower-resource ones, and open multilingual models outperform language-specific fine-tuned models, which score only slightly better than random guessing.

MILU does not test Google's AI Overviews or AI Mode directly, and its multiple-choice exam format is not the same task as answering a local-business search query. What it does establish, independently of any single company's claims, is that the underlying capability gap between English and Indic-language understanding in large language models is real, measured, and uneven across languages, which is consistent with, though not proof of, the specific translation and context errors TechCrunch documented in Hindi AI Overviews.

What is not yet measured, and how to read this

No public, audited dataset currently tests identical local-business queries, the kind an MSME's own customer would type, across all 13 languages Google currently lists as supported, scoring whether an answer appears at all, whether it names specific businesses, and whether its content matches an English-language or source-document baseline for accuracy. The 2024 Hindi testing that TechCrunch published is the closest public precedent, and it covered one language at one point in time, before six additional languages were added to either feature.

A defensible method for that measurement exists and is worth stating plainly rather than skipping past: a fixed panel of local-business query templates (for example, "[service] near me," a government-scheme eligibility question, a wholesale-supplier question), run in each of the 13 currently supported languages across several cities, with each response scored for presence, named-entity citation, and factual accuracy against a source-language control. Until that kind of panel exists publicly, any claim about which of the 13 supported languages actually perform well today, beyond the single documented Hindi case from 2024, should be treated as an open question, not a settled finding.

What can be stated without qualification: coverage expanded from one scheduled language to 13 in about eighteen months, a real and fast pace of change; nine languages spoken by roughly 51 million people remain outside either feature entirely; and the one language with a public, independent quality test on record showed material translation and accuracy problems at the point it was the only Indic language available. Those three facts do not require an original study to state. What they do not yet tell us, and what no one has publicly measured, is how the newer nine languages perform today.