100 Million Weekly Users, How Many Local Answers? Testing ChatGPT's India Scale Against Its India Depth

What Altman actually disclosed, and what it does not cover

Ahead of the five-day India AI Impact Summit that opened in New Delhi on February 16, 2026, Sam Altman told reporters that India had crossed 100 million weekly active ChatGPT users, making it OpenAI's second-largest market after the United States. He added that India has the largest number of student users of ChatGPT of any country. The disclosure landed alongside a broader signal: ChatGPT's worldwide usage had reached 800 million weekly active users as of October 2025 and was reported to be approaching 900 million, which puts India at roughly one-eighth to one-ninth of the platform's entire global weekly base.

The context matters. OpenAI opened a New Delhi office in August 2025, launched a sub-five-dollar "ChatGPT Go" tier for price-sensitive Indian users that same month, and made that tier free for a year in October. Those are deliberate market-entry moves, and by the count that OpenAI itself tracks, they worked: a follow-up TechCrunch report put India at roughly 180 million monthly active users in January 2026, with ChatGPT accounting for more than 60 percent of the country's GenAI in-app revenue. The same report noted a harder number sitting underneath the growth story: India generates around a fifth of the world's GenAI app downloads but only about 1 percent of in-app purchases, and ChatGPT's India revenue actually fell month over month after the free tier launched.

All of that is real, sourced, and worth taking seriously as evidence of adoption. None of it is a measurement of answer quality. Weekly active users, monthly active users, downloads, and in-app revenue are engagement and monetization metrics. They describe how many people show up and how much they pay. They do not describe what the model says back to any one of them when the question is "who sells this near me."

Reach and depth are different measurements

It is tempting to read a headline like "100 million weekly users" as a proxy for capability, on the assumption that a product used by that many people must be good at most of what people ask it. That inference does not hold for a specific, commercially important class of question: local business discovery. A model can be broadly useful for writing, coding, tutoring, and general knowledge, the categories that plausibly explain India's scale and its outsized student user base, while still being thin, generic, or simply wrong on "which tailor in Indiranagar does same-day alterations" or "which Udyam-registered supplier in Surat ships GSM-graded fabric."

The industry term for the second question is share of answer: not whether an engine returns results, but whether it names a specific, real, correctly described business inside the answer it generates. A 2024 peer-reviewed study, "GEO: Generative Engine Optimization," established that this is a measurable and manipulable construct. The researchers found that adding cited statistics, direct quotations, and citations to authoritative sources measurably raised a source's visibility inside the answers of the generative engines they tested. That result matters here for a narrow but important reason: it proves answer-engine visibility is not random or unknowable. It responds to specific, identifiable properties of how a business is described online. It also means the absence of those properties, thin, uncorroborated, inconsistent business listings, plausibly suppresses visibility in the same systematic way.

A second data point sharpens why this stakes matter more than it used to. Pew Research Center's July 2025 browsing-panel study found that users clicked a traditional search result in about 8 percent of searches when a Google AI summary was present, against 15 percent when it was not, and clicked a link inside the summary itself in only about 1 percent of cases. Whatever a generative engine chooses to name inside its answer increasingly is the outcome for the businesses involved, not one signal among several that a human will still click through and compare.

What we do not have: an audited, India-specific local-answer number

This is a real gap. No academic, standards-body, or government dataset currently benchmarks how often ChatGPT, or any single generative engine, correctly names a real Indian MSME in response to a local commercial query, broken out by city tier, language, or business category. That number does not exist in public form. What circulates instead is a widely cited industry-blog figure claiming ChatGPT recommends local businesses in only about 1.2 percent of relevant queries, against roughly 35.9 percent for Google's local search surfaces. That comparison is not India-specific, was not produced by a peer-reviewed process or an audited methodology, and should be read as directional industry monitoring, not a settled measurement. It is included here for the same reason a compass is useful without being a map: it points toward the likely shape of the gap without fixing its size.

The evidence that does exist, and what it implies without proving

Absent a direct measurement, the surrounding evidence is still informative, as long as it is read for what it is: context, not a substitute finding. On the supply side, India's Ministry of MSME told Parliament that more than 7.83 crore enterprises were registered on the Udyam Registration Portal as of its most recent Rajya Sabha disclosure, a large and growing base of exactly the kind of business a local query would need to surface. On the readiness side, an independent survey of 2,882 enterprises conducted by Kantar for the Research and Information System for Developing Countries, funded by the Mastercard Center for Inclusive Growth, found that only 19 percent of MSMEs reported having a website, only 29 percent used social media for marketing, and only 18 percent were engaged in e-commerce at all. A business without a website or a maintained public listing is not absent from the world; it is close to absent from the corpus a generative engine reads to decide what to say.

The Indian government has itself flagged the risk in writing. In a Lok Sabha reply reported in December 2025, the Commerce Ministry stated that AI-powered agentic shopping "poses challenges for retailers, particularly those not integrated with major e-commerce platforms, risking reduced visibility and market access." That is a policy body naming the same structural concern this piece is raising, using its own regulatory language rather than marketing framing. None of these facts measure ChatGPT's India-specific answer accuracy directly. Together they describe a plausible mechanism for a gap between scale and depth: a large, real business base that is thinly represented in the exact machine-readable form that answer-engine visibility now depends on.

India already built a machine-readable layer for this problem

The more pointed part of the picture is that India did not wait for generative AI to try to solve local business discoverability. The Open Network for Digital Commerce, launched in 2023, is a government-backed protocol designed explicitly to make sellers discoverable across any compliant buyer app rather than locked inside one platform's search box. As of a government statement reported in December 2025, more than 1.16 lakh retail sellers were live on ONDC across more than 630 cities and towns. The underlying Beckn Protocol that ONDC runs on describes its own catalog, item, and provider schemas as authored, in the schema registry's own words, "with AI, for AI," with every field optimized for machine navigability and agent consumption, not for a human scrolling a page.

That is, in structural terms, exactly the kind of machine-legible, standardized business data that the GEO research suggests a generative engine should find easier to cite. Whether ChatGPT's answers, in practice, draw on ONDC-listed inventory when an Indian buyer asks a local commercial question is, again, not something this piece has tested or can honestly claim to know. It is an open and answerable question, and arguably the single most important one for anyone deciding whether India's public commerce infrastructure functions as a bridge into AI answers or sits, for now, disconnected from them.

What a national number cannot tell a single business

Even a completed, published, India-specific share-of-answer study would report an average, and an average across 7.83 crore registered enterprises hides the only number that actually matters to the owner of one of them: whether their business, specifically, is the one a generative engine names, and whether what it says is correct. A national adoption headline like "100 million weekly users" cannot answer that question for any individual firm, and neither, honestly, could a national share-of-answer study once one exists. Both are population-level reads. The gap they leave is instance-level, and it has to be closed instance by instance.

That is the practical use of everything above: not a verdict on ChatGPT's India performance, which nobody has published, but a case for treating "we're on ChatGPT's radar because it's huge here" as an assumption worth checking rather than a fact worth repeating.