AI visibility for Indian brands · 4 engines · 12 languages · metro and Bharat
Your buyer asks in Hindi. The engine answers from English.
ChatGPT, Perplexity, Gemini and Claude name one or two brands when an Indian buyer asks a buying question. Rovoki measures how often that brand is yours, in 12 languages and 10 Indian cities, and shows you the sources every answer was built from.
An answer in Hindi is not a translation of the answer in English.
Ask a question in Hindi and the engine writes back fluent Hindi. What it read to get there was mostly not Hindi. It reaches for English sources, answers from those, and names the brands those sources name.
This is a retrieval problem, not a fluency problem. Which means it is not fixed by translating your website.
Answers to Hindi questions.
The same measurement for every other language region tested.
An Indian brand is being described to Hindi-speaking buyers at roughly double the error rate, out of sources it has never thought about, in a language it did not choose.
The study, and what it could not answer
A Stanford-led team put 2,100 questions to six commercial chatbots across six language regions in February 2026, producing 12,600 model-question pairs. For Hindi questions, English Wikipedia was the single most-cited domain at 3.1%, ahead of every Hindi-language publication. Of the top Indian sources, only one was primarily Hindi-language. The authors call the pattern an Anglophone retrieval pivot.
The limitation they flag themselves: every query was issued from servers in the United States, which may have made English sources more likely to be returned. Nobody has yet run the same test from Indian infrastructure. Until someone does, the direction is solid and the size of the effect is not settled.
Source: Suzgun et al., evaluating commercial AI chatbots as news intermediaries, February 2026
We publish the limitation because it is the honest reading, and because it is the next thing we intend to measure. Rovoki asks from India.
Two buyers. One category. Different brands named.
"best UPI app for a small business?"
For a small business, most owners start with Paylo01 for the instant settlement, and add a soundbox once volumes pick up.
"छोटे बिज़नेस के लिए सबसे अच्छा UPI ऐप कौन सा है?"
छोटे कारोबार के लिए ज़्यादातर दुकानदार RupeeOne01 चुनते हैं, क्योंकि सेटअप आसान है और सेटलमेंट तुरंत होता है।
12 of the 31 languages Rovoki measures are used in India. Every one is a separate question, asked separately, with the engine’s search filtered to that language. We record what came back and what it cited, in that language, and we do not translate the answer before scoring it.
Hinglish is on that list for a reason. It is a real query behaviour rather than a separate language: Hinglish queries return Hindi-language sources, so we file it under Hindi and report it on its own line. Half of urban India’s buying questions look like this and no other instrument in the category asks them at all.
You are being described to your buyers by pages you have never read.
A national average hides two different results.
An AI answer is always personalised to somewhere. There is no view from nowhere, which is why Rovoki asks from a place rather than pretending location does not exist.
For India we ask from 10 cities: Mumbai, Delhi NCR, Bengaluru, Pune, Nagpur, Chennai, Hyderabad, Kolkata, Ahmedabad, Jaipur. Every number in the report is then cut twice, metro and Bharat, with the answer count printed on each side so you can see how much sample each one is standing on.
Where the two sides agree, you have one visibility problem. Where they do not, you have been running one strategy against two markets.
Illustrative figures for acme.co, a brand that does not exist
Same brand, same category, same week. The gap is the whole point of measuring separately.
The scope, printed on every score.
Engines
ChatGPT, Perplexity, Gemini and Claude, with the citations attached to each answer. Google AI Overviews is next.
Languages
English, Hindi and Hinglish, plus Marathi, Tamil, Telugu, Kannada, Bengali, Gujarati, Malayalam, Punjabi and Urdu. Pick the ones your buyers actually use. A brief that does not need Tamil should not be sold Tamil.
Cities
10, or India as a whole with no city pinned. The metro and Bharat split is computed when you pick cities on both sides of it, and is left out rather than computed wrongly when you do not.
Prompts
Buyer questions in your category, not queries about your brand name. “Best online pharmacy for same-day delivery” is a prompt. “Is Acme good” is not, and it is how most tools flatter their customers.
What returns
A Visibility Standing out of 100, engine by engine and language by language, the prompts you lose with the brand that wins them, the publications AI reads for your category sorted by desk, and a ranked fix list with predicted point gains. Predicted, never promised.
Sampling
Each question is asked once per engine per run today, and we do not yet publish a confidence band. That makes a single run directional rather than conclusive, and we would rather write it here than have you find it in month two.
Raw answers
Every number opens into the text it came from, with a timestamp and the sources that answer cited. Nothing is summarised away.
Walk in with the number already measured.
You already own the client relationship and the fix. What you do not own is a defensible number. That is the part we run.
Measure before you pitch
Not a proposal to find out what AI says about them. The answer, in their categories, in their buyers’ languages, before the meeting.
White label it
Your logo, your colours, your domain. The report says prepared by you.
Defend it in the room
You keep the method: the losing prompts, who wins them and in which language, the publications AI reads sorted by desk, and the fix list with predicted gains. Including the months a number goes down, which is the only kind of number a client eventually trusts.
Bill for the language work
English, Hindi, Hinglish and nine more Indian languages are a line item you can quote, deliver and re-measure. Almost nobody in this category can produce it, in India or anywhere else.
What we don’t do.
No ranking
We don’t guarantee a ranking. Nobody honestly can. Anyone who does is selling you a number they don’t control.
No position
We don’t report a position. Ask the same question twice and you get two answers. A rank implies a stability that does not exist.
No phone screen
We don’t claim to see one person’s phone. We collect through grounded API calls with India set as the location, not by screenshotting a consumer app. Both methods are partial, and ours is the one that can be repeated, dated and shown to you.
Nothing hidden
We don’t hide the raw answers. Every number opens into the text it came from, with a timestamp.
Plain answers.
Do you actually ask in Hindi, or translate an English answer?
We ask the question in Hindi and the engine's search is filtered to Hindi. The answer is scored in the language it came back in, and the sources it cited are recorded as they were cited. Translating first would measure the translation.
Is Hinglish a real thing to measure?
Yes, and it behaves differently from both of its parents. It is Hindi written in Latin script, and those queries return Hindi-language sources, so we file it under Hindi and report it on its own line.
Which engines?
ChatGPT, Perplexity, Gemini and Claude today, with citations. Google AI Overviews is next.
Is Rovoki only for India?
No. Rovoki measures 18 countries and 31 languages, or worldwide with no country pinned. This page is the India view of it, because India is the market where the language gap is widest and the fewest people are measuring it.
We already do SEO. Why this?
SEO earns you a position on a page of links. An AI answer names one or two brands and moves on. Being fourth in search still gets clicks. Being unnamed in an answer gets nothing.
Why does the Hindi score differ from the English score for the same brand?
Because the engine read different sources to write each answer. That gap is the measurement, not an error in it. The report shows which sources each side was built from.
Do you guarantee a ranking?
No, and nobody honestly can. We show predicted point gains, make the changes measurable, and re-measure. Every gain or loss lands on the record.
Measure one brand, in the languages your buyers use.
Pick your category, pick your languages, pick your cities. You get the answers back with the sources attached.
Four engines · 12 Indian languages · metro and Bharat