AI search visibility in India, measured: 420 readings of ChatGPT and Gemini on 16 October 2026, and the usable count behind every rate
210 Indian buying questions put to two named engines produced 420 India readings and 420 usable ones. The usable observation count is the number to check first in any AI visibility benchmark, including this one.
Published by AI Knows Us · Updated 16 October 2026
Two hundred and ten Indian buying questions went to ChatGPT on 16 October 2026, and the same 210 went to Gemini. Nothing came back unreadable: 420 India readings attempted and 420 counted. Across all three markets in the same run that is 630 questions, 1,260 readings and 1,260 usable. The reason the usable count leads this page rather than a mention rate is that the usable count is the only figure that tells you whether the rate underneath it means anything. A benchmark that used 597 of 612 attempted readings is making a different claim from one that used 1,260 of 1,260, and almost no published AI visibility benchmark tells you which of the two it is. Ours does, in both directions: the run before this one, on 15 October 2026, attempted 612 readings and used 597, and the 15 it lost were not spread evenly.
Three words that are usually treated as one
Before any count means anything, three things have to be kept apart, and the same engine answer can be all three or none of them.
- Mentioned. The name appears in the text of the answer. No link, no source, no endorsement implied.
- Cited. A page on that domain is one of the sources the answer was built from. This is what we count, and the definition is strict: one domain in one reading, deduped.
- Recommended. The answer puts the business forward as a choice for the buyer. A business can be cited as a source and never recommended, and it can be recommended and never cited.
Every figure on this page is a citation count under that definition. Where a benchmark does not say which of the three it counted, the number cannot be compared to any other number.
The India readings on 16 October 2026
| What was measured | India |
|---|---|
| Questions put to each engine | 210, in 21 batches |
| India readings recorded | 420 in all, split evenly between the two engines |
| Usable India readings | 420 of 420 |
| Readings citing our own domain, ChatGPT | 2 of 210 |
| Readings citing our own domain, Gemini | 0 of 210 |
| Engines read | ChatGPT and Gemini. Two, not five. |
Across all three markets the same run holds 1,260 readings, 20,143 citations and 4,947 distinct domains, with our own domain in 6 of the readings, which is 0.5%. Of those citations, 82.0% went to a page on a vendor's or an agency's own website and 0.7% to a review directory.
Why the usable count is the number to check first
On 15 October 2026 we ran a different question set, 102 questions per market, attempted 612 readings and used 597. Its question set, its controls and its own source breakdown sit together in what AI assistants actually cite across 597 readings of ChatGPT and Gemini in three markets. The 15 readings that could not be used did not scatter. All 15 landed in a single market and engine cell, and the India slice of that benchmark names it and prints the counts around it.
That is the general lesson and it survives the specific case. Missing responses in a benchmark are rarely random. They cluster on one engine, in one market, or on one kind of question, so a benchmark that reports a single usable total without saying where the losses sat has hidden the part most likely to matter. The practical rule for reading anybody's benchmark, including this one: find the attempted count, find the usable count, and ask where the difference went. If the page does not say, the rates on it are unverifiable rather than wrong, and unverifiable is enough reason to leave it out of a plan.
The two runs agree, which matters more than either one
The second question set was written to attack the first one's blind spots across eight directions it had never taken, from procurement depth to questions asked before the buyer knows the category has a name. If the first result had been an artefact of wording, this was the run most likely to break it.
| Figure | 15 October 2026 | 16 October 2026 |
|---|---|---|
| Usable readings | 597 of 612 attempted | 1,260 of 1,260 |
| Vendor or agency own site | 85.6% of citations | 82.0% of citations |
| Review directories | 0.8% of citations | 0.7% of citations |
| Readings citing our own domain | 3 of 597 | 6 of 1,260 |
The second set went after what the first one could not see, on twice the sample, and the columns above agree anyway. An agreement reached that way is itself a finding, and it is the strongest thing in either dataset.
How these India figures were produced
The questions were Indian buying questions, written as an Indian buyer would type them, with India named in the question text rather than set as a parameter. That control matters: the surfaces we read accept a country parameter, return success, and then answer for somewhere else, so the market was locked in the words and then checked on the answers. No question named our own business, because a question that names the business measures recall rather than recommendation. Every cited URL was recorded, and source classification runs through one tested file that refuses to load if any of 24 named examples sorts into the wrong category. For the run before this one, how the question set and the 597 readings were built walks through every control and every way the measurement failed.
One scoring rule is worth stating because it changes the size of every count. A citation here is one domain in one reading, deduped, so the same domain cited four times inside a single answer counts once. Loosen that to every URL occurrence and the total for this identical data is a great deal larger, which is a different measurement wearing the same word rather than a correction of this one. 712 cited URLs in this run were the data provider own fetch infrastructure, an iframe proxy and a mirror of Google documentation, and they are resolved back to the page they serve before counting. A proxy hostname is never listed as a publisher.
What this India dataset does not contain
Nothing in it breaks down by trade. The questions that went after particular trades named no trade, so any software, healthcare or legal figure on this page would be an invention. Which parts of a search programme actually change with the trade is set out in search by industry. City level Indian figures are not in it either; those come from a separate scan that read ten named cities one at a time. The engines here are ChatGPT and Gemini, counted apart. Nothing on this page reads on Google AI Overviews, on Perplexity or on Claude, none of which saw these questions, and the five surfaces we cover for customers are never added into one score.
Questions about the India readings and their denominators
How many prompts are enough for an Indian benchmark?
Enough that a single reading changing cannot change the headline. At 210 questions per engine one reading is worth roughly half a percentage point, which is a workable floor; under about fifty questions the figure moves further than the thing it is measuring. Publish the count and the denominator either way and let the reader judge.
How do I calculate an AI mention rate?
Numerator over denominator, with both printed and the date attached, and one definition applied throughout. Our own India numerator is 2 and the denominator is 210 ChatGPT readings on 16 October 2026. Avoid averaging engines into one figure, because the same Indian questions gave 2 on ChatGPT and 0 on Gemini that day.
How should inconsistent answers be scored?
Score each reading as it came back, keep the raw answers, and report the spread rather than smoothing it. Repeated answers to the same question are not independent observations, so a benchmark that runs one question three times and calls it three samples has inflated its own denominator.
Do the Indian answers put a business forward as a choice?
Whether an answer put a business forward as a choice was not scored here at all. What was scored is which domains the answers were assembled from, and on Indian questions those were small Indian sites. Adding the naming score is the next change to this measurement, and it will arrive with its own N and date.
When is the India figure taken again?
The Indian set is frozen, so the next round is these same 210 questions with a new date on them, the usable count beside the figure, and a version number on any change to the scoring.
What an Indian publisher should take from the denominator
The useful thing about leading with 420 of 420 is that it sets the bar for everything you will be shown this year. An Indian supplier who arrives with an AI visibility score and no attempted count, no usable count and no engine names has not measured your visibility, they have measured something and called it that. Ask for the three numbers. We sell this measurement, so the request applies to us first, which is why our own India numbers are on this page with their denominators attached: 2 of 210 on ChatGPT, 0 of 210 on Gemini, 16 October 2026. Those are the figures we will be held to when the same questions are asked again.