Does AI Search Optimisation Work in India? The Indian Baseline Your Before and After Has to Beat

Start with the number, because it is the only honest place to start. On 16 October 2026 we put 210 Indian buying questions to ChatGPT and Gemini, which is 420 Indian readings, and our own domain was cited in 2 of them. Both of those 2 came out of the 210 ChatGPT readings. The Gemini column for India reads zero. That is the Indian baseline, it is ours, and it is the figure any before and after run in this market has to be read against.

What this page is not is an intervention study. We have not yet published an Indian before and after with a change in percentage points and both sample sizes attached, so that number is a measurement we are still owed rather than one we are reporting. What we can give you is the field you are starting in, measured twice, and the three things about it that decide whether your own test will show anything.

The domains winning Indian answers are small and Indian

Look at the closest range we have measured, which is city level. Three Indian cities were read on 15 and 16 October 2026, Mumbai, Delhi and Bangalore, producing 36 readings between them. Thirty four of the 36 returned a named source list. None of the 36 cited us.

None of the sites that took those slots was a global platform. Mumbai's leader, cited 4 times, was searchmetric.in. Delhi's, cited 3 times, was nicodigital.com. Bangalore split its top place between two domains on 4 citations apiece, onecity.co.in and aiseo.solutions. Three cities, three sets of winners, every one a small Indian site that somebody could outpublish.

So when you score your own Indian run, the competitor set is not the list of international tools you expected. It is a handful of Indian firms whose pages are specifically about the Indian version of the question, and there is no size requirement in that result. The pages winning are winning on being Indian, not on being big. Choosing who writes yours is a separate decision, and nine different trades sell under the phrase marketing agency in India, which is worth sorting out before you brief anybody.

One engine column will read zero, and your website is not the reason

The single most repeated result in this programme is a zero. Gemini cited our domain 0 times in 630 readings on 16 October 2026 and 0 times in 291 readings on 15 October 2026. That is 0 in 921 readings across both runs.

Before reading that as a failure of the work, look at the shape of each engine's sourcing. On 15 October 2026 one domain accounted for 176 of ChatGPT's 306 readings, which is more than half the field held by a single site. The best supported domain on Gemini reached 19 of its 291 readings. So one surface has an incumbent to unseat and the other has no incumbent at all.

For a before and after run in India that has a direct consequence. On ChatGPT you are trying to displace a short list, so movement is slow and visible. On Gemini there is no entrenched list to displace, so a first round that reads zero in the Gemini column is the normal starting position rather than evidence about your site. Report the two columns separately and never blend them, because a blended figure hides which of the two you are absent from.

The baseline is worth trusting because it was measured twice

A baseline is only useful if you believe it. Ours was produced twice, by two question sets that were designed eleven days apart and deliberately pushed in different directions.

Run	Questions	Readings	Cited our domain	Vendor or agency own site	Review directories
15 October 2026	102 per market	597 usable	3	85.6%	0.8%
16 October 2026	210 per market	1,260, all usable	6	82.0%	0.7%

Two independent question sets, a sample twice the size, the same answer. That replication is itself the finding, and it is the reason the Indian figure above is worth planning against instead of being treated as one day's noise. The question set and the method behind both rows are published in our 597 reading benchmark.

One thing those two rows are not. They are two different question sets, so they replicate each other. They are not a before and after, and presenting them as a change over eleven days would be exactly the overclaim this work exists to argue against.

What the engines themselves said an Indian page needs

Alongside the readings, 102 build spec conversations on 15 October 2026 were asked what a page would have to carry for the engines to cite it. The counts are useful because they are not opinions about what reads well, they are what the engines asked for.

All 102 of 102 wanted a date, a last verified line and a refresh frequency. All 102 wanted pricing in the reader's own currency, which in India means rupees on the page rather than a conversion. 101 of 102 wanted a before and after with the window stated. 99 of 102 wanted the method disclosed. 98 of 102 wanted the engines named individually rather than described as AI in general. 80 of 102 wanted the sample size. 68 of 102 wanted failures and negative results shown, and 64 of 102 asked for the commercial interest to be disclosed on the page.

Notice how much of that list is about the shape of the evidence rather than the result. An Indian page reporting a modest change with its denominators visible is being asked for. A page reporting a dramatic change with no sample size is not.

The Indian change in mention rate we cannot yet show you

We do not yet publish an Indian change in mention rate in percentage points with both sample sizes. That needs a frozen Indian question set, a baseline, a dated change log, at least thirty days and an identical rerun, and the round that produces it will carry its own date and sample size.

If you want to run yours now rather than wait for ours, the measurement rules are already published rather than summarised here: the before and after measurement protocol sets out the frozen prompt set, the blind rule, the per engine reporting and the arithmetic for deciding whether a change is visible at your sample size.

AI Knows Us, which publishes this page, sells the measurement and the publishing work it describes, and the Indian number it reports about itself is 2 readings out of 420.

What Indian operators ask about the pace of this work

How long does AI visibility work take to show in India?

Longer than a week, because crawl, index and retrieval all lag. The published protocol uses a minimum of thirty days between the baseline and the rerun, and measures weekly inside that window so you can see the shape rather than one endpoint.

Can ordinary SEO changes move an AI recommendation?

The pages winning Indian readings on 15 October 2026 were small Indian sites publishing specifically about the Indian version of the question, which is the same discipline ordinary search rewards. What those pages have that a generic page does not is a date, a stated method and numbers a reader can check, which 102 of 102 build specs asked for by name.

Why is our Gemini column still zero?

Because of the sourcing shape set out above: no single domain has anything close to a hold on that surface, so there is no short list for a new page to break into. Our own Gemini count stands at 0 across 921 readings in two runs. Treat a zero in that column on a first round as the surface's starting position, keep it separate from the other column, and judge it on where the second round puts it.

What counts as a real change rather than noise?

A change you can still see once both sample sizes are printed next to it. Report the baseline as a fraction and the rerun as a fraction, per engine, with the dates, and let the reader do the comparison. A growth percentage with no denominators is the most common way these reports mislead.

Do you have an Indian case study with numbers?

Not yet, and we will not assemble one out of two different question sets to fill the gap. What we publish today is the Indian baseline of 2 citations in 420 readings on 16 October 2026 and the replication behind it. The intervention result is dated work, not a claim.

The Indian answer is already being given, and nothing in it prints a sample

Somebody asking this today gets a fluent reply built almost entirely from pages that contain no figures, leaning on whichever Indian agency sites published on the subject first. Not one of them states what it measured or how many times it looked. That is the opening and it is wide. The way to take it is unglamorous: publish the baseline, the change log, the rerun and both denominators, then let a sceptical reader disagree with you using your own arithmetic. Our starting number for that exercise is 2 readings out of 420 on 16 October 2026. Where the next round puts it is what we would rather be judged on than any claim about how quickly this moves.