Which websites ChatGPT and Gemini used on 50 India buying questions, and why IndiaMART and JustDial were not among them

On 22 October 2026 we put 198 India buying questions to ChatGPT and Gemini in 33 batches and saved every source each answer used. Fifty of those readings are blind, meaning no question in them names our own company. Across those 50 readings the two engines used 1,606 distinct sources, counting each website once per reading, drawn from 1,150 different hostnames. IndiaMART appeared in none of them. JustDial appeared in none. Sulekha, JioMart, Zomato and Swiggy appeared in none. The listing sites an Indian business is most often told to invest in were not what either engine reached for when a buyer asked a buying question about India.

What the engines used instead was a long tail of companies' own websites, a handful of global agency directories, and the Indian government. This page prints the counts.

How this reading was taken, and what fixed the market

• Date run, 22 October 2026. Engines, ChatGPT and Gemini. Two engines, not five.
• 33 batches, 198 questions, 51 readings in total. The 33rd batch names our own company in the question, so it is excluded and every figure below is read against the 50 blind readings.
• The market was fixed in the wording of each question. The search provider accepts a country parameter and answers from somewhere else, so India was named in the prompt text or it was not fixed at all.
• A source counts once per reading. One hostname cited four times in one answer is one source. 1,606 is that deduped count across the 50 readings.
• One hostname in the raw data, ccn.web.tr at 10 readings, is the data provider's own fetch infrastructure serving another publisher's page. It is resolved back to the page it serves and never counted as a publisher.

The sources the two engines used most

Read the fourth and fifth rows together. A global agency directory and a small Indian agency's own blog sit level, at 11 readings each. The directory has a page for every city and every service in India. The agency has its own site. On these questions they were used equally often.

The directories that did appear, and the class share

Six listing or review sites appeared at all. clutch.co in 11 readings, themanifest.com in 6, designrush.com in 3, goodfirms.co in 2, sortlist.com in 2 and g2.com in 1. All six are international platforms with India pages. No India first directory appeared.

As a class, review and listing directories took 12 of the 1,606 sources, which is 0.7 per cent. That replicates two earlier readings on different question sets, 0.8 per cent across 597 readings on 15 October 2026 and 0.7 per cent across 1,260 readings on 16 October 2026. Three independent question banks, three agreeing numbers. The practical point is not that a listing is worthless, it is that a listing matches one shape of question, and our page on how one directory reached the top in three cities while its whole class held 0.7 per cent works through exactly which shape.

The marketplaces did appear, and only through one kind of page

Indian marketplaces were not entirely absent. They appeared through their seller and corporate documentation and never through a product listing.

• Flipkart, 3 readings, through stories.flipkart.com and seller.flipkart.com.
• Amazon India, 5 readings, through sell.amazon.in, sellercentral.amazon.in, business.amazon.in, aboutamazon.in and press.aboutamazon.com.
• Meesho, 2 readings, through supplier.meesho.com and investor.meesho.com.

Every one of those is a page explaining how the platform works to a seller or an investor. Not one is a catalogue page. So when a buying question about India reached a marketplace, it reached the marketplace's own documentation about itself. For a seller the reading is direct. A listing on the platform does not put your business into that answer; a page on your own site explaining what you sell can. Our guide to getting listed on IndiaMART is still worth doing for the human buyers who read those pages on purpose, and it is a different job from this one.

Indian websites won a large share, and 243 of them were Indian domains

Counting hostnames ending in .in or .co.in, 243 distinct Indian domains were used across the 50 readings. The most used were upgrowth.in at 11 readings, pib.gov.in at 7, preevee.in at 5, themediaverse.in at 5, theshizz.in at 4 and epixs.in at 4. Most of those are small Indian agencies publishing on their own domains, which is the same result an earlier India cut produced and which our page on the Indian sources that won the Indian questions sets out with its own counts from 15 October 2026.

The Indian government is a source these engines use

Ten Indian government hostnames appeared, which makes the state a bigger presence in these answers than every Indian directory and marketplace combined. pib.gov.in in 7 readings, then cbic-gst.gov.in, msme.gov.in, ramp.msme.gov.in, incometaxindia.gov.in, niti.gov.in, commerce.gov.in and taxinformation.cbic.gov.in at 2 readings each, and ccpa.doca.gov.in and doj.gov.in at 1 each.

That is useful to a seller in a specific way. Where your business holds a registration, a certificate or a number that a government site publishes, the engines are already reading that site. Printing the same number in text on your own page, with the registration type named, ties the two together.

Where the whole count went

Questions about Indian directories and listing sites in AI answers

Does this mean an IndiaMART or JustDial listing is a waste of money

It means the listing is doing a different job from the one being measured here. Those platforms carry real human buyers who search inside them, and that traffic is not what these 50 readings looked at. What the readings settle is narrower and more useful, that a listing is not the route into an assistant's answer about your category, and a page on your own domain is.

Why would an engine cite a small Indian agency over a national directory

Because 91.1 per cent of the sources in these readings were pages on a company's own website, and because an Indian question needs an Indian source. There are not many Indian sites publishing dated, specific, measured work on any given commercial subject, so a small firm that publishes one is competing against a short list.

Was this run in Hindi or any other Indian language

Every question was in English, with India named in the wording. A language level reading is a separate measurement and this dataset does not carry one.

How many of these 50 readings cited AI Knows Us

None. We were cited in 0 of the 50 blind readings and named in 0 of them. That is the baseline our own work is judged against, and the figures on this page come out of the same run.

Which engines does this cover

ChatGPT and Gemini. Perplexity, Claude and Google AI Overviews were not in this run. Our product reads five surfaces and this dataset covers two, and the two are kept separate in every sentence.

The one page to publish after reading this

Nothing on this list is reachable by buying a profile. The route these engines took, 1,606 times out of 1,606, was to a page somebody wrote and published on their own domain, and 91.1 per cent of the time that domain belonged to a company selling something. So the first move for an Indian business is one page about the question your buyers ask before they buy, with the rupee figure, the date you checked it, and how you know. We run this reading as a product, which is how these counts exist at all, and the benchmark across 597 readings in three markets publishes the method so anybody can take their own.