AI Assistant Referrals in India: Why Google Analytics Has No Row for a Citation, and the Five Row Report That Does

Google Analytics 4 has no row for the event that matters most here, and the reason is mechanical rather than mysterious. Analytics records arrivals. Read four names in a chat, pick one, call them, and no browser ever leaves the assistant, so the acquisition report has nothing to count. Our survey of 16 October 2026 asked 210 Indian buying questions of ChatGPT and the same 210 of Gemini, kept every one of the 420 readings, and found our own domain cited twice on ChatGPT and not once on Gemini. Both of those citations happened inside an answer. Neither was a session. This page covers what to put in an Indian monthly report once you accept that, and how to classify the arrivals that do happen.

Where an assistant arrival actually lands in GA4

Two channels compete for an assistant visit, and published rules decide between them. Referral takes anything whose medium reads "referral", "app" or "link". Organic Search is the one with a gatekeeper outside the rules: a downloadable spreadsheet of sources and categories. Fetched on 16 October 2026 that spreadsheet held 819 rows, 129 of them marked as search, running through Bing, DuckDuckGo, Yahoo across more than twenty country hosts, Baidu, Naver, Yandex and Sogou. Assistants are absent.

An Indian property therefore behaves like any other: a referrer that survives the hop puts the visit in Referral, beside blog and news links, and never in Organic Search. Add a custom channel group keyed on the assistant hostnames and keep the default group untouched, so figures already quoted from last year still reconcile.

Then there are the arrivals with nothing attached. Google's own list of what produces (direct) and (none) covers untagged links, redirects and shorteners that discard the parameters, somebody typing the address, "links from offline documents such as PDFs and Word documents", and ad blockers. An assistant installed as an application belongs on that list, since a program is not a page and usually passes on no referring URL. And no tag of yours can ride inside somebody else's sentence, which is the boundary our Campaign URL Builder guide runs up against.

What the Indian readings showed, and why it changes the report

The India cut of that run is published as 420 readings of ChatGPT and Gemini in India, and its definitions are the ones used here: mentioned, cited and recommended are three states, and the same answer can be all three or none. A citation means one domain in one reading, deduped.

Widen the frame to all three markets and the run holds 630 questions, 1,260 readings with none lost, and 20,143 citations spread over 4,947 distinct domains. Three kinds of page absorbed nearly all of it: somebody's own commercial website took 82.0%, the platforms' documentation 12.3%, review directories 0.7%. The pass run the day before, from questions written independently, kept 597 of 612 attempted readings, logged 10,652 citations and put the first and last of those three at 85.6% and 0.8%. Two unrelated designs agreeing is stronger than either pass alone, and the 597 reading benchmark of ChatGPT and Gemini in three markets holds every row.

One Indian finding from that earlier pass decides what belongs in your report. The domains the two engines leaned on hardest were not platforms or directories. Four small Indian agency websites took the top places: upgrowth.in in 52 readings, themediaverse.in in 48, nicodigital.com in 44, hawkorbit.com in 33. Every one of those is a working firm whose own pages are carrying Indian answers. Set that beside review directories at 0.7% in the later pass and 0.8% in the earlier one, and a monthly plan built mainly on getting listed is working the thinnest part of the field.

The five row monthly report, and where each row comes from

Row	Source	What it proves
Sessions from assistant hosts	GA4, custom channel group, one condition per host	How many buyers clicked through, the smallest of the five
Unsourced sessions on narrow pages	GA4, (direct) and (none) by landing page	Visits with no referrer, read against the question that page answers
Citation count per engine	Your own question set, run and recorded	Whether a page on your domain was credited, engine by engine
Named without a link	The same answers, scored by a person	The state that never reaches analytics, and the one most reports miss
Cited domain list, counted	Every source in every answer, tallied	Who is read instead of you, which is next month's work plan

Four rules keep those rows comparable month to month. Freeze the question set and give it a version number, because changing the questions resets your history. Keep your own company name out of every question, since naming it returns it and the reading becomes a measurement of your own prompt. Run each question more than once per engine, because two answers on the same day can differ. And report each engine on its own line, never blended, since the two engines we measured did not agree often enough to share a number. The rules for separating a mention from a citation from a clickable link are in our 30 answer citation and link audit.

What an Indian team should do with row five

The cited domain list is the output that pays for the exercise. Take the top three domains an assistant reached for in your category, check whether your business is represented on any of them, and fix the ones where that is a matter of submitting accurate information rather than buying anything. Then look at what those pages do that yours does not. In the Indian set the winners were agency owned pages, so the thing being rewarded is a published page on a specific subject, with dates and figures a reader can check, on a domain clearly about that subject.

If your business already pays a monthly retainer, this changes the brief rather than the budget: the measurement half tells the writing half what to write. The rupee lines that actually move a quote are in Indian pricing for AI visibility monitoring tools. If the website itself is the weak link once traffic does arrive, the four kinds of supplier who fix that are in how to judge a conversion optimisation agency.

Questions Indian teams ask about AI referrals and analytics

Why does Google Analytics 4 show no traffic from ChatGPT or Claude?

Two reasons, both documented. Most assistant answers end without a click, so there is no session to record. And where a click does happen from an application rather than a browser tab, there is often no referrer, which Google's own help page lists among the causes of (direct) and (none) traffic. Set up a custom channel group for the hostnames and read the unattributed deep page arrivals beside it.

Will assistant traffic show up as organic search?

Not without help. That channel is gated by a spreadsheet, and when we counted it on 16 October 2026 the 129 search entries did not include a single assistant. What is left for an assistant referrer, with its medium of "referral", is the Referral channel.

How do I measure AI visibility for an Indian business without buying software?

Twenty questions is enough to start, phrased the way a buyer would phrase them and carrying no mention of your firm. Put each one through two assistants three times over, opening a fresh chat every time, and log the date, the engine, the naming result and the cited domains on one row apiece. An afternoon of that produces a month you can defend.

Can I compare my citation count with a competitor's?

Only if both counts come from the same question set, the same engines, the same repeat count and the same definition of a citation. Change any one of those and the two are not comparable, which is also why a figure quoted from a vendor's marketing cannot be set beside your own.

What does the report look like after three months?

Three rounds of the same five rows, question set unchanged, every round published including the ones that got worse. A series with a missing month reads as a steady month, which is a wrong impression created by omission rather than by any number.

The Indian answer nobody is logging

A procurement manager in Pune asks an assistant which suppliers to consider, reads four names, opens two tabs and closes the chat. Your analytics may record one of those tabs. The decision happened in the part it did not see. The Indian evidence says the companies winning that moment have published pages specific enough to be quoted, which is why the citation column belongs in the same report as the sessions column rather than in a file nobody opens. We sell that measurement and the writing that follows it. AI Knows Us works through five surfaces, ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, keeps your firm's name out of every question it asks them, and lists the sources that decided each answer. Its own pricing page showed Rs 499 as the lowest monthly plan when we fetched it on 16 October 2026. Measurements are published for two of them, and the Indian result was 2 citations across 420 readings. We print it because a supplier asking for your trust in any number should start with its own worst.