The Multi Engine AI Citation Audit Framework: One Column Per Engine, Never a Blended Score

The output of a multi engine AI citation audit is a table with one row per engine and one row per market, and it is never a single score. We can show you why from our own readings. 210 questions per market went out on 16 October 2026 to both ChatGPT and Gemini across the United States, the United Kingdom and India: 630 questions, 1,260 readings, none of them discarded. Our own domain was among the sources in 6 of them. The Indian half reads 2 of 210 on ChatGPT and 0 of 210 on Gemini. Average those two cells and you get a figure that describes neither engine and hides the one thing an agency needed to know, which is which column is the zero.

This page is the audit design, not a result you can borrow. We sell the measurement and the publishing work that follows it, so the framework is written to be run by somebody who has not paid us.

The framework in one table

Eight fields, all decided before the first question is asked, because four of them are the denominator and a denominator decided afterwards is not a denominator.

Field	What we set it to
Questions per market	102 on 15 October 2026, then 210 on 16 October 2026, frozen before the run
Engines, named	ChatGPT and Gemini. Two, written as names rather than as a category
Markets	United States, United Kingdom, India, with the country written into the question text rather than set as a parameter
Readings attempted	612, then 1,260
Readings usable	597 of 612, then 1,260 of 1,260
Counting rule	One domain in one reading, deduped
Brand rule	Nothing we sell is named anywhere in the file
Date and next reading	On the report, with the interval set before the result is seen

Those 102 build spec conversations of 15 October 2026 produced three counts worth naming. 98 of 102 wanted each engine written as a name rather than as AI in general, 80 of 102 wanted the sample size, and 78 of 102 wanted the question set itself handed over so a stranger could run it again. All three land in the first three fields above.

Why the two columns cannot be averaged

The engines do not source alike, and the difference changes what an agency sells. ChatGPT took 306 of the 15 October 2026 readings and Gemini the remaining 291. Over those 306 ChatGPT readings one domain turned up in 176 of them. Gemini's busiest source reached 19 readings out of its 291.

So one surface has a short, entrenched list holding more than half the field, and the other has no entrenched list at all. On the first, an Indian agency is displacing a handful of incumbents and movement is slow and visible. On the second there is nothing to displace, and a first reading of zero in that column is where the surface starts rather than a verdict on a client's website. Gemini is the clearest case. Across the two runs it produced 921 readings carrying nothing of ours.

Sizing the run, and the two counts that go together

Attempted and usable are two numbers, and a report giving only one has left out the part a sceptical client asks about. The 15 October pass aimed at 612 readings and finished with 597. The 16 October pass went out to 1,260 and lost none. The second figure is the stronger claim precisely because the attempt total is printed beside it.

Size matters less than repetition. The two passes came from question sets designed eleven days apart and pushed in deliberately different directions, and they land in the same place: 3 citations of our domain in 597 readings against 6 in 1,260, a vendor or agency own site share of 85.6% against 82.0%, and review directories at 0.8% against 0.7%. Different questions, double the readings, one answer.

Two rules that set the headline before any reading is taken

One reading plus one domain is one citation, so a domain quoted four times inside the same answer still scores one. Switch the unit to raw URL occurrences and the same 16 October rows total 36,576 rather than 20,143. Neither number is wrong and the two are not comparable, so the rule goes at the top of the report and does not change between rounds.

What the Indian column looks like at baseline

Indian answers are being assembled from small sites with .in on the end. Across the 597 readings of 15 October 2026 the Indian set reached for upgrowth.in in 52 readings, themediaverse.in in 48, nicodigital.com in 44 and hawkorbit.com in 33. Ten cities were read one at a time on 15 and 16 October 2026, six buyer questions apiece on both engines. We were cited in none of them. Mumbai's leading local source was searchmetric.in on 4 readings and Delhi's was nicodigital.com on 3, while Bangalore gave its top place jointly to onecity.co.in and aiseo.solutions.

Two things follow. The competitor list an Indian audit produces is not the list of international tools anybody expects, so the scoring sheet needs a free text column for domains rather than a fixed list. And the pages taking those slots win on being specifically about the Indian version of the question, which makes the decision of who writes yours a budget line rather than an afterthought, and the five different businesses behind the phrase marketing agency in Mumbai is the shortest route through it. Outside those ten cities we hold no figure at all. For an unmeasured city, carry the market figure and label it as the market figure.

The two cells this framework leaves open, named on purpose

Being used as a source and being put forward to a buyer as a choice are two different events. Everything above scores the first. We have not yet scored recommendation with its own numerator, denominator and date, so the framework carries that column empty and says so.

The second open cell is a multiplier. A common claim in this category is that appearing on three or more independent third party directory listings multiplies your citations several times over. We have not run a controlled test that adds listings and re-measures, so we publish no multiplier. What we hold points the other way: review directories took 0.7% of the 20,143 citations recorded on 16 October 2026 and 0.8% of the 10,652 recorded on 15 October 2026, while 82.0% went to a vendor's or an agency's own website. An Indian agency choosing between listings work and dated publishing can settle it on 0.7% rather than on opinion. Both runs, their controls and the corrections applied to them are written up in our 597 reading benchmark of ChatGPT and Gemini across three markets.

Questions about running an AI citation audit in India

How many questions does an Indian audit need?

Fewer than you will be told, run more than once. Our first pass was 102 questions per market and our second was 210, and the second agreed with the first rather than correcting it. A set you will re-run next month on the same frozen wording tells a client more than a larger set run once.

Should we hand a client a column of zeros?

It does. A cell reading 0 of 210 has been measured; an empty cell has not. Our Indian Gemini cell reads 0 of 210 on 16 October 2026, and that engine has reached 921 readings in total with our name absent from all of them. A report that drops the zero column has removed the cell telling a client which surface the next quarter of work is aimed at.

Can this framework cover five engines?

The design extends to any engine you can name and read on a schedule. Our own two runs covered ChatGPT and Gemini only. Perplexity, Claude and Google AI Overviews were not asked these questions, and a framework claiming five surfaces from a dataset of two would fail the standard it asks an agency to hold itself to.

What goes into the pack an Indian client receives?

The unit, both reading counts, every engine with its own fraction, the market, the date it was taken, the date of the next one, and any cell that came back empty. The question set goes in the same folder, because 78 of 102 build specs asked for exactly that.

The framework is worth more than our score, and our score is 2 of 420

Run this design on your own category in India and the first reading will probably be small, because ours was: 6 of 1,260 readings across three markets, with the Indian ChatGPT column at 2 and the Indian Gemini column at nothing. We publish that because an audit framework from a supplier whose own column reads well is one nobody can test. AI Knows Us reads five surfaces for a business and writes the pages that change what those surfaces find, and the two columns above are the two we have read often enough to print. Fix the eight fields, run the questions blind, and the number you end up with is one a client can argue with using your own arithmetic.