Sign inJoin the waitlist

Brand Citation Rates in AI Answers: What 1,260 Readings Show About Which Names Get Repeated, and the Industry Baseline Nobody Has

Across 630 buying questions put to ChatGPT and Gemini on 16 October 2026 we recorded 20,143 citations, and eight of the ten most cited domains were an operator's own documentation or a code host. The order is a publishing ranking, not a brand size ranking.

Published by AI Knows Us · Updated 16 October 2026

Ask what share of businesses get named in an AI answer and nobody can hand you a sourced figure, because no such study exists with its questions, its engines and its date attached. What you can get is a count with a visible denominator. Here is ours. We put 630 buying questions to ChatGPT and Gemini on 16 October 2026 across three markets, took 1,260 readings, and found our own domain among the sources of 6 of them, which is 0.5%. Inside the United States, ChatGPT returned us once over its 210 questions and Gemini never did over the same 210. A wholly separate question set on 15 October 2026, 597 usable readings, had already put us at 3. Two question sets designed eleven days apart landed in the same place, and the fact that it replicated is worth more than either run on its own.

The counting rule, before the numbers

Every figure here counts one domain in one reading, with repeats inside a single answer collapsed. That rule, rather than the data, is what decides the total: counted the other way, by raw url occurrence, the same run yields 36,576 instead of 20,143. Neither is wrong and the two are not comparable, which is the trap when you set one vendor's citation figure beside another's. The arithmetic and the reason the rule is almost never printed are in what counts as one citation. Read that page before quoting any citation rate, ours included, and note while you are there that being in the answer text, being one of its sources and being put forward as a choice are three separate events. This page counts the middle one.

The ten domains that were actually repeated

Same run, same date, 4,947 distinct domains behind it.

readings that cited it, of 1,260domaintype
446developers.google.comofficial or platform
328support.google.comofficial or platform
325developers.google.cnofficial or platform
309blog.googleofficial or platform
269help.openai.comofficial or platform
243github.comsocial or community
230developers.openai.comofficial or platform
199semrush.comvendor or agency own site
197theverge.comnews or publisher
147searchengineinfo.orgvendor or agency own site

Eight of the ten are an operator's own documentation or a code host. Two are a company's own website, and they are the pair worth sitting with. One is semrush.com at 199. The other is searchengineinfo.org at 147, which is not a household name in anybody's market and which arrived in the same top ten off its own pages. If citation frequency tracked brand size, the second row would not exist.

Sorting all 20,143 citations by the kind of page they landed on points the same way.

kind of pageshare of citations, 16 October 2026
a company's or an agency's own website82.0%
official or platform documentation12.3%
news or a large publisher3.2%
social or community1.9%
review directory0.7%

The 15 October question set produced 85.6% and 0.8% for the first and last of those rows. So what is rewarded in the aggregate is having published a page that answers the question, which is a different asset from being well known and, unlike being well known, one a smaller company can go and build this quarter.

The naming order in one category, counted

Most people asking about brand citation rates want to know how often the engines name the leaders in their own field. We can report that for exactly one field, which is ours, from the 597 usable readings of 15 October 2026 and the 10,652 citations in them.

  • semrush.com, cited in 95 readings
  • tryprofound.com, 61
  • otterly.ai, 58
  • depra.ai, 53
  • peec.ai, 49
  • Scrunch, 11, and Rankscale, 3
  • our own domain, 3

That list measures one thing only: how frequently an engine typed a name. It settles no question about quality, it is not the output of a test, and we have put none of those products through a shared benchmark. We make one of them, and we sit near the bottom of our own list, which is the plainest reason to read it as a tally rather than a verdict.

The pattern inside it is the same one the top ten shows. The order tracks how much of the category each company has published pages about, and the one established search platform on the list leads a group of specialists built entirely for this job. Anyone asking an engine to rank suppliers is reading a publishing ranking and calling it a shortlist, which is also the reason a supplier's own evidence has to be read rather than trusted. The questions worth putting to one are set out in how to read the evidence a generative engine optimisation agency shows you.

The baselines that do not exist, named individually

This is the part a benchmark page owes its reader, because the gaps are where invented numbers come from.

  • There is no per industry rate in this dataset. The vertical questions were generic, so there is no software number, no healthcare number and no legal number. A sentence implying one would be false.
  • There is no per city rate in it either. City figures come from a separate scan: 10 named cities, 6 buyer questions each, both engines, 120 readings, all usable. Our domain was cited 0 times in all ten. There is no figure for a city outside those ten.
  • Two engines were read, not five. ChatGPT and Gemini. Nothing here describes Perplexity, Claude or Google AI Overviews, which sat outside both runs. The product we sell reads all five surfaces, and that does not lend a single figure to a dataset built on two.
  • There is no average brand. Swap the wording of the questions, let a brand name into one of them, change the app, change the week, or count recommendations instead of sources, and the figure moves while nothing in the world has. A rate belongs to its test, not to a company.

What does exist, and it is the only baseline that will survive a challenge, is your own first reading on a frozen set. Take between thirty and a hundred questions in your buyers' words, name no company in any of them, send the same wording to each engine, report each engine in its own column with its own denominator, and write the date on it. Your second reading is then comparable to your first, which is more than any industry average can offer you.

Questions about brand citation rates in AI answers

What is the average citation rate for a business in AI answers?

Nobody has published one with a question set, an engine list and a date attached, so there is no figure to quote. The honest substitute is a count with its denominator visible: ours is 6 of 1,260 readings on 16 October 2026, and 3 of 597 on 15 October 2026.

How often do AI assistants name the top companies in a category?

In the one category we have measured, the leader was named in 95 of 597 readings and the seventh name in 3. That spread is wide, it is a count of naming rather than of quality, and it was taken on a single date on two engines.

Why is the Gemini number zero?

Because it is. Gemini cited our domain 0 times in 630 readings in the second run and 0 times in 291 readings in the first, which is 0 in 921 across both. It is the most repeated single result in this programme and it is the reason we never publish a blended figure across engines.

Does a bigger brand automatically get cited more?

Size correlates with how much a company has published, which is the thing being rewarded. The clearest evidence against size as the cause is in the table above: searchengineinfo.org, cited in 147 of 1,260 readings, sits in the same top ten as semrush.com.

How do I measure my own rate without inventing a benchmark?

Freeze the questions, keep every brand name out of them, name the engines, print the numerator and denominator for every rate, and date it. Then compare yourself to your own last reading rather than to a published average, which is the approach argued in what a good AI mention rate is without a fake industry benchmark.

Why we publish the run that makes us look worst

We sell the software that takes this measurement, and running it twice has told us mainly that the engines almost never reach for us: 6 readings in 1,260, none of them on Gemini. That is printed here for a practical reason rather than a modest one. A reader who has worked out that we have an interest will test the uncomfortable number first, and if it is there the rest of the table survives the test. Everything behind the first run, the wording, the controls and the source breakdown, is published as 597 readings across ChatGPT and Gemini in three markets. Your own rate starts the same way ours did, with a denominator you chose and a date you wrote down.

Read next

AI Visibility Audit Deliverables for a UK Agency: The Five Artefacts, and Why the Zero Column Is One of ThemWhat a British agency should hand a client at the end of an audit, including the findings that did not survive a second reading.A Workflow Standard for AI Search Optimisation: Six Stations, and Who Owns Each OneThe execution sequence an Indian agency can hold a team to, measured against what our own site serves eleven named crawlers.The Multi Engine AI Citation Audit Framework: One Column Per Engine, Never a Blended ScoreA reproducible audit design for an Indian agency, built from 1,260 readings on 16 October 2026 and the 597 that came eleven days before it.A Small UK Agency Against a Large Consultancy on AI Search Optimisation: What the Citation Data Says Each One Can OwnThree British agency domains took dozens of citations each in our readings, and none of them is a large firm. Here is what that result does and does not tell a ten person shop.Build or Buy an AI Visibility Stack: The Five Silent Wrong Answers an In House Build ProducesWhat it actually takes to run a multi engine tracker yourself, measured from the defects in our own build, with rupee prices for the alternative.AI Search Optimisation and SEO in India: What Each One Costs, and What Each Return Is Counted InMost of the production work is already inside your search invoice, two budget lines are genuinely new, and the two results are counted in units that cannot be divided by each other.

Or browse the whole archive.

See what AI says about you.

The first scan is free, it needs no account, and it asks the questions your buyers ask.

Free. No card. We ask 5 real buyer questions on 2 AI apps.

Already sure you want this? Join the waitlist. We open on 31 October 2026. It costs an email address. No card, and nothing is charged. Or talk to a person if you have a question first. Or read how we fix what it finds.