Which sites AI assistants actually cite
Every article on this subject publishes a leaderboard of the most cited domains. We downloaded the raw data behind one of the few studies that releases it, counted the citations ourselves, and found the leaderboard is answering the wrong question.
Written by Ron Kloth, founder of RankJudge.
Across 1,050 citations from 100 prompts, 784 different domains were cited, and 669 of them appeared exactly once. The ten most cited domains took 11.5% of all citations between them. The single most cited domain in the entire dataset took 2.1%.
By the end of this page you will know how concentrated AI citations really are, which well-known sites were absent, how differently the three engines behave, and what that means if you are about to pay for a tool that tracks your brand in AI answers. The data was collected on and we analysed it on.
The finding that surprised us is not in the numbers above. It is which two of the biggest names in software comparison never appeared at all.
What we did, and what we did not do
RankJudge runs no testing and never claims to. We did not run a single prompt through any engine.
David Quaid at GEO Search Console ran 100 prompts across ten everyday topics through ChatGPT (gpt-5-6, read by a scraper with search forced on), Claude (Haiku 4.5 via the API, search forced on) and Gemini (3.6 Flash via the API with Google Search grounding, left to decide for itself) on . Hepublished the study and, unusually, the raw CSV.
We downloaded that file and counted what was in it. His article reports how the engines behaved. It does not ask who got cited, which is the question that decides where anyone should spend effort, so we asked it of his data. Everything below is arithmetic over a file you can download and check.
Citation scatters where ranking concentrates
| Total citations | 1,050 |
|---|---|
| Distinct domains cited | 784 |
| Domains cited exactly once | 669, or 85.3% |
| Share of citations going to once-only domains | 63.7% |
| Share taken by the top ten domains combined | 11.5% |
| Share taken by the single most cited domain | 2.1% (youtube.com, 22 citations) |
| Share going to a list of 22 household-name domains | 13.1% |
This is the finding, and it is why a "top 50 most cited websites" list is close to useless as a plan. On a Google results page the first position takes a large share of the clicks. In this dataset the first position takes two percent of the citations, and nearly two thirds of all citations go to domains that were cited once and never again.
Ranking is a contest for a small number of slots. Citation, on this evidence, is not.
Watch: Why the AI citation leaderboard is the wrong question
Recorded 22 October 2026. Prices quoted in the audio were the listed prices that day. The figures on this page are the current ones, and where the two disagree, the page is right. Plays from YouTube, and nothing loads from YouTube until you press play.
Three corrections to the video, and the newest one is not an error it made. Its sampling example, 50 prompts weekly against 100 daily for roughly 200 readings a month against 3,000, was read off Moz Pro and SE Ranking pricing pages on 26 September 2026 and was true then. Re-read on 30 October 2026 those plans are 150 weekly and 25 daily, roughly 650 readings against 760, so the arithmetic in the video is right and the two tools it describes have moved much closer together. The point it makes about sampling a scattered surface is unaffected. The other two: it opens by calling this "a data set of 150 AI citations" and forty seconds later says "roughly a thousand", which is the right figure: 1,050 citations across 784 domains. The 150 is wrong and nothing on this page rests on it. It also sends viewers to rankjudge.com/ai-visibility, which is not a page on this site. The address is the one you are on, rankjudge.com/who-ai-assistants-cite, and we have pointed the spoken one at it so anyone typing what they heard still arrives. Two smaller things: ChatGPT searched on 98 of the 100 prompts rather than every time, and the study is by David Quaid at GEO Search Console, which is his own site and not Google Search Console.
Read the full transcript
Our data shows AI citations actually scatter wildly across 784 different domains, with 85% of them cited exactly once. The leaderboard strategy is completely broken. Let’s look at the raw data together and I’ll show you exactly why.
Breaking down the data
Okay, so let’s break this down. We analyzed a data set of 150 AI citations, all generated from just standard everyday prompts. And right off the bat, the raw data was honestly staggering.
You know how in traditional SEO, everything kind of pools around a handful of mega-sites? Well, that doesn’t happen here. Instead, citations scattered wildly across those 784 distinct domains I just mentioned.
I mean, really think about that scale for a second. Out of roughly a thousand citations, almost 800 completely different websites were pulled in. It’s like throwing a handful of darts and hitting almost every single spot on the board.
And the way this scatter actually breaks down is an absolute bombshell. 669 of those domains, which makes up a massive 85.3% chunk of the total, appeared exactly one time. Just once. And then poof, never again.
What does this mean for us? Well, it means that if you’re looking at popular leaderboard lists, you’re literally answering the wrong question entirely. Traditional ranking is a tight contest for a few top slots, but AI citation is incredibly, almost unimaginably, fragmented.
Bashing the leaderboard myth
Let’s take a look at why these popular lists are failing us. To figure all this out, we have to rely on hard math, not just industry hype. And that’s exactly what we do here at Rank Judge with our honest AI, SEO, and AEO tool reviews.
So, we took this raw data from a brilliant Search Console study by David Quaid, sat down, did the hard arithmetic, and asked the one absolutely crucial question: Who actually got cited in the end?
Where do citations go? Mapping the tops
The contrast here with traditional Google search is just massive. Even at the very top, AI citations are incredibly fragmented.
Get this: the top 10 domains combined took just 11.5% of all the citations. And the single most cited domain overall, it was YouTube, capturing a measly 2.1%.
I mean, think about that. Winning an entire data set with just 2%. It’s kind of like winning a marathon by taking one single step. Nearly two-thirds of all citations went to domains that were cited once and then completely ignored.
The missing comparison giants
Seriously, where are the big names? We went ahead and isolated 10 software-specific prompts within the data set to see what would happen. And across those prompts, massive comparison giants, I’m talking heavy hitters like Capterra and TrustRadius, got exactly zero citations. Nada. G2 managed to scrape by with just one.
Meanwhile, 109 completely obscure domains, places with names like costbench.com, popped up once or twice. It’s concrete, undeniable evidence that just being the dominant name in your category doesn’t automatically secure you an AI citation.
Three very different engines
We’re looking at ChatGPT, Claude, and Gemini. When you actually break down the behavior engine by engine, the differences are wild.
- Claude: searched exactly once per prompt, but cited roughly 4x as many sources as ChatGPT
- Gemini: declined to search at all on nearly half the prompts; when it did search, it cited very widely
- ChatGPT: searched every time but cited far fewer sources
The big takeaway here: you simply cannot treat AI as a single monolith. They all have completely different personalities.
The Reddit myth debunked
And this next stat completely debunks a really popular marketing narrative going around right now, the whole idea that AI loves Reddit. Yeah, total myth.
Out of Reddit’s 20 citations and YouTube’s 22 citations in this entire data set, every single one of them came exclusively from Gemini. ChatGPT cited neither of them, not a single time in 216 searches.
So when people out there confidently say "AI cites Reddit," they’re really just describing Gemini, not the whole AI landscape.
ChatGPT never searches your exact query
But wait, the study adds an even crazier detail about ChatGPT specifically. In 216 searches, ChatGPT never once, literally zero times, searched for the question exactly as the user typed it in. It rephrased. It dug around for context. But it never used the exact input.
So think about this: if you’re writing a web page perfectly tailored to match a specific typed question, you’re essentially optimizing for a prompt the engine isn’t even looking for. Crazy, right?
Buying AI visibility tools: what specs actually matter
So, because 85% of these sources appear only once, a tool’s sampling rate is quite literally the only specification that matters.
- Weekly, 50 prompts: roughly 200 monthly readings, missing the vast majority of sources that flash into existence and disappear
- Daily, 100 prompts: roughly 3,000 monthly readings, far better coverage of a wildly scattered data set
If a tracking tool is only checking 50 prompts once a week, giving you roughly 200 readings a month, it’s missing the vast majority of sources that just flash into existence and disappear. You’re basically operating completely in the dark.
But a high-frequency tool that samples daily, you’re looking at roughly 3,000 monthly readings. That massive gap in data collection is for sure the only way to actually catch such a wildly scattered data set.
Final thought
Which leaves us with the ultimate question: Is your AI visibility tool actually seeing the whole picture? Because if it isn’t sampling constantly, you are missing all the action.
Like I said at the start, I highly encourage you to head straight over to rankjudge.com/ai-visibility right now. Go read our full data breakdown and check out our honest tool comparisons.
Because honestly, if the AI engines aren’t even searching for the exact questions your customers are typing, what makes you think your current visibility strategy isn’t already entirely obsolete?
Something to think about.
The comparison giants were not there
We probed the dataset for the best-known software review and comparison properties, across all 100 prompts:
These counts, and the Reddit finding below, are entries 6 and 7 in our freeAI SEO Fact Check 2026, alongside nine other claims checked against the document each one rests on.
| Site | Citations in 100 prompts |
|---|---|
| G2 | 1 |
| Capterra | 0 |
| TrustRadius | 0 |
| Software Advice | 0 |
The ten software prompts cited 109 domains once or twice each, among them aicomparison.ai, costbench.com, saascompared.io and stackfyi.com. Names most people in the industry have never heard of appeared where the category leaders did not.
Absence across 100 prompts is not proof that Capterra is never cited, and we will not claim it is. It is evidence that being the biggest name in a category does not secure a citation the way it secures a search ranking.
The three engines barely resemble each other
| Engine | Searched | Searches | Citations | Distinct domains |
|---|---|---|---|---|
| ChatGPT (gpt-5-6) | 98 of 100 | 216 | 223 | 165 |
| Claude (Haiku 4.5) | 100 of 100 | 100 | 424 | 385 |
| Gemini (3.6 Flash) | 54 of 100 | 91 | 403 | 332 |
Claude searched once per prompt and cited roughly four times as many sources per search as ChatGPT. Gemini declined to search at all on nearly half the prompts, and when it did it cited widely.
And the two most cited domains overall belong to one engine. youtube.com at 22 citations and reddit.com at 20 were cited by Gemini and by nothing else. ChatGPT cited neither, once, in 216 searches. Every article telling you that AI assistants love Reddit is describing Gemini and calling it AI.
The study itself adds the detail that makes this concrete: ChatGPT"searched the prompt as typed 0 times in 216 searches". Not rarely. Never.
What this means if you are buying an AI visibility tool
This is the practical end of it. If citations were concentrated in twenty domains and the engines agreed with each other, tracking would be easy and a weekly check would be plenty. The data says the opposite on both counts, and that makes one specification matter more than any other: how often the tool actually samples.
A tool checking 150 prompts once a week is taking roughly 650 readings a month of a surface where 85% of sources appear once. A tool checking 25 prompts daily is taking roughly 760. Those were 50 weekly and 100 daily when we first wrote this in September, and both vendors moved; the principle holds and the gap between those two particular plans has closed, as set out inMoz Pro vs SE Ranking, where exactly those two sampling rates are what separates the two tools.
The two tools that do nothing else are compared onProfound vs AthenaHQ, including the thing neither of them publishes: how either one samples prompts in the first place. That gap matters more after reading the table above than it did before.
What we could not confirm
- Whether any of this holds beyond one day. The data is a single snapshot from 7 October 2026. Its author repeated a subset of 20 prompts to test stability and reports that not one of ChatGPT's 122 different searches came back in all three runs.
- Whether the pattern holds for real users. These are API and scraper results, not a logged-in person's app with memory and history, which the author states plainly.
- How much of this was placed there on purpose. Nothing in the dataset distinguishes a page an engine found from a page somebody worked to put in front of it, and getting a brand mentioned by an assistant is a service with a price list. Practitioners who sell it name syndicated press releases as the method that works, which means some share of any citation count is a measurement of who paid for placement rather than who answers the question best. We cannot say what that share is here, and neither can the study. It does not affect the finding this page rests on, because scatter across 784 domains is scatter either way, but it does change what a domain's presence in the list proves about that domain.
- What happens inside Google AI Mode, whose own fan-out cannot be observed from outside Google.
- Whether a citation produces a visit. Nothing in the dataset measures what happened next, and a citation is not a click.
- Whether our niche behaves like the sample. Ten of the 100 prompts were about software. That is a small base for a conclusion about software buying.
- Any of the competing leaderboards. We read the study and the dataset. We did not open the agency roundups that publish most-cited lists, so we take no position on their numbers beyond noting that one of them reports a single domain's share as both 47.9% and 7.8% on the same page.
Where to go next
If you want to know whether assistants can reach your site at all, which is the question underneath all of this, our free crawler check answers it in a few seconds and stores nothing. If you are choosing a tool rather than reading research, our comparison by business situation starts from the situation. And every figure we publish carries the date we read it, logged atwhat these tools used to cost.
Questions people ask
- Which websites do AI assistants cite most?
- Across 100 prompts run through ChatGPT, Claude and Gemini on 7 October 2026, the most cited single domain was youtube.com with 22 citations out of 1,050. That is 2.1%. The top ten domains together took 11.5%. There is no dominant source, which is the opposite of what a "most cited websites" list implies.
- Do AI assistants cite the pages that rank on Google?
- Mostly not. The study behind this data checked the engines’ citations against Google’s US top ten for the search each engine had just run, across 30 prompts and 122 searches. Only 8% of ChatGPT’s cited pages, 3% of Claude’s and 14% of Gemini’s were in that top ten. At site level the figures were 31%, 9% and 24%. The study’s author calls these rough floors.
- Does ChatGPT search for the question you typed?
- No. The study reports that ChatGPT "searched the prompt as typed 0 times in 216 searches". It ran about 2.2 searches per prompt and none of them was the question as asked. A page written to answer the typed question is answering something no engine searched for.
- Do Reddit and YouTube get cited by every AI assistant?
- No, and this is the most repeated mistake in coverage of this topic. In this dataset youtube.com was cited 22 times and reddit.com 20 times, and every one of those citations came from Gemini. ChatGPT cited neither, once, in 216 searches. "AI cites Reddit" is a statement about one engine.
- Are G2 and Capterra cited by AI assistants?
- Barely, in this sample. Across 100 prompts G2 was cited once. Capterra, TrustRadius and Software Advice were not cited at all, while the ten software prompts alone cited 109 obscure domains once or twice each. Absence across 100 prompts is not proof they are never cited, and it does show that category dominance does not secure a citation.
- Did RankJudge run these tests?
- No. RankJudge runs no testing and never claims to. David Quaid at GEO Search Console ran the prompts on 7 October 2026 and published the raw CSV. We downloaded that file on 22 October 2026 and counted what was in it. Every figure here is arithmetic over somebody else’s published data, and the file is linked so you can repeat it.