AI Citation Study: What ChatGPT, Gemini and Claude Actually Cite
The three engines barely agree on anything. Across 27,854 citations we logged from ChatGPT, Gemini and Claude answering the same 374 buyer questions, 80.6% of cited sites were named by just one engine and only 5.2% by all three. Authority is not the gate either: 84.1% of cited sites sit outside the web's top 50,000.
So track each engine separately, and read it monthly rather than weekly, because re-asking the same engine the same question a fortnight later changes roughly half the sources on its own. Method, four more findings and the downloadable dataset are below.
Almost everything written about getting cited by AI is written from the outside: someone types a prompt, screenshots the answer, and generalises. That produces advice about one prompt on one day in one engine.
We are in an unusual position to do better. MentionAgent runs an AI visibility scanner for its customers, which asks ChatGPT, Gemini and Claude a fixed panel of buyer-intent questions on a schedule and records every source URL each engine returns. That log had been used for one thing, checking whether a customer's brand came up. The rest of it, every source the engines cited for every question, was sitting untouched.
This is what is in it.
Method
Every figure on this page comes from the same log and can be recomputed from it.
| Questions | 374 buyer-intent questions |
| Businesses | 23, in unrelated verticals |
| Engines | ChatGPT, Gemini, Claude |
| Answers logged | 3,450 (ChatGPT 1,157, Gemini 1,155, Claude 1,138) |
| Citations | 27,854 |
| Distinct domains cited | 6,779 |
| Period | 8 July to 17 August 2026 |
The questions are not a designed sample. They are the questions 23 live businesses actually want to be found for, which is why the set spans foundation repair, sobriety support, travel photography, HR services, social media scheduling, email validation and local search. Each engine answered all 374. Every question was asked more than once, on separate dates, which is what makes the repeatability measurement below possible.
A citation means the engine returned that URL as a source alongside its answer, through its own API. It is not a scrape of the answer text and not a guess at what the model read.
1. The three engines agree on 5% of their sources
Taking the questions all three engines answered, and asking how many engines cited each domain:
| Cited by | Domains | Share |
|---|---|---|
| Exactly one engine | 4,197 | 80.6% |
| Two engines | 741 | 14.2% |
| All three engines | 269 | 5.2% |
Four sources in five are seen by one engine and no other. This is the single most consequential number here, and it has a direct practical reading: a report on your visibility in one AI engine tells you close to nothing about the other two. If you are tracking ChatGPT only, you are measuring roughly a fifth of the picture.
It also means the engines are not converging on a shared canon of trusted sites. They are reading different indexes.
2. ChatGPT answers a quarter of questions without citing anything
Answers that returned zero sources:
| Engine | Answers | With no sources | Share |
|---|---|---|---|
| ChatGPT | 1,157 | 305 | 26.4% |
| Gemini | 1,155 | 16 | 1.4% |
| Claude | 1,138 | 16 | 1.4% |
Roughly a quarter of the time, ChatGPT answered a commercial buyer question entirely from what it already knew, citing no one. Gemini and Claude did that on fewer than 2% of answers.
For anyone trying to earn AI visibility, this splits the job in two. Against Gemini and Claude you are competing to be a retrieved source, which is won by being on pages those engines fetch. Against ChatGPT, a quarter of the time there is no retrieval step to win, and presence has to come from the model already knowing the brand. Those are different problems with different levers, and no dashboard that reports a single AI visibility score separates them.
3. Engines cite a very different number of sources per answer
| Engine | Distinct domains per answer | Most in one answer |
|---|---|---|
| Gemini | 12.3 | 38 |
| ChatGPT | 7.0 | 13 |
| Claude | 5.0 | 12 |
Counted across answers that cited at least one source. Gemini cites about two and a half times as many distinct sites per answer as Claude, and it caps far higher. On raw odds of appearing at all, Gemini is the most winnable of the three and Claude the least.
4. Ask again a fortnight later and half the sources change
Because the panel repeats on a schedule, the same engine gets asked the same question by the same business on different dates. Comparing each later run against the earlier one, and counting what share of the domains cited on the later run were not cited before:
| Engine | Repeat pairs | Sources new on the later run |
|---|---|---|
| Gemini | 1,658 | 54.7% |
| ChatGPT | 1,219 | 51.7% |
| Claude | 1,607 | 46.3% |
Same question, same engine, same business, roughly half the sources are different. Nothing about the sites changed in between. The engines are simply not deterministic.
This is the number that should change how AI visibility is reported. A week-on-week movement in an AI visibility score is mostly the engine reshuffling, not your content winning or losing. If a tool shows you a line chart of your AI rank with weekly granularity, most of what you are looking at is noise, and so is most of the reassurance or panic it produces. Read these numbers over months and across engines, not week to week in one.
5. Authority is not the gate
We matched every cited domain against the Majestic Million, a free public ranking of the top million sites by referring subnets.
| Band | Domains | Share of domains | Citations | Share of citations |
|---|---|---|---|---|
| Not in the top 50,000 | 5,698 | 84.1% | 18,665 | 67.0% |
| Top 50,000 | 540 | 8.0% | 3,110 | 11.2% |
| Top 10,000 | 379 | 5.6% | 2,605 | 9.4% |
| Top 1,000 | 162 | 2.4% | 3,474 | 12.5% |
Five sites in six that these engines cited are outside the top 50,000 sites on the web, and two citations in three go to them. A further detail in the same direction: 3,490 of the 6,779 domains, 51.5%, were cited exactly once.
The top 1,000 sites are still overrepresented, taking 12.5% of citations from 2.4% of domains. Big sites do get cited more per site. But the bulk of AI citation is not going to household names, and a small site is not structurally shut out the way it is on a competitive Google SERP. This is the most encouraging finding here for anyone doing GEO on a small domain.
6. Reddit is a Gemini source, not an AI source
Splitting the best-known citation targets by engine:
| Domain | Gemini | ChatGPT | Claude |
|---|---|---|---|
| reddit.com | 297 | 0 | 0 |
| youtube.com | 186 | 37 | 0 |
| quora.com | 24 | 0 | 3 |
| medium.com | 144 | 72 | 25 |
| g2.com | 64 | 34 | 25 |
| en.wikipedia.org | 41 | 53 | 12 |
| techradar.com | 16 | 405 | 4 |
Reddit was cited 297 times across 162 different questions, every single time by Gemini, and never once by ChatGPT or Claude in this corpus. The advice to go and post on Reddit for AI visibility, which is now everywhere, is on this evidence advice about one engine.
TechRadar runs the other way: 405 of its 425 citations came from ChatGPT. These are not general-purpose citation targets. They are engine-specific ones, and picking the wrong one for your audience means working a channel your buyers' engine never reads.
7. A quarter of citations point at a homepage
By path depth of the cited URL:
| Depth | Citations | Share |
|---|---|---|
| Homepage or one level deep | 7,112 | 25.5% |
| Two levels | 14,444 | 51.9% |
| Three levels | 4,036 | 14.5% |
| Four or more | 2,262 | 8.1% |
Most citations are article-depth pages, which is what you would expect. But a quarter point at a homepage or a top-level page, and when we looked at those they are largely the product sites being recommended rather than the publisher doing the recommending. Being named in a roundup gets your own homepage cited, not only the roundup.
What this means if you are choosing a GEO tool
Three of these findings bear directly on the GEO tools now being sold.
| Common claim | What the data says |
|---|---|
| Track your AI visibility score | A single score across engines hides an 80.6% disagreement rate. Ask for it per engine or it is not telling you much. |
| Weekly AI rank tracking | Roughly half the cited sources turn over between runs with no change on your side. Weekly granularity mostly reports engine variance. |
| One engine as a proxy for the rest | 5.2% of sources are shared by all three engines. There is no proxy. |
| You need authority before AI cites you | 84.1% of cited domains are outside the top 50,000 sites. |
None of this makes monitoring worthless. It makes the reporting period longer and the per-engine breakdown mandatory. If you want a free starting point you can run today, our AI Mention Checker shows what the engines currently say about your brand.
Being cited starts with being on the page that gets cited
Two citations in three go to sites outside the top 50,000. Those are reachable. MentionAgent finds the blogs the engines are already citing in your niche and pitches you for a mention on them. Start free, or read how the agent works.
Limits of this study
Stated plainly, because they change how far you should take the numbers.
- The questions are commercial and English. They are what 23 businesses want to be found for, which skews toward buying-intent questions and away from research, news and non-English queries.
- Three engines, not all of them. Perplexity, Copilot, Grok and Google AI Overviews are not in this corpus. Given how little the three studied engines agree with each other, do not assume these findings transfer to them.
- Majestic Million rank is a popularity measure, not Ahrefs DR or Moz DA. It ranks by referring subnets. A domain being absent from it means it is outside the top million by that measure, which is missing data rather than a proven low score.
- Six weeks. 8 July to 17 August 2026. Engine behaviour changes; we watched Copilot citations to our own site fall by 93% overnight in July with nothing changed on our end. Treat every figure here as dated.
- Citations are what the API returned. If an engine used a source without disclosing it, we cannot see that.
Get the data
The 250 most-cited domains, with per-engine counts, the number of distinct questions each was cited for, and its Majestic Million rank where present:
Published under CC BY 4.0. Use it, chart it, argue with it. Please credit MentionAgent and link back to this page. We have not published the raw corpus because the questions belong to our customers.
Frequently asked
Which AI engine is easiest to get cited by?
Gemini, on these numbers. It cites 12.3 distinct domains per answer against ChatGPT's 7.0 and Claude's 5.0, it returned sources on 98.6% of answers, and it draws most heavily on sites outside the top 50,000. Claude is the hardest of the three: fewest sources per answer and the lowest turnover between runs.
How often should I check my AI visibility?
Monthly at the most frequent, and read the trend over several months. Roughly half the sources an engine cites change between two runs of the same question with no change on your side, so a weekly reading is dominated by that variance rather than by anything you did.
Do I need a high-authority domain to be cited by AI?
No. 84.1% of the 6,779 domains cited in this study are outside the Majestic Million top 50,000, and those domains took 67.0% of all citations. Large sites are cited more per site, but the majority of AI citation goes to small ones.
Does posting on Reddit help with AI visibility?
For Gemini, on this evidence, yes: Reddit was cited 297 times across 162 questions. For ChatGPT and Claude it was cited zero times in this corpus. Whether it is worth the effort depends entirely on which engine your buyers use.
Can I reproduce these numbers?
The aggregate tables and the top 250 cited domains are downloadable above. The underlying corpus is 374 questions belonging to 23 businesses, so it is not published. Every figure on this page is a count over the source URLs the three engine APIs returned between 8 July and 17 August 2026.