The 57 Sources Every AI Engine Cites Were Cited Together 9 Times in 7,133
Across 85 days, 69 fixed queries and five AI engines, 57 domains were cited by all five. On the same question, on the same day, all five cited one of them 9 times out of 7,133 appearances — 0.13%. Corpus-level consensus and answer-level consensus are different measurements, and only one of them is a buying decision.
Ask five AI answer engines the same question on the same day, and note every source they cite. Do that 2,781 times over 85 days. Then find the domains that every engine cited at some point in the window — the consensus set, the sources you would call safe.
There are 57 of them. They appear on 7,133 question-days between them. All five engines cited one of them for the same question on the same day nine times.
Nine. Out of 7,133. That is 0.13%.
Both of the following are true of the same dataset, and the distance between them is the whole problem with how AI visibility is currently measured:
- Over the window: 77% of the domains with real citation volume were cited by more than one engine. Consensus looks common.
- Within an answer: 13.5% of the domains cited for a given question were cited by more than one engine. Consensus is rare.
Same citations. Same engines. Same 85 days. The only thing that changes is whether you count a domain once per window or once per answer. One of those numbers describes a retrieval system. The other describes what happens when a buyer asks a question — and it is the second one that decides whether being cited on Perplexity does anything for you on ChatGPT.
The instrument
This is a fixed-panel measurement, not a crawl. Every day since 2026-06-24, the same query set has been put to the same engines and every cited URL stored.
| Field | Value |
|---|---|
| Observation window | 2026-06-24 to 2026-09-22 |
| Runs | 89 across 85 dates |
| Distinct queries | 69 |
| Retrieval engines | 5 (ChatGPT, Claude, Gemini, Google AI Mode, Perplexity) |
| Question-days after de-duplication | 2,781 |
| Question-days with two or more engines citing | 2,704 |
| Distinct third-party domains cited | 3,217 |
| Mean citations returned per answer | Perplexity 6.0, Gemini 4.6, ChatGPT 3.5, Claude 3.5, Google AI Mode 3.2 |
Four dates carry two runs; the second run of each is dropped. Every number in this piece excludes our own six domains, because the query set is drawn from our own displacement register and our pages are the subject matter — leaving them in raises apparent agreement from 13.5% to 23.0% and would be measuring ourselves.
These are citations: anchored URLs attached to an answer. They are immune to the failure class we audited and published yesterday, where an unanchored string match counted a model saying it had never heard of a brand as evidence the brand was present. A cited URL either is or is not in the response. Nothing here rests on reading a name out of prose.
The same panel produced the per-engine citation churn measurement published this morning, which measures the other axis: whether one engine keeps citing the same sources from day to day. This piece measures whether different engines cite the same sources as each other on one day.
Answer-level: the engines barely intersect
For each question-day, take the set of domains each engine cited and compare them.
| Engine pair | Mean shared share of combined sources | Question-days sharing nothing at all |
|---|---|---|
| Claude × Perplexity | 10.7% | 54.2% |
| Google AI Mode × Perplexity | 10.4% | 46.5% |
| Gemini × Google AI Mode | 7.5% | 57.2% |
| ChatGPT × Claude | 7.2% | 72.7% |
| Gemini × Perplexity | 6.6% | 62.1% |
| Claude × Gemini | 6.6% | 65.7% |
| Claude × Google AI Mode | 6.2% | 65.8% |
| ChatGPT × Perplexity | 5.4% | 73.6% |
| ChatGPT × Google AI Mode | 3.1% | 81.1% |
| ChatGPT × Gemini | 2.7% | 84.2% |
The best-agreeing pair in the panel shares about a tenth of its sources and cites nothing in common on more than half of all questions. The worst pair — ChatGPT and Gemini, the two engines with the largest consumer footprints — cites nothing in common on 84.2% of questions.
Shared-share arithmetic punishes engines that return different numbers of sources, so here is the same comparison as containment: of the smaller of the two source sets, how much does the larger one also cite? Google AI Mode and Perplexity 20.9%, Claude and Perplexity 20.8%, down to ChatGPT and Gemini at 6.3%. The correction moves the numbers and not the finding.
Per engine, the share of its cited domains on a given question that no other engine cited that day:
| Engine | Same-day exclusive share |
|---|---|
| ChatGPT | 83.2% |
| Gemini | 78.0% |
| Perplexity | 71.3% |
| Google AI Mode | 70.7% |
| Claude | 68.0% |
At the URL level rather than the domain level, 10.1% of cited pages appear on more than one engine.
This part is not new, and it should not be sold as new. Wellows' study of 22.7 million citations across five engines found 79.6% of sources cited on a question appear on exactly one engine and 0.31% on all five, with ChatGPT the most exclusive at 76.3%. Writesonic's 161,286-prompt study found 3.8% of sources shared by all four engines it tested and pairwise Jaccard between 0.119 and 0.237, with ChatGPT and Gemini bottom of its table too. Temso's analysis of two million answer-engine citations found 71% of sources appear in only one model's responses. Frase's summary of BrightEdge's analysis puts pairwise top-100 source overlap at 16% to 59%. Our panel runs at or below those figures — a narrower query set and smaller citation lists will do that — and reproduces the ordering exactly, including ChatGPT as the outlier. Four independent instruments, one finding.
The mechanism is not a mystery, and the platforms describe it themselves. Google's AI features documentation makes eligibility for AI surfaces a property of its own index and snippet rules, which is why two Google products can behave differently from each other. Each engine runs its own crawler, its own index and its own source-selection step, as the per-engine source-selection comparison sets out. Different pipelines produce different evidence. The surprise is not that agreement is low; it is how confidently the market prices a single score on top of it.
Corpus-level: the same data says the opposite
Now count each domain once for the whole window, the way a citation index does.
| Minimum citations in the window | Domains | Cited by one engine | Mean engines per domain |
|---|---|---|---|
| 1 (all) | 3,217 | 71.8% | 1.49 |
| 2 | 1,771 | 48.7% | 1.89 |
| 3 | 1,359 | 40.8% | 2.09 |
| 5 | 969 | 30.7% | 2.37 |
| 10 | 613 | 23.0% | 2.74 |
| 20 | 378 | 15.3% | 3.12 |
| 50 | 159 | 6.3% | 3.78 |
44.9% of the domains in this panel were cited exactly once in 85 days. A domain cited once can only be attributed to one engine — there is no second citation available to come from a second engine. Strip that tail out and single-engine exclusivity falls from 71.8% to 23.0%, and at fifty citations it is 6.3%.
That is a real and important correction, and it is the one the Machine Relations Index published on September 21 against its own headline finding, on a larger corpus: 46.2% of 22,377 cited domains were cited exactly once, and conditioning on an evidence floor took single-engine share from 67.4% to 1.5%. This panel reproduces that shape independently.
It is also, read alone, an invitation to conclude that the engines mostly agree. They do not.
The bridge
Group the domains by how many engines ever cited them, then ask how often more than one engine cited them on the same question, on the same day.
| Engines that ever cited it | Domains | Question-day appearances | Cited by more than one engine that day |
|---|---|---|---|
| 1 | 2,309 | 7,089 | 0.0% |
| 2 | 493 | 4,809 | 4.0% |
| 3 | 215 | 5,222 | 10.8% |
| 4 | 143 | 8,957 | 19.5% |
| 5 | 57 | 7,133 | 27.0% |
A domain that every engine in the panel has cited is still cited by a single engine alone on 73% of the occasions it shows up. And the strictest version of the question — all five engines citing it for the same question on the same day — happens nine times in 7,133 appearances.
Applying a citation floor instead of an engine count gives the same answer. Domains with at least ten citations are co-cited by more than one engine on 16.1% of their appearances; at fifty citations, 21.5%. The floor selects domains that are heavily cited. It does not select domains that are cited together.
The consensus set is real, and it is worth having: those 57 domains are 1.8% of the panel's domains and carry 24.8% of all citations. But membership in it means each engine reaches for that source on its own schedule, not that the engines reach for it at the same moment. Even inside the set, the per-engine rates diverge by an order of magnitude or more — medium.com was cited 196 times by Gemini and 4 times by ChatGPT; searchengineland.com 117 times by Perplexity and 5 by ChatGPT; muckrack.com 126 by Gemini and 5 by ChatGPT.
What this bounds
The evidence-floor correction says: stop counting thin sources as proof that engines disagree. Correct.
It does not say: engines agree. On the same question on the same day, they do not, and the floor does not change that. Conditioning on citation volume tells you which sources are substantial. It tells you nothing about co-occurrence, because it has aggregated away the only unit where co-occurrence exists — the answer.
The two measurements answer different questions:
| Question | Measurement | This panel |
|---|---|---|
| Which sources carry real weight in this market? | Corpus-level, with an evidence floor | 613 domains at ten-plus citations; 57 cited by all five |
| Will a citation on one engine show up on another for the same query? | Answer-level co-occurrence | 13.5% of sources; 0.13% for all five |
| Is one engine a usable sample of the others? | Per-engine exclusivity | No: 68% to 83% of each engine's sources are its own that day |
An index ranking sources by citation weight is doing the first job and should not be read as doing the second. A vendor quoting a single cross-platform "AI visibility score" is implicitly claiming the second, and the second is the one that fails.
What an operator does with this
Do not buy one number. A single AI visibility score averages five systems that share roughly a tenth of their evidence on any given question. The average describes no engine.
Pick the engine that matches the decision. If your buyers arrive through ChatGPT, measure ChatGPT — it is the most exclusive engine in every study including this one, which means it is both the least substitutable and the least informative about the others.
Use the consensus set as a floor, not a plan. The 57 domains every engine touches are where earned coverage has the best chance of being seen more than once. That is worth aiming at. It is not a guarantee of simultaneous presence, and 0.13% is what simultaneous presence actually costs.
Judge a source by its own citation record. "Cited by AI" is not a property a domain has. Per engine, per question shape, per week — that is the resolution at which the data is real, and it is the resolution the selection mechanics themselves operate at.
Limits
Sixty-nine queries, all within AI visibility, earned media, and PR measurement. This is a claim about this topic space, not about the web.
One snapshot per query per engine per day. A 19,556-query repeat-sampling analysis reports engines are only 58% to 65% self-consistent when asked the same question three times, so some of the disagreement measured here is within-engine noise rather than between-engine difference. That cuts against the finding and is not corrected for; the external studies at far larger prompt counts land in the same place, which suggests it is not doing much work.
Answers returning no citations are excluded from the overlap sets, which affects Google AI Mode most. Our own six domains are excluded throughout.
Direction over time is not clean. The share of sources seen by more than one engine per question ran 15.9% in June, 15.6% in July, 12.0% in August and 11.0% in September, while the all-engine share moved the other way, 0.3% to 1.4%. Wellows, over a longer window and a much larger corpus, reports average pairwise agreement rising from 8.49% in January to 11.25% in June. Four months on 69 queries is not enough to call a trend, and we are not calling one.
FAQ
Do AI engines cite the same sources for the same question?
Rarely. Across 2,704 question-days on five engines, 13.5% of cited domains appeared on more than one engine and 1.0% on all of them. The largest published studies agree: Wellows found 79.6% of sources on a question appear on exactly one engine, Writesonic found 3.8% shared across four engines.
Why do some studies say cross-engine agreement is high?
Because they count each source once for the whole measurement window instead of once per answer. Over 85 days, 77% of the domains in this panel with ten or more citations were cited by more than one engine. On the same question on the same day, the same domains were co-cited 16.1% of the time. Both numbers are correct; they answer different questions.
Which two AI engines agree least?
ChatGPT and Gemini. In this panel they share 2.7% of their combined sources per question and cite nothing in common on 84.2% of questions. Writesonic's much larger study also puts ChatGPT and Gemini bottom of its pairwise table.
Can one AI engine be used as a proxy for the others?
No. Between 68% and 83% of each engine's cited sources on a given question were cited by no other engine that day. Measuring one engine measures one engine.
What is Machine Relations?
Machine Relations is the discipline of making an organisation legible, credible, retrievable and citeable inside AI-mediated discovery systems, through earned authority, entity clarity, citation architecture, distribution and measurement. It was coined by Jaxon Parrott, founder of AuthorityTech, in 2024.