Cross-Engine Citation Agreement: Which Sources All AI Engines Cite

Among sources AI answer engines cite often enough to measure, the engines largely agree. The Machine Relations Index grades 508 of 22,377 cited domains A, B or C across six engines — ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, and Perplexity — and no graded domain is cited by a single engine. Among them, the closest engine pair, Claude and Perplexity, shares 90.6% of its source set. Across the whole index, 67.4% of cited domains are observed on only one engine, and that figure is set by the long tail: 46.2% of the index was cited exactly once in 128 days, and a domain cited once is single-engine by arithmetic.

Correction, September 26, 2026. Gemini is a third engine with dated collection gaps, not a continuously collected one: 22 blind days inside the scoring window (May 11–15, then scattered single days from August 11 to September 16), carrying 442 eligible successful runs that produced no citations. The intermittent shape is why it went unnoticed while the two Google outages were found. Three arms collected without a gap — ChatGPT, Claude and Perplexity — and three are gapped: Gemini, Google AI Mode and Google AI Overviews. (An earlier form of this note, posted the same day, named only ChatGPT and Perplexity. Claude's single flagged date carried one answer run and was an artifact of a zero-citation scan with no run floor, not an outage.) Every Gemini count on this page is therefore a floor and every Gemini absence a ceiling, and no ranking that places another engine above Gemini is stated here. Rankings that place an engine below Gemini still hold, because a floor above a fixed count can only move further above it.

Engine Agreement Rises With Evidence #

Release mri_score_v2.0+2026-09-21+d73e239e7c5b, window May 10 through September 21, 128 observed days, 126,924 source events, 16,146 observed answer runs, 22,377 cited domains, 113,112 domain-run citations.

On this release, 10,334 of the 22,377 cited domains — 46.2% of the index — were cited exactly once across the whole 128-day window. A domain cited once can be attributed to exactly one engine; there is no second citation available to come from a second engine. Those 10,334 domains are single-engine by arithmetic, not by retrieval behaviour, and they are 68.5% of the single-engine population. Together they carry 9.1% of all citation volume.

Condition the count on evidence and the engines converge:

Source: Machine Relations Index v2, release mri_score_v2.0+2026-09-21+d73e239e7c5b. The engine count is whole-window per-engine observation in the v2 public release — which engines were observed citing the domain — not the retired v1 composite component of the same name. The release does not publish per-engine citation counts per domain.

At ten citations — still a low bar, roughly one citation every thirteen days — single-engine sources are 1.5% of what remains. At thirty they are none.

This is why the Index reads the number of engines observed citing a domain against evidence depth. Counting how many distinct things a sample contains, when most of them appear once, is the singleton problem in species-richness estimation: the count of once-observed species dominates the tally and is driven by sampling effort rather than by what is actually out there, which is why estimators such as Chao's correct for it and why Good–Turing frequency estimation assigns the singletons to unseen mass instead of treating them as findings. An AI-citation index is the same shape of data. Any study that reports "what share of cited sources appear in only one engine" without stating how often those sources were cited is largely reporting its own sampling depth.

The exclusive pools sit in the tail. Computed over the whole index and over domains with at least ten citations:

† Floors, not measurements. Both Google surfaces have collection gaps inside this release's window — see the coverage note below — and missing days can only shrink an exclusive pool. Neither Google row is ranked against the others here.

Two engines have no exclusive source at all once a source has to have been cited ten times, and both of them are gapped arms, which is why that is reported and not explained. The largest remaining pool is eighteen domains.

What the Index's own grade says #

The Machine Relations Index grades a domain A, B or C when enough evidence stands behind it, and marks it collecting otherwise. On the September 21 release, 508 of 22,377 domains carry a grade. The lowest citation count among them is 30. The highest citation count of any single-engine domain anywhere in the index is 27.

The two populations do not touch. No graded domain in the Machine Relations Index is cited by a single engine. The evidence floor and the number of engines observed citing a domain are, empirically, the same filter.

Among those 508 graded domains, pairwise overlap is not thin. Measured as the shared share of two engines' combined sets:

† Both Google columns understate their engines and neither should be read as its overlap. The Index recorded no citations from Google AI Overviews on any monitored query between August 15 and September 24, 2026, covering 37 days of this release's 128-day window, and none from Google AI Mode between September 14 and September 24, covering 7 days of it. Missing days can only remove domains from an engine's set, and removing domains can only lower its shared share with every other engine. Every Google AI Overviews and Google AI Mode figure in the table above is therefore a lower bound, and the same holds for their rows and columns on any release generated during those periods. The cause in both cases was a defect in the Index's own Google collection path, replaced on September 25; the consequences are set out in How Many Sources Each AI Engine Cites.

Read across ChatGPT, Claude, Gemini and Perplexity, the furthest pair in the graded set is recorded as ChatGPT and Gemini at 66.5% — a pair whose figure cannot be signed, because Gemini's blind days remove citations from both the shared set and Gemini's own set and the net direction is not determinable. Every pair in the matrix that includes Gemini carries the same caveat. The furthest pair between two exactly measured arms is ChatGPT and Perplexity, and even that furthest recorded pair is still more than twice as close as the closest pair in the unconditioned index. Two earlier forms of this paragraph named a furthest pair with a gapped arm in it: first ChatGPT and Google AI Overviews at 57.3%, then ChatGPT and Google AI Mode at 64.2%. Both readings are withdrawn. A pair cannot be called the furthest when one side of it is a lower bound, and the rule that follows is the general one: rank only the arms whose collection ran through the whole window. Claude's graded set of 470 domains and Perplexity's of 498 intersect on 460.

What this means for source strategy #

Measuring one engine samples a shared source set. By domain count, a single engine sees a minority of the index. By citation volume, and among sources with enough evidence to be graded, one engine sees most of what the others see. The cost of measuring one engine is real, and it is a sampling cost: a brand's own citation count is small, and at small counts one engine is a bad estimator of six. That is the reason to measure all six.

Measurable AI visibility wins are portable. A source cited often enough to clear the evidence floor is cited by a mean of 5.17 engines. Once a source is cited ten or more times, Perplexity holds 18 domains no other engine cites; the two Google surfaces hold none, which is what a gapped arm would also produce and so is not read here as an engine property. Citation that does not travel is citation that has happened once.

Cross-engine consensus correlates with source role clarity. Domains that AI engines can categorize — editorial media, market databases, analyst research, vendor-owned authorities — achieve 6-engine citation at 3-4x the rate of uncategorized sources. The MRI's source-role taxonomy shows this pattern across every measured category. Vendor-owned sources are observed on 2.55 engines on average against 1.54 for uncategorized ones.

Real engine disagreement lives among heavily cited sources. On the September 21 release reddit.com is the most-cited domain in the index at 2,021 cited runs, carries an A grade, and is still absent from ChatGPT and Claude across the whole 128-day window. It is one of 105 graded domains missing two or more engines — nytimes.com is on three of six, facebook.com and techradar.com on four. That is the population worth studying: 105 domains, not 15,080.

Engine set sizes differ. On the September 25 release, mri_score_v2.0+2026-09-25, Perplexity cites 10,667 distinct domains of 23,872 and ChatGPT 5,355 — both exactly measured arms, so this factor of 1.99 is a measurement rather than a bound, which holds at 1.97 among domains cited ten or more times. Both Google surfaces are excluded from those figures and read as floors: Google AI Overviews at 2,304 and 1,025 after 41 days of the window with no recorded citations, Google AI Mode at 7,458 and 1,585 after 11. Excluding Google AI Mode moves neither factor, because Perplexity and ChatGPT bound the range either way. Earlier forms of this paragraph put the spread at 4.2 and then 4.6 with Google AI Overviews as the low end, and reported it narrowing under conditioning; those figures are withdrawn, because the narrowing was that collection gap closing arithmetically rather than tail sampling falling away. The gap and its consequences are set out in How Many Sources Each AI Engine Cites, which also withdraws an earlier per-engine citation-volume ranking this index cannot support. The companion overlap study reports the same conditioned measurement for the May–July window.

Whole-Index Counts Across Three Releases #

Counting every cited domain once, whatever its frequency, gives the whole-population view. It is dominated by the once-cited tail, and it is stable across releases for that reason.

July 18 Release #

Of 14,680 domains the Machine Relations Index measured during its 66-day observation window, the distribution of how many engines cite each domain breaks down like this:

Counted this way, nearly two-thirds of all cited domains are observed on a single engine, and fewer than 5% on five or more.

External studies that count the same way report the same whole-population shape. Temso AI's analysis of 2 million answer-engine citations found that 71% of sources appear in only one model's responses, and even the two most overlapping engines agreed on just one in five cited sources. SatelliteAI's cross-engine verification study found citation volumes for the same brand differing by a factor of 615x across platforms.

The Same Count, 63 Days Later #

A single measurement of a structural pattern is a snapshot. This one was re-run on the release of September 19, 2026 — mri_score_v2.0+2026-09-19+0cad03121f60, window May 10 to September 19, 126 observed days, 125,115 source events, 15,883 observed answer runs, 22,213 cited domains. The universe grew by 7,533 domains. The structure did not move.

Two months of additional observation, a 51% larger measured universe, and every band is within a percentage point of where it was. That stability is a property of how the universe grows: each release adds thousands of thin, once-cited domains that preserve the band shares, which is why the Index reads the count against evidence depth.

The concentration is sharper than the domain counts suggest. On the September release, the 347 domains cited by all six engines hold 24,623 of 111,510 domain-run citations — 22.1% of all citation volume sits with 1.6% of domains — while the 14,985 single-engine domains hold 23,901, or 21.4%. That is 71 citations per full-consensus domain against 1.6 per exclusive one, a 44x difference in what a single domain earns.

What did move is which engine owns the exclusive pool. In July, Google AI Mode held the largest set of domains no other engine cited. In September, Perplexity does:

† The September 19 column for both Google surfaces is a floor. It sits inside the Google AI Overviews blackout that began August 15 and, for Google AI Mode, five days into the blackout that began September 14. A July-to-September change on either row is not readable as engine behaviour.

Perplexity's exclusive pool more than tripled and ChatGPT's nearly doubled: both engines were collected throughout, so both movements are the engines'. Google AI Mode's barely moved, and that is exactly what a collection gap produces, so no reading is offered for it. Among the continuously collected engines, the retrieval-first products are widening the long tail they reach. These whole-index pools are mostly once-cited domains; conditioned on ten or more citations they nearly vanish, as the table above shows.

Across the whole unconditioned index, overlap between any two engines is thin. Measured as the shared share of two engines' combined domain sets on the September 19 release, the closest pair is Claude and Perplexity at 23.3%, and ChatGPT and Claude share 1,223 domains of 8,819 combined, 13.9%. An earlier form of this sentence named ChatGPT and Google AI Overviews at 8.6% as the furthest pair; that reading is withdrawn, because that release sits inside the Google AI Overviews blackout and a gapped arm is a lower bound, not a rank. Among the four continuously collected engines on the current release, the thinnest pair is ChatGPT and Gemini at 10.2% and the closest is Claude and Perplexity at 22.8%.

Among the 508 graded domains, the same pairs measure 66.5% to 90.6%, as the matrix above shows.

Exclusive Pools, July 18 Release #

The 9,749 single-engine domains reveal which engines pull from the most distinct source pools:

Exclusive-pool sizes counted over every domain are almost entirely once-cited domains, so they describe how far into the long tail each engine reaches rather than a set of sources one engine trusts and the others reject.

SurfacedBy's study of 127,198 citations across five engines found similarly low overlap rates, and arXiv research on GEO quality factors showed that engines weight different page-quality signals — metadata freshness, semantic HTML, and structured data were the dimensions with the strongest engine-to-engine disagreement.

Citation Rate by Engine Count, July 18 Release #

Average citation rate rises with the number of engines observed citing a domain:

A domain cited by all six engines averages a 74x higher citation rate than a single-engine domain. A domain's engine count cannot exceed its citation count, so the engine-count column and the citation-rate column are mechanically coupled at the low end; the table is a description of the measured universe rather than evidence that adding engines raises frequency. Among sources that clear the evidence floor, cross-engine presence is close to universal, so engine count stops discriminating between them.

Which Sources Achieve Full Consensus? #

The 237 domains cited by all six engines are not randomly distributed across the web. They skew heavily toward classified, high-authority source types:

By contrast, among single-engine domains, 82.8% are uncategorized sources — pages that have not established a recognizable source role in the citation ecosystem.

The top 6-engine domains by citation rate include LinkedIn (8.49% citation rate), Medium (6.78%), Gartner (4.65%), G2 (4.12%), and Crunchbase (3.35%). These are large platforms with established content across multiple industries and question types — not niche publications optimizing for a single engine.

A telling counterexample: Reddit, the single highest-cited domain in the entire index at a 10.57% citation rate, is cited by only four engines. ChatGPT and Claude rarely surface Reddit in their responses. The most-cited source on the internet is not a consensus source.

That counterexample held through the September re-measurement. On the September 19 release, reddit.com ranks #1 of 22,213 domains at a 12.67% citation rate — 2,012 cited runs of 15,883 observed — and its engine list still holds four names: Gemini, Google AI Mode, Google AI Overviews, and Perplexity. It was cited in zero ChatGPT runs and zero Claude runs across the 126-day window. Two of six engines have not cited the most-cited domain on the open web across the whole measured window, May 10 through September 19.

Machine Relations Measurement Context #

The Machine Relations Index tracks citation rates across six engines specifically because cross-engine agreement is the harder, more informative signal. A domain's MRI citation rate — the percentage of observed answer runs in which it was cited — captures both frequency and cross-engine reach in a single measurement. Sources with high citation rates and high engine counts represent the strongest form of AI source authority: multiple independent retrieval systems, built by different companies with different architectures, independently choosing to cite the same domain.

The data window for this analysis: May 10 through July 17, 2026 — 66 observed days, 69,030 source citation events, 14,680 measured domains. Methodology: MRI Score v2.0. Citation rates are published only for domain-segment combinations that clear the evidence floor of 10 observations across 7 distinct run dates.

The re-measurement window: May 10 through September 19, 2026 — 126 observed days, 125,115 source events, 15,883 observed answer runs, 22,213 measured domains, release mri_score_v2.0+2026-09-19+0cad03121f60, all six engines reported healthy. Same methodology version, same evidence floor. The conditioned measurement window: May 10 through September 21, 2026 — 128 observed days, 126,924 source events, 16,146 observed answer runs, 22,377 measured domains, 113,112 domain-run citations, release mri_score_v2.0+2026-09-21+d73e239e7c5b. Same methodology version, same evidence floor. The difference is not the window but the conditioning: the number of engines observed citing a domain is read against how many times that domain was cited, rather than over every cited domain equally. The engine count used throughout is v2 per-engine observation recorded per domain, not the retired v1 weighted composite, which was superseded on 2026-07-05. v1 scores are not comparable to v2 citation rates.

All three sets of figures are on this page because together they show how a whole-index statistic can reproduce across releases while the conditioned measure tells a different story.

Consensus Over a Window Is Not Consensus Inside an Answer #

Conditioning on evidence shows engines agree over a window. That is a different claim from agreement inside a single answer.

Every figure on this page counts a domain once for the whole observation window. It asks whether six engines were ever seen citing a domain across 128 days. A brand's question is narrower and harder: when a buyer asks one question, do two engines cite the same source in their two answers? That is a co-occurrence measurement, and no whole-window aggregation can produce it, because the aggregation has already dissolved the unit where co-occurrence happens — the answer.

A fixed daily panel can. Over 85 dates, 69 fixed queries and five engines, 57 third-party domains were cited by all five at some point, which is the consensus set this page's evidence floor selects for. Those 57 domains 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: 0.13% of appearances. More than one engine did so on 27%. Restricted to domains with ten or more citations, the same measure is 16.1%.

Both readings are right about their own question. Conditioned on evidence, cross-engine citation is common over a window. Measured inside a single answer it stays rare — 13.5% of the domains cited for a question appear on more than one engine, and ChatGPT and Gemini cite nothing in common on 84.2% of questions. The evidence floor identifies which sources carry weight. It does not make the engines agree at the moment a buyer is reading. Full measurement: The 57 Sources Every AI Engine Cites Were Cited Together 9 Times in 7,133.

FAQ #

What percentage of AI-cited sources are exclusive to one engine? #

Counting every cited domain equally, 67.4% of the 22,377 domains on the September 21, 2026 release appear in only one engine's responses, close to the 66.4% first measured in July. That figure is set by the long tail: nearly half the index was cited exactly once across 128 days, and a domain cited once cannot be cited by two engines. Require ten citations and the single-engine share is 1.5%; require the evidence the Index itself grades on and it is zero of 508.

Which AI engine cites the most unique sources? #

Counted over every domain, Perplexity, with 4,308 domains no other measured engine cited on the September 21, 2026 release. Gemini's 2,842 and ChatGPT's 2,741 are recorded next, but Gemini's blind days move an exclusive-pool count in both directions at once — removing its own exclusives while manufacturing exclusivity for whichever engine still cited a domain it missed — so that pair is reported rather than ranked. Google AI Mode reads 3,371 and Google AI Overviews 364 on the same release, but both are floors rather than ranks: the Index's own collection recorded no citations from Google AI Overviews after August 15 or from Google AI Mode after September 14, and missing days can only shrink an exclusive pool. Counted over domains cited at least ten times, the pools nearly vanish: Perplexity holds 18, ChatGPT 12, Gemini and Claude 2 each, and both Google surfaces none, which a collection gap would also produce. The exclusive pools measure thin observation, not distinct retrieval taste.

Do different AI engines cite different sources? #

Over a whole observation window, less than the headline figure suggests; inside a single answer, very much so. Among the 508 domains the Machine Relations Index grades A, B or C on the September 21, 2026 release, and counting only the arms with enough collection to report at all, of which three — ChatGPT, Claude and Perplexity — ran without a gap through the whole window, the closest engine pair — Claude and Perplexity — shares 90.6% of its combined source set, and the lowest recorded pair — ChatGPT and Gemini — shares at least 66.5%, a floor and not a ranking, because Gemini has collection gaps inside the window too. Neither Google surface is ranked here, and neither is Gemini: all three have collection gaps inside the window, so their overlap figures are lower bounds. The disagreement above the evidence floor is specific and small: 105 graded domains are missing two or more engines, reddit.com among them. That is the window view. Measured answer by answer on a fixed five-engine daily panel, only 13.5% of the domains cited for a given question appear on more than one engine, and the 57 domains all five engines cite were cited by all five for the same question on the same day nine times in 7,133 appearances. See consensus inside an answer.

Does being cited by more engines increase citation frequency? #

The two move together, but the question assumes a direction the measurement cannot supply. Domains cited by all six engines average a 0.74% citation rate against 0.01% for single-engine domains, and the relationship is roughly linear on a log scale. A domain's engine count cannot exceed its citation count, so a large part of that gradient is the citation-count distribution restated — 46.2% of cited domains were cited once on the September 21, 2026 release, and a once-cited domain is single-engine by arithmetic. Read the gradient as a description of the universe, not as evidence that adding engines raises frequency. See engine agreement rises with evidence.

What source types achieve cross-engine citation most often? #

Among 6-engine domains, editorial media (19%), vendor-owned sites (31%), and analyst/market research sources (13%) are heavily overrepresented compared to single-engine domains, where 83% are uncategorized sources without a clear source-role signal.

Updated 2026-09-23: figures reflect the current Machine Relations Index methodology; see the MRI methodology and update log.