The answer engines the Machine Relations Index monitors draw on source pools of different sizes. On release mri_score_v2.0+2026-09-25, across ChatGPT, Claude, Gemini and Perplexity — of which ChatGPT, Claude and Perplexity were collected without a gap — Perplexity was observed citing 10,667 of the index's 23,872 cited domains and ChatGPT (API) 5,355 — a 1.99x spread over the same 132-day window and the same query population.
Both Google surfaces are reported separately and excluded from that spread, because the Index's own collection failed on each of them inside this window: no citations from Google AI Overviews on any monitored query for 41 of the window's 132 days, and none from Google AI Mode for 11. Their pools, 2,304 and 7,458 domains, are floors rather than comparable measurements. Those gaps, and why this page reports 1.99x rather than the 4.6x figure an earlier version of it published today, are set out in Collection Gaps in Two Engines.
That is a measurement of breadth: how many distinct sources an engine was observed drawing on. It is not a ranking of engines by citation volume, and the July version of this page published one.
Correction, September 25, 2026 #
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.
An earlier version of this page, dated July 25, 2026, was headed "Gemini Citations Outpace Every Other AI Engine" and reported that Gemini "accounts for 30.3% of citations to top-ranked market and analyst sources — more than double ChatGPT's 6.7%." That ranking is withdrawn. Three things were wrong with it.
The denominator was six domains. The 30.3% was an aggregate over G2, Crunchbase, Forbes, Deloitte, MarketsAndMarkets and Mordor Intelligence — 436 citations in total. Six hand-selected domains cannot order six engines, and the page presented the figure as a property of each engine rather than of those six rows.
A reader could not check it. The figure was attributed to the Machine Relations Index, and the Index's public release records which engines were observed citing a domain, not how many times each engine cited it. No reader opening the per-domain profiles the page pointed at could reproduce the per-engine split, and neither can this remeasurement. Per-engine citation counts per domain are not published.
Two downstream claims went further than any citation data can, and one of them inverted its own source. The page said that a placement in the wrong outlet "can carry 4.5x less answer influence" than the right one, and attributed it to a publication-level analysis by AuthorityTech, a related party to this site. That page reports an academic study's constructed influence proxy — 0.2713 for ChatGPT against 0.0584 for Google among successfully fetched pages — and states in its own text that the ratio "does not mean that buying or earning one placement produces four times the answer influence on one provider." The withdrawn sentence took the ratio the cited page expressly refused and credited that page with it. No engine publishes how it scores one source against another, and nothing in this dataset recovers it. The page also said brands appearing on MarketsAndMarkets reports "get zero ChatGPT referral traffic from that placement." The Index observes citations in answers. It does not observe referral traffic, and a domain below an engine's evidence floor has not been shown to be absent from it. Both claims are withdrawn.
One classification error travelled with them: the release classifies forbes.com as an editorial publication, and the withdrawn table listed it as analyst research.
What survives is the observation that started the page: the engines differ, and measuring one of them samples the others imperfectly. Below is that observation measured on the current release, in the unit the release actually publishes.
Source-Pool Breadth by Engine #
Release mri_score_v2.0+2026-09-25, window May 10 through September 25, 2026, 132 observed days, 131,670 source events, 16,788 monitored answer runs, 23,872 cited domains. Evidence floor: 10 observations across at least 7 distinct measurement dates.
Every domain in the index carries the set of engines observed citing it at least once in the window. Counting domains per engine gives each engine's observed source pool.
| Engine | Distinct domains observed | Share of the 23,872 cited domains | Domains cited 10 or more times |
|---|---|---|---|
| Perplexity | 10,667 | 44.68% | 2,150 |
| Google AI Mode ‡ | 7,458 | 31.24% | 1,585 |
| Gemini | 7,276 | 30.48% | 1,738 |
| Claude | 5,564 | 23.31% | 1,755 |
| ChatGPT | 5,355 | 22.43% | 1,094 |
| Google AI Overviews † | 2,304 | 9.65% | 1,025 |
† Not comparable to the other rows. The Index recorded no citations from Google AI Overviews on any monitored query between August 15 and September 24, 2026 — 41 of this window's 132 days. Its counts are a floor.
‡ Not comparable either. The Index recorded no citations from Google AI Mode between September 14 and September 24, 2026 — 11 of this window's 132 days. Its counts are a floor, and its position in this table is not a ranking.
Source: Machine Relations Index v2, release mri_score_v2.0+2026-09-25. Each figure is a count of domains whose public profile records that engine, read from the machine-readable release at https://machinerelations.ai/data/machine-relations-index.json. Google AI Mode and Google AI Overviews are counted as two engines because Google documents and supports them as two separate surfaces in its AI features guidance for site owners.
Two readings, and they are different questions. Both are taken over ChatGPT, Claude, Gemini and Perplexity; the endpoints of every spread quoted are ChatGPT and Perplexity, two of the three arms with gap-free collection — the third is Claude, so the spreads are measurements rather than bounds.
Whole-index breadth is dominated by sources cited once. 16,244 of the 23,872 domains — 68.05% — were observed on exactly one engine, and most of those were cited once in 132 days. A domain cited once can be attributed to exactly one engine; there is no second citation available to come from a second engine. Ranking engines on whole-index domain counts therefore partly ranks how deep each engine's tail was sampled.
Conditioning on evidence barely moves the spread. Restricted to the 2,290 domains cited ten or more times across the window, the range runs from Perplexity at 2,150 to ChatGPT at 1,094 — 1.97x, against 1.99x unconditioned. Both endpoints are exactly measured arms. The spread does not narrow. Perplexity and ChatGPT bound the range whether or not Google AI Mode is included, so excluding it changes neither figure; it changes what the figures may be called.
This is a correction to what this page said earlier today. It reported the spread narrowing from 4.6x to 2.1x and read that narrowing as evidence that the whole-index column measures tail depth. Both of those figures included Google AI Overviews, whose pool is depressed by 41 days of missing collection rather than by tail sampling, and almost all of the apparent narrowing was that gap closing arithmetically as the threshold rose. Among the engines actually collected throughout, there is no narrowing to explain.
One real reordering survives it: Gemini leads Claude on whole-index domains, 7,276 to 5,564, and that direction holds despite Gemini's blind days, because a floor above an exact count can only move further above it. Among domains cited ten or more times the recorded figures are 1,738 for Gemini and 1,755 for Claude — a 17-domain difference that is not read as a ranking either way, and now cannot be read as a tie either, since Gemini's 1,738 is a floor.
Neither column is a citation-volume ranking, and this page no longer publishes one.
Collection Gaps in Two Engines, and What They Cost This Comparison #
Both of the Google surfaces this Index monitors stopped returning observable citations inside this window, and neither stopped for the same length of time.
Between August 15 and September 24, 2026, the Index recorded zero citations from Google AI Overviews on every monitored query — 41 of the window's 132 days. Between September 14 and September 24 it recorded none from Google AI Mode — 11 days. Collection resumed on both lanes on September 25, the release this page is measured on: Google AI Overviews returned 51 citation URLs across the seven queries it completed of ten planned, and Google AI Mode returned 29 across ten of ten. The cause in both cases was a defect in the Index's own Google collection path, replaced on September 25; the prior collector graded responses as successes whether or not they carried content, and 111 of its last 181 Google AI Mode responses carried none. Citations reappeared immediately on the same query basket, which is why both gaps are treated as the instrument's and not as engines that stopped citing. The Index cannot distinguish those two states from the outside, and it does not claim to.
Three consequences, stated plainly because they bound every figure on this page.
Both Google source pools are floors. Google AI Overviews was collected on roughly 91 of the window's 132 days and Google AI Mode on roughly 121. A count of distinct domains accumulates, so missing days remove domains from a pool and can never add them. The 2,304 and 7,458 figures are minima, not measurements, and neither engine's position in the table above is a ranking.
Any spread computed against either of them is partly the gap. The 4.6x whole-index spread and the 2.1x conditioned spread that this page published earlier today used Google AI Overviews as the low end. Those figures are withdrawn. Computed between ChatGPT and Perplexity, two of the three arms that ran without a collection gap, the spread is 1.99x and 1.97x — unchanged by the exclusion of Google AI Mode, which sits inside the range rather than at either end.
The release manifest currently overstates both engines' health, and that is being fixed. The public release manifest for mri_score_v2.0+2026-09-25 reports every engine, including both Google surfaces, with "status": "healthy" and a last_success_date of 2026-09-25, under a coverage statement that all six engines reported successful runs within the observation window. Each of those is true as written and none of them is the question a reader needs answered: the lanes were reachable and returned successful responses throughout, and those responses contained no citations. A health check that asks whether a source responded cannot see a source that answers with nothing, and one engine went 41 days and another 11 in exactly that state without the manifest changing. The fix belongs to the manifest, not to this page, and until it ships the coverage statement should not be read as evidence that a between-release delta on either Google surface reflects engine behaviour.
Readers comparing this release against the September 18 or September 23 releases should treat every Google AI Overviews and Google AI Mode movement across those three releases as a collection artifact. Both of those editions disclose the two blackouts in their own coverage notes.
Why the Engine Count Rises With Evidence #
The same conditioning applies to how many engines observe a given domain.
| Minimum citations across the window | Domains | Observed on one engine | Observed on all six | Mean engines per domain |
|---|---|---|---|---|
| 1 — the whole index | 23,872 | 68.05% | 1.48% | 1.62 |
| 2 | 12,625 | 39.58% | 2.80% | 2.17 |
| 5 | 5,046 | 10.82% | 7.00% | 3.20 |
| 10 | 2,290 | 1.70% | 15.28% | 4.08 |
| 25 | 686 | 0.00% | 40.38% | 5.02 |
| 50 | 232 | 0.00% | 59.48% | 5.36 |
| 100 | 75 | 0.00% | 69.33% | 5.48 |
This relationship is partly mechanical and is stated that way deliberately: a domain observed more often has more opportunities to be observed by more engines. It is not evidence that appearing on more engines raises a source's citation rate, and no such claim is made here. What it does establish is that any published figure of the form "N% of AI-cited sources appear in only one engine" is reporting its own sampling depth unless it states how often those sources were cited. This is the singleton problem familiar from species-richness estimation: the count of once-observed members dominates the tally and tracks sampling effort rather than the population, which is why estimators of that kind correct for it rather than reporting it as a result. Independent datasets that count the same way report the same whole-population shape: Temso AI's analysis of 2 million answer-engine citations found 71% of sources exclusive to a single model, and GeoXylia's 500-site benchmark found engines citing different sources for the same query in 73% of cases.
The conditioned agreement measurement — pairwise overlap between engines at each evidence threshold, and what the Index's own grades say about it — is reported in full in Cross-Engine Citation Agreement, and is not repeated here.
Where the Engines Genuinely Disagree #
Real disagreement sits among heavily cited sources, not in the tail. Three examples from the current release, each checkable at its own public profile.
reddit.comis the highest-ranked domain in the index, cited in 2,063 of 16,788 monitored answer runs at Confidence A, on 126 of 132 observed days — and across the whole window it was observed on Gemini, Google AI Mode, Google AI Overviews and Perplexity, and not on ChatGPT or Claude. Its profile carries the figures.techradar.comis the reverse shape: cited in 464 of 16,788 runs at Confidence A, observed on ChatGPT, Claude, Gemini and Perplexity, and on neither Google answer surface — a non-observation rather than a measured absence, since both Google lanes carry blind days in this window. Its profile carries the figures.- 353 domains are observed on all six engines. They include
youtube.comat 1,452 cited runs,linkedin.comat 1,128,medium.comat 818,forbes.comat 686,nih.govat 493,g2.comat 476 andgartner.comat 463.
The population worth studying is that middle group — heavily cited domains missing one or two engines — not the 16,244 domains whose single-engine status is an artifact of being cited once.
What This Means for Source Strategy #
Measure all six, and expect the gap to be a sampling gap. A brand's own citation count is small. At small counts, one engine is a poor estimate of six, and the cost of measuring one is a sampling cost rather than proof of engine-specific behaviour.
Do not buy a per-engine strategy from a whole-population overlap statistic. The published overlap figures that circulate in this category are computed over every cited domain, most of which were cited once. They describe a tail, and a media plan built on them is a plan built on noise.
Treat a source absent from one engine as a question, not a finding, until it clears the evidence floor. An engine that has not been observed citing a domain may not have been observed enough times to say so. The Index marks a domain collecting rather than absent for exactly this reason: on this release 520 of 23,872 domains carry a confidence grade at all, 14 of them A.
Methodology and Limits #
The Machine Relations Index v2 measures source-segment citation rates across six answer engines — ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity — over buyer-intent query baskets. The September 25, 2026 release covers 131,670 source events across 23,872 cited domains and 16,788 monitored answer runs in a 132-day window opening May 10, 2026. A segment publishes a citation rate only after clearing an evidence floor of at least 10 observations across at least 7 distinct measurement dates; domains are graded A, B or C on the evidence behind them, and marked collecting otherwise.
Every figure on this page is read from the public release at https://machinerelations.ai/data/machine-relations-index.json or from the per-domain profile pages under https://machinerelations.ai/index/domains/, which any reader can open. Three limits bound what it can support.
The release publishes per-engine presence, not per-engine volume. For each domain it records which of the six engines were observed citing it in the window. It does not record how many citations each engine contributed. No table on this page splits citations by engine, and the withdrawn one above could not be checked against this release.
Presence is bounded below by observation. An engine missing from a domain's set means that engine was not observed citing it in this window at this sampling depth. It is not a statement about the engine's index or its policy toward the source.
The engine count is an observation, not a score. It is the set of engines seen citing the domain in the whole window. It is not the retired v1 scoring component of a similar name, and nothing here treats it as a quality signal or as a predictor of citation frequency.
Other published datasets measure adjacent things on different populations and are not interchangeable with this one: Profound's 377,765-citation dataset reports per-engine shares of its own tracked citations, LLM Pulse's 5.3 million-citation set covers 470,380 answers, Attrifast's vertical study reports unique domains per answer, Presenc AI's 2,400-brand benchmark reports citation rates by content type, Foglift's Q2 2026 benchmark reports first-party share per engine, QuickSEO's cross-engine analysis compares publications cited per engine, The Boring SEO's comparison describes response construction, AI Citation Monitor's engine comparison describes retrieval integration, and Searchless Journal's 500-query benchmark uses a separate multi-engine framework. Each has its own prompt set, window and counting rule. Figures are not comparable across them, and none of them is the source of any number in the tables above.
FAQ #
Which AI engine cites the most sources? #
On release mri_score_v2.0+2026-09-25, Perplexity was observed citing the most distinct domains: 10,667 of the 23,872 in the index, against 5,355 for ChatGPT (API), the lowest of any arm in the index — a ranking no collection gap can touch, since every other count is either exact and larger or a floor already above it. Restricted to domains cited ten or more times, Perplexity leads at 2,150 and ChatGPT trails at 1,094. Google AI Overviews is reported at 2,304 and 1,025 and Google AI Mode at 7,458 and 1,585, but neither is ranked against the others, because the Index recorded no citations from Google AI Overviews for 41 of the window's 132 days and none from Google AI Mode for 11. That is source-pool breadth. This release does not publish per-engine citation volume, so no engine is ranked here by how many citations it produced.
Did Gemini outpace other engines on citation volume? #
That claim appeared in an earlier version of this page and is withdrawn. It was computed from 436 citations across six domains, and the per-engine split behind it is not published in the Index release it was attributed to.
Why are most cited domains observed on only one engine? #
Because most of them were cited once. 68.05% of the 23,872 cited domains are single-engine across the whole index, and a domain with one citation can only be attributed to one engine. Among domains cited ten or more times, the single-engine share is 1.70%; above twenty-five citations there are none.
Is a source missing from an engine proof that the engine will not cite it? #
No. It means the engine was not observed citing that source in this window at this sampling depth. The Index marks thinly observed domains collecting rather than absent, and only 520 of 23,872 domains on this release carry a confidence grade at all.
Do different engines cite different sources for the same question? #
At the level of a single answer, yes, and the structural reasons are set out in Why AI Search Engines Cite Different Sources for the Same Question. Measured over a window and conditioned on evidence, the engines converge sharply; that measurement is in Cross-Engine Citation Agreement.