Only 508 of 22,377 AI-Cited Domains Are Cited Often Enough to Score

In the September 21, 2026 release of the Machine Relations Index, AI answer engines cited 22,377 distinct domains. Five hundred and eight of them were cited often enough, on enough separate days, for the Index to publish a confidence grade.

That is 2.27% of the observed web. Those 508 domains hold 34.66% of every citation the Index recorded. The other 21,869 sit below the evidence floor, and the single most common thing that happens to a domain in AI answers is that it gets cited once, in 128 days, and never again.

Ten thousand three hundred and thirty-four domains — 46.18% of everything cited — were cited exactly once.

This is the number under every AI-visibility leaderboard in the category, including ours, and almost nobody states it. A list of "domains AI engines cite" is mostly a list of accidents. The interesting population is the one an instrument can stand behind, and it is very small.

What was measured #

Every figure below comes from the public release at machinerelations.ai/data/machine-relations-index.json, contract machine_relations_index_public_view_v2.0, methodology mri_score_v2.0, release mri_score_v2.0+2026-09-21+d73e239e7c5b, generated September 21, 2026 and verifiable against the release manifest by SHA-256.

The observation window runs May 10 to September 21, 2026 — 128 days observed, 16,146 observed answer runs across six engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity), 126,924 source events and 22,377 cited source domains.

Two definitions carry the piece.

Domain-run citation. For each domain the Index records runs_cited: the number of observed answer runs in which at least one engine cited that domain. Summed across all 22,377 domains that is 113,112 domain-run citations, and it is the denominator for every citation share below.

The evidence floor. The Index publishes a rate only where it has seen at least 10 observed runs across at least 7 distinct dates. Below that line a domain is marked collecting rather than scored. A graded domain carries confidence A, B or C. Collecting is not a zero rate and never has been — that distinction has its own page — it is the instrument declining to publish a number it cannot support yet.

The scoreable population #

Two thirds of citation volume sits on domains the Index will not score. That is not a defect in the index; it is the shape of the thing being measured. An engine answering a long-tail question reaches for a page it will never reach for again, and 128 days of observation across six engines produces a very large number of one-time events.

The consequence for a reader is blunt. Appearing in the citation record is not a position. It is a coin that landed.

The median graded domain was cited on 20 of the 128 observed days. The median collecting domain was cited on 2.

Three denominators, three different markets #

The same release answers "which source class dominates AI citation" three ways, and the three answers disagree with each other. Which one a vendor quotes is usually a function of which one flatters the product.

Count domains and the market looks like an undifferentiated mass: 84% of it is unclassified. Count citations and editorial publications lead the classified layer at 11.99%. Count scoreable domains and vendor-owned sources overtake them, 104 against 99.

None of these is wrong. They answer different questions, and a leaderboard that does not say which one it is counting cannot be checked.

The one comparison that survives class size #

Shares of a whole are the wrong instrument here, because these classes are not remotely the same size. Community and social platforms hold 27 domains; "other observed source" holds 18,811. Any top-N column over groups that different reads 100% for the small ones by construction.

The size-safe question is a rate inside each class: what share of the domains in this class were cited often enough to be scored at all?

Six classes carry more than 300 domains each, which is enough for the rate to mean something. Three classes do not — community and social platforms (10 scoreable of 27), search or media platforms (2 of 10) and wire and press-release distribution (3 of 9) — and their rates are excluded from the comparison above because a single domain moves each by three to eleven points. They are reported in the table above them and should be read as counts, not rates.

Among the six comparable classes the ranking is this: a vendor's own domain is about 1.6 times more likely than an editorial publication's to be cited often enough for an index to score it — 12.64% against 8.04%. Analyst research, market databases and academic sources follow. The unclassified mass trails at 1.02%, which is another way of saying the long tail is where one-time citations live.

That result cuts against the working assumption in most earned-media strategy, and it needs the right caveat attached: it is a statement about the odds a cited domain in that class is cited repeatedly, not about how much citation the class receives. Editorial publications still take more citation volume than vendor-owned sources, 13,560 against 11,280. The two facts coexist. The editorial class is larger and flatter; the vendor class is smaller and more concentrated in domains that get cited again.

Per-engine presence splits on the same line #

The floor separates the population on engine count too. Of the 508 scoreable domains, 233 were cited by all six engines and none was cited by fewer than two. Of the 21,869 collecting domains, 15,080 — 68.96% of that group — were cited by exactly one engine.

This is the same defect we corrected in public across ten of our own pages earlier today, arriving from the other direction. A headline that "most cited domains appear on only one engine" is a statement about domains cited once, not about how engines select sources; measured among domains with real evidence behind them, cross-engine agreement is high. The full correction and the conditioned overlap figures are on the cross-engine citation agreement page, and the dated record is in the Index correction record.

Why a floor at all #

Thresholds before publication are ordinary practice in every measurement field that has had to defend a number.

Statistical agencies evaluate reliability before official release rather than after complaint: Statistics Canada's quality guidelines make certification of data quality a release step, with minimum requirements covering coverage error, response rates and sampling error on key characteristics. The Media Rating Council's Minimum Standards for Media Rating Research — the basis on which audience measurement services are accredited — is built on the same principle: disclose the method and its limits inside the report. Survey publishers document precision as a matter of course; Pew Research Center's methods pages exist for exactly that. Metrology has its own vocabulary for the boundary between a signal and noise, maintained in the JCGM publications.

The arithmetic behind a floor is not exotic. Estimating a proportion to a stated precision requires a minimum sample; the NIST/SEMATECH e-Handbook gives the standard derivation. One observed run gives a domain a citation rate of either 100% or 0% with no useful interval around it. Ten runs across seven dates is a low bar, deliberately — it is the point at which a number stops being a single event.

Where a heavy-tailed distribution is involved, choosing that cutoff is itself the hard part of the analysis. Clauset, Shalizi and Newman's Power-law distributions in empirical data is explicit that characterising such distributions is complicated by "the large fluctuations that occur in the tail" and by "the difficulty of identifying the range over which power-law behavior holds." Publishing a ranking over the whole tail, cutoff unstated, is the error their paper was written about.

What this does not say #

It does not say the 21,869 are worthless. They hold 65.34% of citation. Collectively they are most of the market; individually none of them has a measurable position yet.

It does not say a scoreable domain is a good source. The floor measures whether an engine came back, not whether the answer was right. Whether the cited source actually supported the claim is a separate instrument.

It does not rank classes by value. The scoreable rate is an observability statistic. A class can be cited rarely and matter enormously when it is.

It does not exempt us. machinerelations.ai, which publishes this index, sits below its own evidence floor in this release. The instrument measures every domain an engine cites on equal terms, including the ones we own, and we do not lift our own rows above the line.

Where the classified layer ends. Source-role classification covers 39.23% of citation; the "other observed source" class is the unclassified remainder and is reported as such rather than folded into a conclusion. That boundary, and a classification defect we corrected in it this week, are documented in the correction record linked above.

How to reproduce this #

1. Fetch https://machinerelations.ai/data/machine-relations-index.json and confirm its SHA-256 against the release manifest. For this release the digest is d73e239e7c5b6195637e0c3c2fff24103e70969202261c0080eca4b74f127266.
2. For each entry in domains[], read mri_score_v2.overall.confidence and mri_score_v2.overall.runs_cited. Confidence A, B or C is scoreable; collecting is below the floor.
3. Group by source_role_label. Compute three columns: share of domains, share of summed runs_cited, and scoreable count divided by class domain count.
4. Do not compare the scoreable count across classes of different size; compare the rate, and drop any class with fewer than a few hundred domains from the comparison.

Every figure in this piece is a direct read of those fields on that release. The class totals close exactly: 22,377 domains, 113,112 domain-run citations, 508 scoreable.

What a reader should take from it #

If a tool reports that your domain is cited by AI engines, the first question is not the rate. It is how many times, across how many days, on how many engines. Below the floor, a rate is a description of a coin flip, and the distribution of measured citation rates is the wrong place to put your number until you are above it.

For a category that mostly sells dashboards built on the ungraded 97.73%, that is an uncomfortable sentence. It is also the one we hold ourselves to: the same floor keeps our own domains off our own leaderboards, keeps a shopping leaderboard honest about which of its ranks are still collecting, and is why a founder reading a first-of-its-kind shortlist gets told how thin the evidence under it is.

Being cited is not the bar. Being cited enough to be measured is — and 508 domains out of 22,377 clear it.

Engines that select sources are described at their own publishers: Google documents its crawlers and AI Overviews, OpenAI documents web search, Anthropic documents citations, and Perplexity documents its answer stack. None of them publishes a citation rate per domain. That is the gap this index exists to fill, and the floor is what makes the filling checkable.