How Concentrated Are AI Citations? It Takes 1,357 Domains to Reach Half

In the September 18, 2026 release of the Machine Relations Index, the most-cited domain on the open web — reddit.com, cited in 2,012 of 15,782 observed answer runs — holds 1.81% of all citations recorded. The top ten domains together hold 7.41%. To account for half of every citation the Index observed, you have to count down to the 1,357th domain. To account for 80%, you count down to the 6,643rd.

That is a concentrated distribution. It is not the distribution the category talks about.

The working assumption behind most AI-visibility strategy is that answer engines collapse the web to a short list, and that the practical question is how to join a handful of dominant sources. Measured across six engines, 125 days and 22,179 cited domains, the head is real but thin, and the mass of citation sits in a very long middle. This matters for a publisher because it changes the question from can I displace Reddit to which of the 21,679 domains below the grading floor am I, and what would move me.

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, generated September 18, 2026. The observation window runs May 10 to September 18, 2026 — 125 days, 15,782 observed answer runs across six engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Perplexity), 22,179 cited source domains.

One definition carries the whole piece. 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 22,179 domains, that is 110,877 domain-run citations, and it is the denominator for every share below. It is not the same as the raw citation-event count, which is higher because a single answer run can cite one domain several times. Shares of domain-run citations answer "how often was this domain one of the sources", which is the publisher's question.

A domain's citation rate is its runs_cited divided by 15,782. A domain publishes a confidence grade only after clearing an evidence floor of 10 observed runs across 7 distinct run dates.

The concentration curve #

The Gini coefficient of the distribution is 0.663: unequal, but not winner-take-all. A winner-take-all citation layer would put the first ten domains well past a quarter of the total. These ten hold less than a thirteenth.

The ten domains at the top, with their source-role classification from the release:

Six distinct source-role classes appear in those ten slots. There is no single class that owns the head.

The median indexed domain is cited twice #

The other half of a distribution is the part nobody publishes. Across 22,179 domains:

• 10,381 domains — 46.8% of the index — were cited in exactly one answer run across 125 days.
• 14,053 (63.4%) were cited in two runs or fewer.
• 18,166 (81.9%) were cited in five or fewer.
• Only 2,165 domains (9.8%) reach ten or more observed runs.
• Only 71 domains are cited in 100 runs or more.

The confidence grades follow directly. Of 22,179 domains, 14 carry grade A, 56 carry grade B, 430 carry grade C, and 21,679 are still collecting — below the evidence floor, with no publishable rate. Being in the index is not an achievement; almost every domain the engines touched is in it. Clearing the floor is the achievement, and 500 domains have.

Breadth is just as thin. 17,186 domains — 77.5% — appear in exactly one segment (one subject category paired with one question shape). The median domain appears in one. Only 710 domains appear in five or more. The maximum is 98.

Read together: the typical cited domain is cited once or twice, in one segment, and has not been observed often enough to earn a grade. It was a source for one answer to one kind of question, once.

Concentration is a property of the source class, not of AI answers #

This is the finding that changes what a publisher should do, and it is invisible in any single aggregate number. Split the 22,179 domains by their source role and measure concentration inside each class:

The release assigns every domain to one of nine source roles. Three of the nine hold fewer than thirty domains each, so the top-ten measure saturates inside them; the share held by each class's single most-cited domain is the measure that compares across all nine.

The nine rows account for all 22,179 domains and all 110,877 domain-run citations. Among the eight identified classes the top domain's share runs from 94.6% down to 2.7% — a thirty-five-fold difference in how much the leader of a class takes.

Two rows in that table depend on how the release classifies individual domains, and the notes below say how to read them: the release files prnewswire.com in the editorial class rather than the wire class, and the editorial row's leader is an open publishing platform.

Search or media platforms look like a class and behave like one domain. Ten domains, 1,513 citations, and youtube.com holds 94.6% of them. The other nine were cited 82 times between them across 125 days; eight of those nine sit below the evidence floor and carry no grade at all. This is the most concentrated position measured anywhere in the release, and it is not a position a publisher can enter.

Community and social platforms are the next-tightest. The class has only 27 domains in it, and ten of those take 97.3% of everything the class is cited for. For a new community property, the class is effectively closed: the seats are taken by Reddit, LinkedIn, Substack, dev.to and Facebook, and the tail behind them is negligible.

The paid press-release layer is the smallest classified citation pool in the index. Counted with prnewswire.com in the wire class, the class is ten domains, and on the release generated 2026-09-20 it holds 414 of 112,518 cited-domain observations — 0.37% of classified citation, below search or media platforms at 1,517 on ten domains. Of the nine wire domains in the table, three clear the evidence floor, all at grade C: businesswire.com in 97 runs, globenewswire.com in 71, financialcontent.com in 37. The other six have too little evidence to grade. Whatever the wire layer buys, appearing as a source in an AI answer is not measurably it.

How to read the wire row. The table's wire and press-release row counts nine domains, because the release files prnewswire.com as an editorial publication. The 0.37% figure above counts it in the wire class, and it is the figure to use: at 414 the wire and press-release class is the smallest classified citation pool in the release by a wide margin.

Vendor-owned sources are the flattest of the identified classes, the residual bucket aside. 823 domains share 11,153 citations, the top domain takes 2.7%, and the median vendor domain is cited three times — the highest median of any class. A vendor page competes against a wide, shallow field rather than against an incumbent.

Editorial publications sit between: 1,222 domains, the largest classified citation pool at 13,309, and a leader on 6.1%. There is room in the class, but the median editorial domain is cited twice, which is the same as the median across the whole index. Being an editorial publication is not, by itself, an advantage.

How to read the editorial-class rank. The editorial-class leader and the class share above are read off the Index's source_role field, which in the current release combines source-type evidence with AuthorityTech's placement catalog; 96 of the 1,231 class members are admitted through the catalog, and the Index is moving editorial classification to source-type evidence alone. The class currently includes six platform domains alongside edited newsrooms: medium.com (class rank 1), amazon.com (7), prnewswire.com (8), apple.com (16), wordpress.com (239) and blogspot.com (330). At the same time substack.com, beehiiv.com and dev.to — the same hosted open-publishing product as Medium — are filed as community platforms. Those six hold 1,358 of the editorial class's 13,525 cited runs, 10.04%, and they include the class's top slot, which is why the 6.1% leader above is medium.com. The class denominator also moves between releases (published figures carry 1,025 to 1,230) because it counts the domains classified in each release. What is unaffected: the citation rate, cited runs, days cited, engine breadth and confidence on this page are direct per-domain observations over a fixed run set, and are the primary measure; read the rate first.

The practical consequence: "how hard is it to get cited" has no single answer. It has nine answers, and which one applies to you is decided by what kind of source you are, not by how good the page is.

The scarcest position is not volume, it is engines #

There is a second concentration in this release, and it is steeper than anything in the volume curve. Split the 22,179 domains by how many of the six engines cited them at least once:

347 domains are cited by all six engines. They carry more citation than the 14,969 cited by exactly one — 24,530 domain-run citations against 23,793. One domain cited by all six carries as much as forty-three cited by one. The median all-six domain was cited in 41 runs against a whole-index median of 2.

That group is not the media list the phrase suggests. Of the 347: 153 are unclassified, 85 vendor-owned, 58 editorial publications, 21 market and company databases, 13 analyst and consulting sources, 11 academic and government sources, 4 community platforms, 1 search or media platform, 1 wire service.

The engines differ enormously in how wide they cast. The release records citations to 9,497 distinct domains from Perplexity, 7,079 from Gemini, 5,138 from Claude and 4,877 from ChatGPT. Perplexity's, Claude's and ChatGPT's counts are exact; Gemini's is a floor, so it is not ranked below Perplexity here. What the exact counts alone establish is a 1.95x spread between the widest and narrowest arms with gap-free collection. The two Google surfaces are recorded at 7,441 distinct domains for Google AI Mode and 2,279 for Google AI Overviews, and both are floors rather than measurements.

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.

Correction, September 25, 2026. An earlier version of the paragraph above ranked all six engines against each other by how many distinct domains each cited. That ranking is withdrawn wherever it places either Google surface against another engine. The Index's own collection recorded zero citations from Google AI Overviews on any monitored query from August 15, 2026 to the end of this release's window, and zero from Google AI Mode from September 14. The cause was a defect in the Index's own Google collection path, replaced on September 25, 2026: the prior collector graded a response as a success whether or not it carried content. A distinct-domain pool only accumulates, so missing collection days can remove domains from an engine's set but never add them; each Google figure is therefore a count the engine reached at least, and neither is ranked here against the other engines or against each other. The four figures above stand unchanged, because all four of those engines collected continuously through the window. The cross-engine agreement study reads the same gap on its own release with the same marking.

A domain cited by one engine has cleared one retrieval system's filter; a domain cited by six has cleared six filters that share very little machinery, each governed by its own user agent under RFC 9309.

This is where effort converts. Moving from one engine to two is a larger structural change than doubling volume on a single engine, and it is invisible in any citation-rate number.

Inside a segment, the entry threshold is low #

The whole-index curve overstates the difficulty, because no publisher competes against 22,179 domains at once. They compete inside a segment: one subject category paired with one buyer question shape. 85 of the 151 measurable segments in this release are published. Across them:

Thirteen cited answers puts a domain in the top ten of the median segment, and seven puts it in the top twenty-five. Restricting to the 78 buyer-question segments and excluding the seven longer-running news segments, the number falls to twelve.

Segments vary far more than the median implies. The most open in the release is the legacy news topic, where 4,025 domains share citations and the top ten hold 8.3%. The most closed is Consumer Finance comparisons, where 170 domains share citations, the top ten hold 42.1%, and acorns.com leads at a 38.41% citation rate. 20,704 of the 22,179 indexed domains — 93.35% — appear in at least one published segment, and 16,270 appear in exactly one of them. Most cited domains are cited about one thing.

How v2.0 updates the v1.1 reading #

Machine Relations published an analysis on June 14, 2026 — AI Citation Concentration: Why Market Databases Capture Disproportionate Share — which described the distribution as a power law and read market databases as capturing disproportionate share. That analysis was run under MRI Score v1.1 over 7,124 domains and 28,870 source events, and it is labelled on the page as superseded by v2.0 on July 5, 2026.

At v2.0, over three times as many domains, the shape reads differently in two ways. The head is weaker than "power law" suggests in ordinary use: 1.81% for the leading domain, 7.41% for the top ten. And market databases are not the most concentrated class — their leader, g2.com, holds 7.3% of the class, which puts them sixth of the nine source roles, behind search or media platforms, community platforms, wire distribution, analyst research, and academic and government sources. Market databases remain strong in absolute terms — g2.com and crunchbase.com are rank 9 and rank 16 overall — but at v2.0 their class is not where concentration lives.

Updating a published reading as the measured universe grows is what a versioned instrument is for. Every figure in this piece is tagged to release mri_score_v2.0 generated 2026-09-18, so any later release can be compared against it.

Where the independent research agrees #

An empirical study of source coverage across LLM-based and traditional search engines, analysing 55,936 queries over six LLM search engines and two traditional engines, reports that LLM-based engines cite domains with greater diversity than traditional search engines, and that 37% of domains are unique to the LLM-based engines (arXiv:2512.09483). That is an independent measurement, on a different population with a different method, pointing the same way as the curve above: the citation layer of AI answers is wider than the intuition that it collapses to a few sites.

Two adjacent lines of work frame why the tail behaves as it does. A measurement framework distinguishing citation selection from citation absorption across AI search platforms argues that appearing as a citation and actually shaping the answer are separate dependent variables (arXiv:2604.25707). And an observational audit of Google AI Overviews on "Your Money or Your Life" queries, using rank- and provenance-conditioned citation measures, finds citation behaviour that is not explained by retrieval rank alone (Kakimov et al., PMLR v318). Neither study measures the concentration curve directly; both support the reading that source selection is its own process rather than a re-ranking of search results.

Access policy is the one structural factor a publisher directly controls. The Robots Exclusion Protocol is standardised as RFC 9309; Google documents its crawlers and the Google-Extended token separately from Search inclusion (common crawlers, crawler overview); OpenAI documents its own bots and their user agents (OpenAI bots, catalogued independently at Dark Visitors); Cloudflare has made per-site blocking of AI crawlers a one-click control (Cloudflare). Google's own guidance on appearing in AI features is published in Search Central (succeeding in AI search), and Search Console documents where AI-surface performance appears in its reports (Search Console help). Google has described AI Mode as a distinct search experience rather than a variant of the results page (Google). On the demand side, survey work finds users click source links less often when an AI summary is present (Pew Research Center), which is why citation share and traffic share have to be measured separately. The broad web corpora these systems are built on are themselves public and enumerable (Common Crawl).

What a publisher can check #

Every one of the 22,179 domains in this release has a public page. A domain profile lists its overall citation rate, its confidence grade, its rank in the full universe, and every segment in which it has been observed — for example reddit.com. Category and question-shape leaderboards list the top domains per segment, such as AI infrastructure. The Index is the entry point, and the full release is machine-readable.

Domains that clear the evidence floor in a published segment also carry a citable badge stamped with this release id, served at machinerelations.ai/index/domains/<domain>/badge.svg and embeddable from the profile page.

Four checks are worth making against the numbers above. Which percentile of the curve your domain sits in, from the full-universe rank on its profile. Whether it clears the evidence floor, or is one of the 21,679 still collecting. Which source-role class it was assigned to, because the class decides which concentration figure applies to you. And how many of the six engines have cited it, because that number, not the rate, is what the steep part of this release's distribution is made of.

Limits #

This is a measurement of the Index's own observation window and question basket, not of the whole web. The Index runs a fixed basket of buying and research questions; a domain cited heavily for questions outside that basket will look thinner here than it is. Domains cited zero times in the window are not in the file at all, so the 22,179 is a population of cited domains, not of eligible ones — the true denominator of "sites that could have been cited" is larger and unknown.

Nothing here is causal. The curve describes what engines cited over 125 days. It does not say why, and it does not say that any intervention moves a domain along it. A domain's position at the next release is the only test of that.

Finally, concentration measured on domain-run citations is not the same as concentration measured on citation events, on answer position, or on whether the cited page actually supported the claim beside it. That last question is measured separately in Answer-Source Fidelity, and its results do not follow from this curve.

Release: mri_score_v2.0, contract machine_relations_index_public_view_v2.0, generated 2026-09-18. Window: 2026-05-10 to 2026-09-18, 125 days observed. Population: 22,179 cited domains, 15,782 observed answer runs, 110,877 domain-run citations. Evidence floor: 10 observed runs across 7 distinct run dates. See also the Machine Relations Index Report for September 18, 2026 for the same release read by segment leader and engine.

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