Half of AI Citations in a Category Take Between 57 and 250 Domains

Across the 20 categories in the Machine Relations Index that clear the evidence floor, the ten most-cited domains hold between 7.6% and 23.2% of citations, and reaching half of a category's citations takes between 57 and 250 different domains. There is no category in which a short list of 15 or 30 outlets covers most of what an answer engine cites. The concentration figure every pitch list is built on is a cross-topic aggregate, and at category level it does not survive. The release before this one, on 16 categories, gave 49 to 247: the band moved by a few domains while the Index grew by half.

What was measured #

Every figure here comes from the public release at machinerelations.ai/data/machine-relations-index.json, contract machine_relations_index_public_view_v2.0, release mri_score_v2.1+2026-10-14+7a624b28cd41, methodology version mri_score_v2.1, generated 2026-10-14. The observation window runs 2026-05-10 to 2026-10-14: 151 observed days, 25,323 observed answer runs across six engines (ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity), 30,757 cited source domains and 190,582 source events.

The Index scores each domain separately inside each subject category and each of six buyer-question shapes: best tools, how to choose, is it worth it, problem-first research, top lists, and head-to-head comparisons. For every domain and every category-and-shape cell it records runs_cited, the number of observed answer runs in which at least one engine cited that domain.

The unit of this study is the domain-shape-run citation: one domain, cited in one observed answer run, for one question shape inside one category. Summed over every domain's per-category cells, the release's 24 category buckets hold 167,832 of them; the 20 categories this study reports hold 148,639. That unit answers the publisher's question, how often was this domain one of the sources for a question in my category, which is not the same as a raw citation-event count, because a single answer run can cite one domain more than once.

The answer, by category #

Sorted by the number that matters to a pitch list: how many distinct domains you would have to be present on to cover half of the citations in your category.

Four numbers in that table do work no aggregate figure can do.

No single domain holds more than 4.0% of its own category. The most concentrated top position anywhere in the 20 categories is nerdwallet.com at 4.00% of citations in Consumer Finance. In Enterprise Software the leader, erpresearch.com, holds 1.65%. A category with a dominant source does not appear in this measurement.

The top 25 domains never reach 38%. The maximum across the 20 categories is 37.46%, in Family Software, which has the second-smallest cited pool in the table. In Industrial the top 25 hold 14.10%.

The spread between categories is more than four-fold. Consumer Finance needs 57 domains to reach half; Enterprise Software needs 250. A quantity that varies four-fold across categories is not a constant, and a strategy calibrated to one category's curve is calibrated to the wrong number in every other.

Concentration is not a function of category size. Three categories of almost the same breadth sit far apart: Local Services has 1,044 cited domains and reaches half in 113, Industrial has 1,240 and needs 205, and Consumer Finance has 1,326 and needs 57. Breadth of the cited pool and the shape of its head move separately.

Two releases, the same band #

The study first ran on the release of 2026-09-23: 130 days, 16,475 answer runs and 16 reportable categories. The release of 2026-10-14 adds 21 days, 54% more answer runs and four reportable categories. The band barely moved.

Most categories need a few more domains to reach half as the window lengthens, because a longer record adds tail domains faster than it adds head citations. Three need fewer. The ordering at both ends held: the consumer-money and family categories are the most concentrated, the broad enterprise and healthcare categories the least.

The first run excluded iGaming and Betting, Industrial and Local Services because no cell in them cleared the evidence floor, and warned that their curves were shaped by the floor rather than the market. Local Services shows why. On 401 sub-floor citations across 236 domains it appeared to reach half in 59 domains; on 3,526 citations across 1,044 domains, every shape now published, it takes 113. iGaming and Betting (97 then, 95 now) and Industrial (188 then, 205 now) moved less. Logistics and Freight is new to the taxonomy.

The claims this measurement tests #

Two concentration figures circulate widely enough to shape buying behaviour.

The first is that 15 websites own 68% of AI citations, which traces to the 5WPR AI Platform Citation Source Index, a synthesis of six earlier studies, announced via PR Newswire and reported by Everything PR. The second is that roughly 30 domains own about two thirds of citations within any single topic, a framing repeated across vertical strategy guides such as AI+Automation's vertical playbooks and the vertical citation hub at ConnectEra, which reports single sources owning 85% or more of individual verticals.

Neither survives contact with a per-category measurement of this size. Among the 20 reported categories, the top 25 domains clear 30% in two and sit below 26% in 15. Reaching 80% of a category's citations takes more than 400 domains in 16 of the 20, and more than 700 in four.

This is not a claim that concentration is absent. It is a claim about where the concentration sits: in a thin head that no strategy can buy into, above a very long middle that is the only part a publisher can actually enter.

Where independent measurements land #

Two 2026 studies measuring the same quantity with different panels agree with this range, which matters more than either study alone.

Analyze's August 2026 study analysed 22,295 AI answers across ChatGPT, Perplexity and Google AI Mode, covering 115,843 citation events and 7,058 unique cited domains. It found the top ten domains covering 11% to 13% of citations per engine, and the top 50 covering 29% to 34%. GetIntel's August 2026 run logged 3,754 citations across ten software categories and five engines, going to 1,258 domains: the single most-cited domain took 3.5%, the top ten took under 15%, and 57% of domains were cited exactly once.

Both land inside the Index's 7.6% to 23.2% band for the top ten, and GetIntel's 3.5% ceiling on a single domain sits just under the Index's 4.0%. Three separate panels, three separate prompt sets, three separate engine mixes, one result.

The broader literature points the same way from other angles. Featured's June-August 2026 citation report found 34.5% of 22,881 citations going to domains with a Domain Authority under 40, across 11,499 different domains. Fahlout's synthesis of more than 30 studies records the four major engines sharing as little as 11% of their cited domains with one another, which is itself a reason a single cross-engine top list cannot describe any one engine's behaviour. Foglift's buyer-intent study found the cited set turning over almost completely as intent moves from discovery to shortlist, which is the same instability in a different axis.

Outside the commercial trackers, an academic result finds the same shape in a controlled setting. "When AI Writes, Who Gets Cited?", from teams at UT Austin, Stevens Institute of Technology, Washington University in St. Louis, Notre Dame and Rice, tested eleven models from three vendors on uniformly random panels and found the top decile receiving 23.3% to 30.2% of citations against 15.6% under indifferent selection. Sharp, measurable concentration — and still nothing like a handful of sources owning the answer.

Where sources disagree with this, the disagreement is usually a unit difference rather than a contradiction. Buffy's read of vertical concentration reports top-three brands holding more than 80% of AI visibility in some categories; that is share of answers mentioning a brand, not share of citations going to a domain, and the two can diverge completely because one brand's mention can be sourced from dozens of domains. Attrifast's 1,200-prompt vertical study, DeltaV's 25,000-citation study and Presenc's industry breakdown all report per-vertical differences in the same direction as ours, at smaller scale and with different category boundaries, which is why their absolute figures are not directly comparable to these.

What is excluded, and why #

Four of the release's 24 category buckets are reported below the line rather than inside the finding.

Three of them, sales-gtm-crm, physical-consumer and professional-services, carry no published stratum in this release. Every domain-and-question-shape cell inside them sits below the Index's evidence floor of 10 observations across 7 distinct run dates, and at 691 to 773 citations each they are the thinnest buckets in the release. As Local Services showed between releases, a curve built on sub-floor cells can be off by half.

The fourth, legacy-unmapped, is not a subject category. It is the bucket for observations not mapped to a taxonomy node, and it is the largest bucket in the release at 17,027 citations across 4,017 domains. Its curve, with the top ten at 8.27% and 303 domains to reach half, is included here only because it is the widest and flattest distribution in the data.

How robust the answer is #

The obvious objection to counting every observation is that the thin tail is noisy. So the same curves were recomputed using only cells the release marks published, those clearing the evidence floor on their own, and the answer barely moves.

In 18 of the 20 reported categories the published-only figures are identical, because every cell in them is published. In the two where they differ, Martech and Advertising moves from a top ten of 12.04% to 12.81% and from 137 domains to half down to 104, and Logistics and Freight, where two of six shapes are published, moves from 12.39% to 16.43% and from 108 down to 54. The top-ten band across all 20 stays 7.6% to 23.2% either way; the half band becomes 54 to 250.

That stability is the point. The finding is not an artefact of counting single observations, because the categories are wide enough that removing the unpublishable cells changes the band by a few domains at one end and not at all at the other.

Two limits belong on the record. First, these are domain-run citations inside a fixed taxonomy, and a category boundary is a modelling choice: a wider category will generally show a flatter curve, so the absolute counts belong to mri_taxonomy_v2.0 rather than to any universal definition of a market. Second, a 151-day window records what the engines did, not what they will do; the Index publishes drift series precisely because the composition of these heads moves between releases.

What a pitch list should do with this #

The practical consequence is not that concentration does not exist. It is that the reachable part of the distribution is much further down than the industry's working assumption, and it is category-specific.

Take the strongest possible version of a short list: the ten most-cited domains in a category, chosen with perfect foresight and including the platforms no comms team can pitch. In Industrial that list covers 7.57% of the citations an engine has to choose from. In Consumer Finance it covers 23.23%. Widen it to 25 domains and Industrial reaches 14.10% while Family Software reaches 37.46%. Neither end of that is a strategy, and the distance between them is the whole reason a single cross-topic number was always going to mislead.

The number that does travel is the shape: a head of a few domains each holding 4% or less, a middle of a few hundred domains that together hold the majority, and a tail of single observations. A publisher's realistic question is not which fifteen sources to join. It is which of the few hundred domains inside its own category's middle it can plausibly enter, and what the ordering of those looks like, which is what per-category, per-question-shape standing in the Index is for.

The whole-web version of this curve, from an earlier release, is at How Concentrated Are AI Citations?. Which kind of source wins inside each category, and how that changes with the buyer's question, is at Source Class Capture by Category.

Method notes #

• Counts are summed from each domain's mri_score_v2.strata[] entries, grouped by category across all six question shapes. A domain contributes to a category once per shape in which it was cited.
• A category is reported when at least one of its cells is published. The 2026-09-23 figures in the comparison table are from that release's public view, mri_score_v2.0+2026-09-23+fdec6388001a, computed the same way.
• "Domains to half" is the smallest number of domains whose citations, taken in descending order, reach 50% of that category's total. "Domains to 80%" is the same at 80%.
• The published-only sensitivity uses the same computation restricted to cells with status: "published".
• Shares are of the category's own total, not of the whole index, so they do not sum across categories.
• Categories and ids are the nodes of mri_taxonomy_v2.0 as carried in the release.