How Concentrated Are AI Citations? Legal: 51–74%, Commerce: 28–34%
Published
This article was first published in Korean by Intrix Lab on September 11, 2026.
Depending on the industry, domains cited in AI answers can be concentrated among a few sources or spread widely. As of 2026-09-11, across 12 Intrix CT studies with at least 300 citations, the top 10 domains' share for OpenAI models was 50.5–73.8% in legal studies, 28–34% in brand and agency studies, and 34% in a home and living commerce study. A recent arXiv audit reported that "the top 10 domains accounted for 43.6% of citations" (English mental health queries, 3 platforms). That figure applies to that query group rather than universally, and our legal studies showed greater concentration.
Why did we not take 43.6% at face value?
The paper featured in GEO Watch covered one topic, mental health, in English, across three platforms. Mental health is an area search engines treat as YMYL (Your Money or Your Life), where institutional sources are strongly preferred. Treating concentration under those conditions as a constant across industries leads to misguided recommendations. We measured the same metric again using CT responses we had already collected, without submitting new queries. There was no cost.
How did we measure it?
The metric is the same as the paper's: the share held by the top 10 domains. Among citation studies conducted by Intrix, we calculated it by study and model for 12 studies with at least 300 citations. We used individual studies because the paper was designed around a single topic. The value for the entire pool combining multiple industries is 27.0%, but we include this only for reference because the value falls as queries become more diverse.
| Study type | OpenAI top 10 share | Gemini top 10 share |
|---|---|---|
| Legal (3 studies) | 50.5 / 62.5 / 66.6 / 70.2 / 73.8% | 29.7 / 42.9 / 60.3% |
| Brands and agencies | 27.9 / 33.6% | - |
| Home and living commerce | 33.9% | - |
| Our own company (Intrix) | 58.1% | - |
Each value in the table represents one study-and-model combination. Even within the same legal study, results sometimes differed by more than 20 percentage points across models, so we did not combine them into a single value.
What does it mean for industry to determine concentration?
For legal queries, citations are shared among case law and legislation sources, associations, and a few large platforms. For brand or home and living commerce queries, the top 10 account for less than one-third of the total, with the rest spread across hundreds of hosts. Because this difference appeared with the same model, period, and language, what determines concentration is the nature of the query group rather than the language or platform. This is consistent with our earlier comparison of citation composition in YMYL and non-YMYL queries.
Why was Gemini's concentration lower?
In the same legal studies, Gemini's top 10 share was 29.7–60.3%, generally lower than OpenAI's. Citations per response were 6.9 for Gemini and 3.5–3.7 for OpenAI. These data cannot distinguish whether a model that provides more citations spreads them more widely as a result, or whether the difference comes from different search backends. We leave that for the next measurement.
Is it true that Korean queries receive fewer citations?
The paper reported fewer citations for non-English queries. In our Korean responses, citations per response were 3.5–3.7 for OpenAI and 6.9 for Gemini, and 9 of 2,178 gpt-5.6-luna responses had 0 citations. This corpus did not show citations drying up because the queries were in Korean. However, many of our queries were commercial queries seeking businesses and services, while the paper used health information queries, so language effects and query-type effects are mixed.
What do we do differently as a result?
In industries with high concentration, there is structurally little room for individual businesses to be cited directly. Before modifying a company's own site, we check external routes for accurate inclusion on the association, directory, and institutional pages that receive citations. In industries with low concentration, getting the company's own pages considered as candidates is itself a realistic goal. Here, we first examine the discovery stage through which pages enter the reference set. The study results for that industry tell us which situation applies. A number from a single paper cannot substitute for them.
Limitations
This measurement is an observation across 12 Intrix studies, not a sample of entire industries. The language, topics, and platforms differ from those in the paper, so we are examining the same metric in different populations, not making a direct comparison. Concentration is sensitive to how diverse the query group is. We do not use it as evidence to guarantee any particular ranking or citation.
Frequently asked questions (FAQ)
Q. How can we find out the concentration in our industry?
You need a citation study using actual queries for that industry. Intrix citation studies also show which domains receive citations in each industry.
Q. Should we give up on GEO if concentration is high?
No. The goal changes. Before pursuing direct citations, we check routes for inclusion in the institutions and directories that receive citations.
Q. Does this mean the 43.6% figure is wrong?
No. It is correct for the paper's population of English mental health queries. The point is that it should not be applied unchanged to other industries.