News Source Citing Patterns in AI Search, Digested: What Citation Concentration Means for Your Brand

The paper is News Source Citing Patterns in AI Search Systems, posted to arXiv as 2507.05301 (v1, 7 July 2025). It asks a narrow question: when AI search systems cite news, which outlets get cited, and how evenly are citations spread?

That question matters beyond newsrooms. If a handful of sources take most citations, then being cited is partly a function of already being big. That is the pattern our brand prominence glossary entry describes, and this paper is one of the clearest measurements of it.

The study compares three provider families: OpenAI models, Perplexity and Google. It reports results per family rather than one blended number, which is the right instinct. Engines behave differently, and averaging them hides that.

The data comes from the AI Search Arena, collected between March and May 2025. Two features of that setup shape how far the results travel.

A self-selected population

Arena users choose to take part and choose what to ask. Their questions are not a random sample of what buyers type into ChatGPT or Perplexity. Any finding describes that population first.

A fixed and now-old window

As of October 2026, the data is roughly 17 months old. Every engine in the study has shipped new models since. The cutoffs, retrieval systems and citation formats in the 2025 window are not the ones running today.

What Gini measures

The Gini coefficient is a standard inequality measure. A value of 0 means every source gets an equal share of citations. A value near 1 means almost all citations go to very few sources. It summarises the whole distribution in one number, which makes it useful for comparing engines.

The headline figures, exactly as reported:

• For OpenAI models, the top 20 most frequent news sources accounted for 67.3% of all citations.
• The Gini coefficient for OpenAI models was 0.83.
• Perplexity's Gini coefficient was 0.77.
• Google's Gini coefficient was 0.69.

All three values are high. In this window, OpenAI models concentrated news citations the most and Google the least. Note that the 67.3% top-20 share is reported for OpenAI models only. Do not extend it to the other two.

Other evidence points the same way. Semrush's index of 126M prompts (January to April 2026) found 36 brands visible across all four surfaces it tracked every month, including YouTube, Google, Reddit, Amazon and Walmart. That finding is, in effect, a measurement of brand size. For a broader view of who gets cited, see which kinds of sites AI search engines cite most.

The paper measures a pattern. It does not measure a cause, and it does not hand you a tactic.

• It does not show why prominent outlets win. Big publishers have more pages, more coverage, more inbound mentions and more presence in training data. Any of these could drive the result. The study cannot separate them.
• It does not show a lever. Nothing here tells a brand what to change this quarter to join the top 20. Prominence builds over years and sits outside a single site's control.
• It does not cover commercial brands directly. It studies news sources. The concentration pattern is plausible for other categories, but it was not measured there.
• It does not describe today's engines. A 2025 Arena snapshot cannot tell you current Gini values.

Why this is the confounder everywhere

Brand prominence is the dominant uncontrolled confounder in correlational GEO studies. When a study finds that pages with some feature get cited more, large brands often have that feature too. The usual "controls", Ahrefs DR and Moz DA, are vendor regressions built from backlinks and traffic. They are proxies for prominence itself, so conditioning on them does not remove the confounding. Google publishes no site-level authority score.

Controlled studies sidestep this by design. The 252,000-trial factorial (arXiv 2605.25517) anonymised brands deliberately and warned that real systems may favour established domains. C-SEO Bench also works on documents placed directly in context. The critical survey of 45 GEO studies concludes that no reviewed technique shows a stable, cross-platform causal effect on discoverability.

Treat concentration as a fact about the terrain, not a score to chase.

1. Benchmark only against like-sized brands. If you compare a 20-person software company with a household name, the gap mostly reports brand size. Pick three to five competitors of similar scale and compare platform by platform.
2. Measure each engine separately. The Gini values differ by provider. A single blended number would hide that. Check how each platform answers your buyer questions, one engine at a time.
3. Fix what is in your control first. Access levers have the strongest evidence: server-side rendering of key facts, per-bot robots.txt permission, firewall and CDN reachability, and index and snippet eligibility. These decide whether you can be a candidate at all. Our guide to how AI search picks sources explains the stages.
4. Get named where engines already look. When engines cite other sites instead of yours, being named on those sites is the route into the answer. This is an association in the evidence, not a promised effect.
5. State specific facts plainly. A price, a number or a one-line definition on a fetchable page gives an engine a sentence it can quote.

If you want a starting baseline, the free checker on this site shows whether AI engines name you for up to three buyer questions. It is a snapshot, so rerun it against the same like-sized competitors over time. For newer brands, read whether a brand-new website can get cited.