Where the Citations Go: Ranking and Being Cited Are Coming Apart

In our report on the state of organic search traffic, we documented the decline of the click. This report is about what replaced it: the citation. When an assistant composes an answer, it chooses a handful of pages to draw on and, sometimes, to name. Those choices are now the scarce placement — and the evidence of the last year says they are being made by rules that are not the ranking rules.

1. The decoupling

The clearest measurement comes from Ahrefs, which has run the same analysis twice.

In July 2025, across 1.9 million citations from one million AI Overviews, 76.1% of the pages cited also ranked in the top ten organic results for the same query; 9.5% ranked between 11 and 100; 14.4% did not rank in the top hundred. The median organic rank of the first-listed citation was 2. If you ranked, you were probably cited.

By March 2026, the second study — 863,000 keyword SERPs and four million cited URLs, more than double the first — found the top-ten share had fallen to 37.9%. Positions 11 to 100 supplied 31.2% of citations, and 31.0% of cited pages were not in the top hundred at all. Filtered to organic blue links only, the figures were 37.1%, 26.2%, and 36.7%.

Source: Ahrefs. July 2025: 1.9 million citations from 1 million AI Overviews (top three citations per response). March 2026: 863,000 keywords, 4 million cited URLs.

Ahrefs notes that Google moved AI Overviews to its Gemini 3 model in January 2026, with better handling of long-tail questions, and that YouTube accounted for 18.2% of the citations that did not rank — 5.6% of all AI Overview citations. Whatever the exact cause, the practical implication holds: a page can keep its ranking and lose its place in the answer, and a page that never ranked can be the one the answer is built on.

2. Each engine has its own diet

The second finding is that “AI search” is not one system with one preference. Different assistants draw on strikingly different sources.

Ahrefs’ September 2026 analysis of more than three million U.S. queries found the most-cited domains in Google’s AI Overviews were dominated by video and community platforms. Among the fifty most-cited domains, YouTube took 22.9% of citations and Reddit 18.5% — together more than 41% — followed by Facebook (10.1%), Google’s own properties (8.8%), Instagram (5.6%), Quora (4.7%), and Wikipedia (4.0%). The top ten domains accounted for roughly 81% of citations among that top fifty.

Source: Ahrefs, The 50 Most-Cited Websites in Google AI Overviews, September 2026; 3 million+ U.S. queries.

ChatGPT and Perplexity look different again. Synthesizing the largest public citation datasets (including Profound’s 680 million tracked citations), 5WPR’s State of AI Citations 2026 reports Wikipedia at roughly 47.9% of ChatGPT’s top-ten source share and Reddit at roughly 46.7% of Perplexity’s — and an estimated 11% overlap between the domains the two engines cite. Ahrefs’ own cross-platform study of the ten most-cited domains found Wikipedia leading everywhere but with very different weights (16.3% of ChatGPT’s citations, 12.5% of Perplexity’s, 8.4% of AI Overviews’), YouTube favored by Perplexity and AI Overviews, and Reddit and Quora ranking high only in AI Overviews.

The pattern under the differences: assistants lean heavily on a small number of very large, very well-structured entity and community sources, and the rest of the web competes for what remains. Pew’s panel found Wikipedia, YouTube, and Reddit together made up 15% of the sources cited in Google’s AI summaries; news sites made up 5%; government sites 6% — three times their 2% share of standard results.

3. What the cited pages have in common

Across platforms, the pages that do get cited share a profile that has now been measured several ways.

They are fresher. Ahrefs analyzed 16.975 million citations across ChatGPT, Perplexity, Gemini, Copilot, and AI Overviews. Cited URLs averaged 1,064 days since publication, against 1,432 days for the organic results on the same queries — 25.7% fresher. ChatGPT skewed newest (958 days for in-text citations); Google’s AI Overviews were the most conservative (1,432 days, matching organic).

Source: Ahrefs, Do AI assistants prefer to cite fresh content?, July 2025; 16.975 million citations.

They contain something quotable. The GEO benchmark (Aggarwal et al., KDD 2024) tested nine page modifications across 10,000 queries in a generative-engine setting. Measured by position-adjusted word count in the generated answer, adding quotations improved visibility by 41%, adding statistics by 33%, and citing sources by 28%. Keyword stuffing — the classic SEO reflex — reduced visibility by 9%. The authors note statistics were especially effective in fact-heavy domains such as law and government.

They are about recognizable entities. The 5WPR synthesis reports a correlation of 0.334 between a brand’s search volume and its likelihood of being cited, and that having a Wikipedia entity increases citation likelihood roughly 2.8 times. Assistants organize knowledge around entities; the pages easiest to cite are those about entities the model already recognizes, described with consistent names and identifiers.

They come from sources that publish verifiable facts. The over-representation of .gov sites in Pew’s data is the cleanest signal. Nothing about a government page is optimized for AI; it is simply the canonical source of a specific number.

4. Being cited is not the same as being credited

There is an uncomfortable coda. In March 2025, the Tow Center for Digital Journalism gave eight AI search engines 1,600 excerpts from published articles and asked each to identify the article, publisher, date, and URL. Collectively they answered more than 60% incorrectly. Perplexity was wrong 37% of the time; ChatGPT Search 67%; Grok 3 94%, with 154 of its 200 citations leading to error pages. Premium tiers were more likely to give a confident wrong answer than to decline.

And even when an assistant uses a source, it may not link to it: ChatGPT mentions brands roughly 3.2 times more often than it cites them, per the 5WPR synthesis. Together these mean two things for a publisher. Attribution is leaky, so a citation strategy that depends on every use being linked will under-count its own success. And brand recognition — being the entity the model associates with a topic — is worth pursuing directly, because it is what survives when the link does not.

Methodology and caveats

Datasets and periods. Ahrefs’ two overlap studies (July 2025; March 2026) use different sample sizes and, in the second, a broader definition of results-page positions; Ahrefs reports both a “blocks” and an “organic links only” view and we quote both. Ahrefs’ most-cited-domain shares (September 2026) are shares among the fifty most-cited domains, not of all citations. The 5WPR report is a synthesis of vendor datasets (Profound, Goodie, Surfer, Semrush, Peec, Ahrefs, BrightEdge) collected between August 2024 and March 2026; its platform figures are top-ten source shares and should be read as orders of magnitude, since vendors track different prompt sets. Ahrefs’ freshness figures use publication dates where available and last-modified dates otherwise. The GEO benchmark measured visibility inside generated answers, not clicks. Pew’s source shares are for Google AI summaries in March 2025.

What is stable and what is not. Citation mixes move quickly: ChatGPT’s Reddit share collapsed from roughly 60% to 10% in September 2025 before stabilizing, and AI Overviews changed model in January 2026. The direction of the decoupling is consistent across sources; the exact percentages will not hold for long.

What to do about it

1. Measure citations directly. Track your domain’s appearances in AI Overviews and in the major assistants, and treat a ranking without a citation as a loss.
2. Optimize for the answer, not the list. The properties the data rewards — specific statistics, named sources, quotable statements, freshness — are properties of the page’s content, not of its link profile. Most classic SEO work does not move them; keyword stuffing moves them backward.
3. Be an entity, not just a URL. Consistent naming, an unambiguous “about” for every page, structured data that identifies what the page describes, and a presence on the entity sources assistants trust. Wikipedia presence and brand search are measurable and improvable.
4. Refresh on the data’s schedule. A 25.7% freshness gap is a gap you can close by design: pages that regenerate when their underlying facts update are permanently on the right side of it.
5. Cover the long tail of the question. Assistants answer specific questions and pull from pages that answer them specifically. Depth across an entire subject matters more than authority on a single page.

Sources

1. Ahrefs — 76% of AI Overview-cited pages rank in the top 10 (Linehan & Guan, July 21, 2025) — 1.9 million citations from 1 million AI Overviews
2. Ahrefs — Google AI Overview citations and top-10 rankings, updated (Linehan, March 2, 2026) — 863,000 keywords, 4 million cited URLs
3. Ahrefs — The 50 Most-Cited Websites in Google AI Overviews (September 2, 2026) — 3 million+ U.S. queries
4. Ahrefs — The 10 most-cited domains across AI assistants
5. Ahrefs — Do AI assistants prefer to cite fresh content? (Law, July 28, 2025) — 16.975 million citations
6. 5WPR — The State of AI Citations 2026 (May 2026) — Synthesis of Profound (680M citations), Goodie, Surfer, Semrush, Peec, Ahrefs, BrightEdge datasets
7. Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results (July 22, 2025)
8. Aggarwal et al. — GEO: Generative Engine Optimization (KDD 2024) — GEO-bench, 10,000 queries
9. Columbia Journalism Review / Tow Center — AI Search Has a Citation Problem (Jaźwińska & Chandrasekar, March 6, 2025) — 8 engines, 1,600 queries