What Makes Websites Get Cited by AI? Meta-Analysis Results

• Share via Twitter (opens in a new tab)
• Share via LinkedIn (opens in a new tab)
• Share via Facebook (opens in a new tab)
• Share via Email (opens in a new tab)

A meta-analysis of five independently published studies, tracking which domains large language models cite, how often, and why.

Cross-referencing five studies reveals two main citation levers

Read individually, each of the five studies describes one domain’s performance on one measurement window. 

Read together, every datapoint sorts cleanly into just two separate levers that can be pulled to improve AEO performance. No single study says this outright, because no single study looks at more than one domain at a time.

The table below maps each datapoint against the lever it measures and against which of Stanford Libraries’ three citation factors (retrievability, factual consistency, structural clarity) it tests.

Four of the six datapoints land on structural clarity. Two land on retrievability, and both trace back to something a site owner can’t buy with better markup: a licensing deal or a faster re-crawl schedule. None of the five studies test factual consistency, Stanford’s third factor, at scale. 

That gives us room to form a conclusion none of the five studies states on their own: licensing moves a domain’s citation share faster than structure does, but licensing access sits with the platform, not the page. 

Reddit’s climb traces to its licensing deal with Google, not to any change in how Reddit’s threads are formatted. Structure, by contrast, is the lever available to every page regardless of who owns the domain, and it’s the one the schema study and the BrightEdge findings confirm through two different methods: a 6-million-URL statistical correlation and a semantic-layer analysis. 

Therefore: Structure is the lever most sites can pull, and it’s the one place where two unrelated methodologies land on the same conclusion.

A checklist for earning LLM citations reliably

Five items, in order:

1. One H1 per page
2. Heading levels that nest without skipping
3. Schema markup matched to content type
4. Semantic markup instead of styled divs
5. A visible last-updated date

That list is a WCAG 2.2 conformance checklist wearing a different label. The Stanford Libraries framework for source selection names retrievability, factual consistency, and structural clarity as what a model weighs, and every item above serves that third measure directly.

Accessibility audits already measure most of what makes a page citable, and I think that’s pretty cool.

What counts as evidence here

Three kinds of evidence qualify for this analysis:

1. Peer-reviewed research on retrieval-augmented generation systems.
2. Named vendor and platform citation-tracking studies that publish their method alongside their findings.
3. Direct multi-query testing logs run against named models, sampled repeatedly.

The model scope here covers GPT-5-class systems, Claude 4.x, Gemini 2.x, and Perplexity’s answer engine. Citation behavior for live-retrieval systems, the ones that re-crawl the web on a schedule, is reported separately from training-data recall in a fixed model checkpoint.

Wikipedia and reddit reward the same habits

Strip away the domain names and the two case studies show one pattern:

• Defined headers that break a topic into retrievable chunks
• Timestamps or revision dates that show the page reflects current fact
• Structured question-and-answer or discussion threads
• Almost no decorative markup standing between the text and the point

Any page, regardless of size or brand recognition, can carry all four of those.

Wikipedia runs a nearly identical heading structure across millions of articles, and internal links that connect one topic to the next without ambiguity. Pages keep the same URL from creation onward, so a system indexing an article in June finds it at the same address in December. It all contributes to consistency a machine can count on every time it looks.

But Wikipedia’s dominance had a shelf life. Ahrefs measured Wikipedia at 29.7% of all ChatGPT citations on October 27, 2025, well ahead of every other domain in the dataset.

Reddit’s citation share grew by at least 73% between October 2025 and January 2026. That climb lines up with confirmed data deals: Reddit’s arrangement with Google alone is reported at about $60 million a year, and OpenAI struck a comparable indexing arrangement around the same period. Wide, fast indexing did what editorial reputation alone never could.

And Reddit’s format does something Wikipedia’s prose can’t: a question sits at the top of a thread, and the highest-voted answer sits directly under it. That’s the shape a retrieval system is built to lift, no summarizing required.

Despite that, ChatGPT has sharply cut its citations of both Reddit and Wikipedia. These numbers can shift almost at a moment’s notice.

A site with no brand recognition and no twenty-year domain history gains the same structural legibility by adopting the same formatting habits. A page that reads cleanly to a machine competes on that basis alone, regardless of who owns it. The harder question is how, without that infrastructure? 

Structured markup raises citation odds regardless of domain authority

Machine-readable structured data is part of the semantic layer large language models draw from when assembling an answer. Models increasingly lean on it to work out what a page is about before starting to parse any of the content.

“Machine-readable” in this case means schema.org types like Article, FAQPage, and HowTo sitting in the code, semantic HTML5 elements such as article and nav in place of generic div containers, and one H1 followed by H2s and H3s nested in the order the content actually argues. Tables carry real header cells instead of bolded text pretending to be headers, and lists use proper list markup instead of paragraphs with hand-typed dashes.

Given two pages that carry identical facts and sit on identical domain authority, the one wrapped in quality markup is likely to be the one that wins the citation.

Retrieval engines and training checkpoints run on different clocks

Perplexity, AI Overviews, and Bing Copilot re-crawl the web on a schedule; a base model’s training cutoff doesn’t move until the next round of training.

About 44% of AI Overviews citations traced back to content published in the last 12 months, with older pages losing share fast.

A page edited this morning can show up in a live-retrieval answer by afternoon. That same edit sits invisible to a static checkpoint until it retrains, whenever the maker schedules that, since the two systems don’t share a calendar.

Know which clock you’re optimizing for. Chasing live retrieval makes a fast re-crawl matter more than raw page age, but chasing the next training run means speed doesn’t help at all.

The numbers here will keep moving as platforms retrain and cut new licensing arrangements, the way Wikipedia’s and Reddit’s did within a few months of each other. The structure underneath a page doesn’t have to chase any of that. Build a page a machine can read once, and every future model that comes looking will find one already waiting.

Frequently asked questions

Does schema markup guarantee an LLM will cite a page?

No. Schema helps a page parse cleanly and raises its odds of getting pulled into an answer, but it can’t fix weak topical relevance or a domain a model doesn’t trust yet. Structure buys a fair look, not a guaranteed spot.

Do all LLMs cite the same kinds of sources?

No. Retrieval-based answer engines, such as Perplexity and AI Overviews, pull from a live index and can cite something published only hours ago. A base model answering purely from training data can only draw on whatever made it into that training cycle, however old that gets.

How often should a site get re-audited for AI citability?

Match the cadence to the target. A live-retrieval engine rewards checking structure every time a page goes up or changes. A training-run cycle can wait longer between audits.

Does writing for human readers conflict with writing for LLM citation?

No. Clear headings, real list markup, and semantic HTML let a person scan a page and let a model parse its DOM at the same time. In practice, an accessibility audit and a citability audit are the same audit run twice.

References

1. Cornwell, J. (2026, March 11). AI citation data shows there is no universal top source for brands. Search Engine Land. https://searchengineland.com/ai-citation-data-no-universal-top-source-brands-471285 
2. Guides: AI in academic research: How to cite AI. (2026a). Springshare. https://guides.library.stanford.edu/ai_research/how_to_cite_AI 
3. Guides: AI in academic research: Large language models. (2026b). Springshare. https://guides.library.stanford.edu/ai_research/large_language_models 
4. Harsel, L., Aleksandr Drozdov, & Skopec, C. (2025, November 10). The most-cited domains in AI: A 3-Month study. Semrush. Semrush Blog. https://www.semrush.com/blog/most-cited-domains-ai/ 
5. Initiative (WAI), W. C. W. A. (2024). Web accessibility evaluation tools list. Web Accessibility Initiative (WAI). https://www.w3.org/WAI/test-evaluate/tools/list/ 
6. Linehan, L. (2025, October 28). 67% Of ChatGPT’s top 1,000 citations are off-limits to marketers (+ more findings). SEO Blog by Ahrefs. https://ahrefs.com/blog/chatgpts-most-cited-pages/ 
7. Linehan, L. (2026, May 11). We tracked 1,885 pages adding schema. AI citations barely moved. SEO Blog by Ahrefs. https://ahrefs.com/blog/schema-ai-citations/ 
8. Structured data in the AI search era. (2025). Brightedge.Com. https://www.brightedge.com/blog/structured-data-ai-search-era 
9. Tabeling, J. (2026, April 8). AI Overviews optimization guide: Ranking in Google AI Overviews. Search Engine Land. https://searchengineland.com/guide/how-to-optimize-for-ai-overviews 
10. Tong, A., Wang, E., & Coulter, M. (2024, February 22). Exclusive: Reddit in AI content licensing deal with Google. Reuters. https://www.reuters.com/technology/reddit-ai-content-licensing-deal-with-google-sources-say-2024-02-22/