# How Do AI Search Engines Decide Which Pages to Cite?

> AI citation selection is how an AI search engine picks which of the pages it read to cite in its answer. It happens in two stages. Links, relevance and freshness decide which pages the engine finds and reads. Among those pages, it cites the ones other sites link to and that agree with the other sources, not the ones with better formatting or tone.

Format: explainer · Published: 2026-10-31
URL: https://cognovaai.com/insights/how-do-ai-search-engines-decide-which-pages-to-cite

**Term:** AI citation selection

## Key points

- AI engines run their own web searches, read a shortlist of results, and cite only some of the pages they read.
- Links pointing to the specific page are the clearest sign: in our study of 440 AI answers, ChatGPT cited the page with more linking sites 62% of the time.
- Engines favor pages that agree with the other sources they read, because a widely shared claim is safer to repeat; those pages won 54% of head-to-heads in our study.
- Newer pages have an edge, and controlled tests show AI rankers prefer newer dates even when the text is identical.
- Formatting, schema markup, llms.txt and persuasive copy made no measurable difference once a page was on the shortlist.
- For smaller sites, a title that matches the buyer's question and question-style headings help, while a heavy 'we and our' voice hurts.

## How it works

When you ask ChatGPT, Gemini or Google's AI Mode a question that needs current information, the engine does not read the whole web. It turns your question into a few web searches, takes the top results, reads a shortlist of pages, and then cites the ones it relied on. AI citation selection is two contests: getting onto the shortlist, and then getting cited from it.

The first contest looks a lot like classic search. Pages need to be crawlable, relevant to the searches the engine runs, recent and well linked. A 2026 study of 252,000 head-to-head trials across six AI models found topic match, recency and a source's position in the list acted as gatekeepers, while content structure had no effect.

The second contest is where most GEO advice goes wrong. In [our study of 440 AI answers](/insights/what-makes-ai-cite-a-page-a-study-of-440-ai-answers), we compared the pages ChatGPT, Gemini and Google AI Overviews cited with the pages they read and skipped inside the same answers. The page with more sites linking to it won 59% of these head-to-heads, and 62% in ChatGPT. Pages that agreed more closely with the other pages read won 54%. Formatting, tone and persuasion cues stayed close to 50%, the same as a coin flip.

## Why it matters

For a small or mid-sized business, this changes where to spend time. Rewriting pages with more headings, FAQ blocks, schema and punchy copy feels productive, but it does little once AI has already found your page. Earning links to a specific article, covering what the trusted sources in your category cover, and keeping the page current do more.

It also explains why big vendors and publishers keep winning. AI search leans toward established, third-party sources: one study found AI engines drew 74% to 82% of their citations from earned media in the categories tested, far more than Google's own results. Ahrefs found branded web mentions correlated more strongly with AI Overview visibility than backlinks did (0.664 against 0.218). A smaller business has to earn that outside validation rather than claim it.

Finally, it explains a pattern many owners notice: AI-written listicles get cited a lot. In our study, listicles were cited 35% of the times an engine read them, against 17% for comparison pages. Research on neural search models found they rank LLM-generated text higher than human-written versions of the same content, likely because it is more predictable for the model to process. The engines are not penalizing that style.

## Examples

- **The vendor listicle that keeps winning.** In the AI voice agent category we track, the most-cited article was a vendor's own list of the best AI voice agents for lead generation, cited mostly by Google's AI Overviews and Gemini. It names many competitors, matches what the other top pages say, was updated weeks before our test, and other sites link to it. Ranking its own product in the list did not stop it being cited.
- **The well-formatted page that keeps getting skipped.** Many pages engines read but rarely cited had FAQs, tables and schema markup. Their median was zero sites linking to them and 30 days since the last update, against four linking sites and 14 days for the most-cited pages.
- **A small site's opening.** Once we set aside the handful of sites that dominate the category, page-level choices started to matter. Pages using question-style headings won 55% of head-to-heads, and pages whose title matched the buyer's question won 54%. Pages written mostly about 'we' and 'our platform' won only 46%, because they read as a pitch rather than a source.

## Common mistakes

1. **Treating formatting as the fix.** FAQ sections, extra headings and summary boxes help readers, but they do not decide AI citation selection. Fix relevance and links first.
2. **Counting on llms.txt or schema.** Both showed no clear effect in our data once a few popular sites were set aside, and SE Ranking found llms.txt had negligible impact on ChatGPT citations.
3. **Writing for persuasion instead of answers.** Hype, testimonials and calls to action are aimed at buyers, not at an engine choosing sources. Answer the question plainly and keep the pitch in its own section.
4. **Contradicting the consensus to look original.** Engines prefer claims several sources agree on. Cover the common ground first, then add your original data in a clearly labelled section so it builds on the consensus.
5. **Trusting older lab results.** The widely quoted finding that adding statistics and quotations lifts visibility by up to 40% came from AI-rewritten test pages run through an older model. We did not see that effect in real answers, so test changes on your own pages before rolling them out.

## How Cognova AI can help

We measure AI citation selection for your category the same way we did for our own: which pages AI engines read for your buyers' questions, which they cite, and why yours are passed over. You get the pages to earn links from, the questions your content needs to answer, and a way to prove whether each change worked.

See our [GEO & AEO Consulting](/services/geo-aeo) service, or read how we measured our own starting point in [AI Visibility Baseline: 0 Mentions in 2,034 Non-Branded Answers](/insights/ai-visibility-baseline-0-mentions-in-2-034-non-branded-answers).

## Frequently asked questions

### Do backlinks matter for getting cited by ChatGPT?

Yes. In AI citation selection, links to the specific page were the clearest factor in our study of 440 AI answers: ChatGPT cited the page with more linking sites 62% of the time. SE Ranking's study of 216,524 pages found referring domains were the top factor for ChatGPT citations. Links likely help in two ways: they lift the page in the web search results the engine reads first, and they signal that other sites vouch for it.

### Does formatting help a page get cited by AI search?

Not much, once the page is relevant. A 2026 study of 252,000 head-to-head trials across six AI models found content structure had no effect on which source was cited, and our own comparison of cited and skipped pages agreed. Clear formatting still helps human readers, but AI citation selection depends more on links, relevance, agreement with other sources and freshness.

### Does llms.txt or schema markup help AI engines cite my site?

Not on their own. In our data, sites with llms.txt looked more likely to be cited, but the effect disappeared once a handful of popular vendor sites were removed. SE Ranking also found llms.txt had negligible impact on ChatGPT citations. Schema markup showed no clear effect on AI citation selection either. Both are cheap to add, but neither replaces links or relevant content.

### How often should I update a page to keep AI citations?

Update important pages at least every few months with real changes, not just a new date. SE Ranking found pages updated within three months averaged 6.0 ChatGPT citations, against 3.6 for outdated pages. Controlled tests also show AI rankers prefer newer dates. Freshness is a modest factor in AI citation selection, so it works best alongside links and relevance.

## Sources

- [Cognova AI: What Makes AI Cite a Page: A Study of 440 AI Answers](https://cognovaai.com/insights/what-makes-ai-cite-a-page-a-study-of-440-ai-answers)
- [SE Ranking: How to optimize for ChatGPT (ranking factors study)](https://seranking.com/blog/ranking-factors-for-chatgpt/)
- [arXiv: What Gets Cited: Competitive GEO in AI Answer Engines (Vishwakarma et al., 2026)](https://arxiv.org/abs/2605.25517)
- [arXiv: Do Large Language Models Favor Recent Content? (Fang et al., 2025)](https://arxiv.org/abs/2509.11353)
- [arXiv: Generative Engine Optimization: How to Dominate AI Search (Chen et al., 2025)](https://arxiv.org/abs/2509.08919)
- [arXiv: Neural Retrievers are Biased Towards LLM-Generated Content (Dai et al., 2024)](https://arxiv.org/abs/2310.20501)
- [Ahrefs: An Analysis of AI Overview Brand Visibility Factors (75K Brands Studied)](https://ahrefs.com/blog/ai-overview-brand-correlation/)
- [arXiv: GEO: Generative Engine Optimization (Aggarwal et al., 2024)](https://arxiv.org/abs/2311.09735)
