Short answer
AI assistants usually search first, then pick pages that are relevant to the question, look trustworthy, are recent enough, state the answer clearly and agree with other sources, and then write an answer that cites some of them. The exact formulas are not public, so what we know comes from documentation and studies that show patterns, not rules.
What is an AI citation, and how does one get chosen?
An AI citation is a link or a source name shown beside part of an AI-written answer. To get there, a typical assistant that can search the web goes through the same steps, though each product differs and none publishes its full recipe:
- Interpret the question
The assistant decides whether it can answer from memory or needs live information.
- Search
It turns the question into one or more searches. Google calls the multi-search version query fan-out.
- Retrieve
It pulls candidate pages from an index, either its own or a search provider's.
- Filter and read
It ranks the candidates and reads the most relevant passages, often not the whole page.
- Write and attribute
It writes the answer from those passages and shows links for some of them. Many pages are read and never cited.
Which signals seem to matter?
| Signal | What it means | How well supported |
|---|---|---|
| Relevance | A passage answers this exact question closely | Strong: it is how retrieval works |
| Authority and trust | The site is a known, credible source on the topic | Plausible, backed by correlations, not proven |
| Freshness | The facts are current, which matters most for fast-changing questions | Plausible, little direct evidence |
| Clarity | A direct answer and specific facts, with one idea per section | Supported by the Princeton and SE Ranking studies |
| Structure | Clear headings and sections of about 120 to 180 words | Supported by SE Ranking, observational |
| Consensus | Several independent sources say the same thing | Plausible, little direct evidence |
What does the research show, and how far can you trust it?
- The 2024 Princeton and IIT Delhi GEO paper tested nine ways of rewriting a page. Adding statistics and cited sources raised visibility by up to roughly a third, and keyword stuffing was the worst. Limit: it was a research setup, and real products change over time.
- Ahrefs analyzed about 1.9 million AI Overview citations and found that about 76% of the cited pages ranked in Google's top ten. Limit: it shows a link, not a cause, and it covers Google's feature only.
- SE Ranking found that pages with many concrete data points were cited more than data-light pages, and that sections of roughly 120 to 180 words between headings performed best. Limit: a vendor's observational study.
- An Ahrefs study of 1,885 pages found no meaningful lift in citations after JSON-LD was added. Limit: every page in it already had 100 or more AI Overview citations.
- A critical survey in July 2026 says evidence standards in this field are still unsettled, and most percentages in vendor blogs are unverified.
Why does ranking in Google still matter?
Because retrieval starts from search indexes. Google says its AI Overviews and AI Mode run on its core Search ranking systems, so pages that rank well are the natural candidates. Other assistants use their own indexes or search providers, and OpenAI has not published the full mix behind ChatGPT. You can read what is known in whether ChatGPT uses Google. Either way, a page that is easy to find in several indexes has more routes into an answer.
What can you control?
- Answer first. Put the question in a heading and a direct answer straight below it.
- Be specific and checkable. Use numbers, dates and named sources.
- Publish original material, such as your own data, photos and process.
- Keep facts current and show the date.
- Put the text in the page itself, and let search and AI crawlers in.
- Be consistent and be mentioned. Use the same business details everywhere, and earn reviews, listings and mentions on other sites. Several sources agreeing is the consensus signal.
The practical version of this list is in how to get cited by ChatGPT, and the Google side is in how to appear in AI Overviews.
What should you not read into a citation?
So measure over time, across many questions. Arcline Premium stores AI answers with their sources and shows whether you are mentioned or linked, and it labels the numbers as a guide, not an exact count. See Features or what AI visibility is.