AI engines choose which sources to cite in three stages: they retrieve candidate pages from a search index, select the passages that best answer each part of the question, and attach links to the pages those passages came from. A page gets cited when the engine can access it, finds it relevant to the query or a sub-query, can extract a clear answer from it, and judges it trustworthy.
The citation pipeline
flowchart TD
A[User prompt] --> B[Query rewriting and fan-out into sub-queries]
B --> C[Retrieval from search index: Google, Bing or own crawler]
C --> D[Candidate pages fetched]
D --> E{Passage selection: relevance, clarity, trust, freshness}
E -->|Selected| F[Passages used to write the answer]
E -->|Not selected| G[Discarded]
F --> H[Answer with source links to selected pages]
Stage 1: Query rewriting and fan-out
The engine turns the prompt into one or more search queries. Complex prompts are split into sub-queries. Google describes this as query fan-out for AI Overviews and AI Mode, and other engines do something similar. A prompt such as “best SIEM for a 500-person healthcare company with a small SOC” may become searches for SIEM vendors, healthcare compliance requirements, managed SIEM options and pricing.
Implication: you can be cited for a sub-question even if you do not rank for the full prompt.
Stage 2: Retrieval
Each engine searches an index:
| Engine | Main retrieval source |
|---|---|
| Google AI Overviews, AI Mode, Gemini | Google’s index |
| Microsoft Copilot | Bing |
| ChatGPT search | Its own search system and crawler (OAI-SearchBot), plus partner data |
| Perplexity | Its own crawler (PerplexityBot) and index |
| Claude, Grok, others | Their own search integrations |
If a page is not in the index the engine uses, it cannot be cited. This is why blocked crawlers, JavaScript-only rendering and indexing problems are the most common reasons for zero citations. OpenAI, for example, states that pages blocking OAI-SearchBot will not appear in ChatGPT search answers.
Stage 3: Passage selection
From the retrieved pages, the model picks the passages that best answer each sub-query. The signals that matter in practice:
- Relevance: the passage addresses the exact sub-question.
- Clarity and extractability: a direct statement, definition, list or table that stands on its own.
- Specificity: concrete facts beat general claims.
- Trust: the source looks authoritative for the topic (known brand, expert author, cited elsewhere).
- Freshness: for time-sensitive topics, recently updated pages are preferred.
- Consistency: claims that match what other sources say are safer to repeat.
Research supports the content-side signals. The GEO paper from Princeton and collaborators found that adding citations, quotations and statistics improved a source’s visibility in generative engine responses, while keyword stuffing did not.
Stage 4: Attribution
The engine writes the answer and links the pages that supplied the passages. Some engines link nearly every claim; others show a short source list. Brands may also be mentioned without a link when the model draws on training knowledge.
Why third-party sources get cited so often
For evaluation prompts (“best X”, “X vs Y”, “alternatives to X”), engines tend to prefer sources that compare several vendors: review platforms, analyst-style roundups, comparison articles and community discussions. A vendor’s own page is one-sided by design. That is why B2B brands with good websites can still be missing from AI shortlists. The fix is to be present and described accurately on the sources engines already cite for your category.
Why your page might be skipped
| Symptom | Likely cause |
|---|---|
| Never cited by any engine | Crawler blocked, page not indexed, content rendered client-side |
| Cited by Copilot but not Google | Stronger in Bing than in Google’s index, or vice versa |
| Retrieved but not quoted | Answer buried, vague or split across tabs and images |
| Cited for definitions, not recommendations | Educational content is strong, but third-party sources do not list you |
| Cited with wrong facts | Outdated information on your site or third-party pages |
Checklist to become a citable source
- Allow search-focused AI crawlers (OAI-SearchBot, PerplexityBot, Bingbot, Googlebot) in robots.txt and at your CDN or WAF.
- Serve key content in server-rendered HTML.
- Make sure priority pages are indexed in both Google and Bing.
- Answer each target question in the first 40 to 60 words of the relevant section.
- Use lists, tables and definitions for key facts.
- Add sourced statistics and named expert input where relevant.
- Keep update dates current on time-sensitive pages.
- Get listed and described accurately on the review sites, roundups and communities engines cite for your category.
How to find out what engines cite today
Run your priority prompts on each engine and record the cited URLs. Group them by domain type: your site, competitors, review sites, media, forums. The pattern tells you whether your gap is on-site or off-site. The how AI engines cite page goes deeper on per-engine behavior, and AI citation patterns explained covers common patterns by query type. GrackerAI captures the cited URLs for every tracked prompt across up to 10 engines, which makes this analysis repeatable.
For the metrics that follow from this, see what citation rate in AI search is.
Frequently asked questions
Do AI engines cite sources from training data?
Usually not. Citations come from live retrieval. When an answer relies only on what the model learned in training, it may mention brands with no source links.
Why do different AI engines cite different sources for the same question?
They use different indexes (Google, Bing or their own crawlers), different query rewriting and different source preferences. A page strong in Bing may be cited by Copilot but not by Google AI Overviews, and vice versa.
Can I pay to be cited?
Not in organic answers. Some engines are testing labeled ads near answers, but organic citations come from retrieval and selection, not payment.
Does domain authority decide citations?
It helps a page get retrieved, but it does not decide alone. Smaller sites with a clearer, more specific answer are often cited over larger sites with vague pages.
