Why do AI assistants cite some sources and not others?
AI assistants cite the pages their search step retrieved that directly support what they are writing: pages that rank for the searches they run, answer the question plainly, can be fetched, and are recent and trusted. Citations show only part of what shaped an answer; engines also use pages they read without citing, and what they learned in training.
Where does an AI answer come from?
An AI assistant's answer draws on two reservoirs. The first is parametric knowledge: what the language model absorbed from its training data, a large snapshot of text taken at some point in the past. The second is retrieval: pages the assistant fetches at answer time through a web search or a grounding service. Only the second can be cited, because only there does the system know which page a statement came from.
So a citation is not a list of everything the model "knows". It is a pointer to the retrieved pages the answer leaned on. An answer can repeat an opinion formed years ago in training data while citing a recent page that happens to agree.
How does retrieval choose pages?
Engines differ in the details, but the flow is broadly similar:
- The question is turned into searches. The assistant rewrites the user's question into one or more search queries, sometimes several with different angles ("best X 2026", "X vs Y review").
- A search index returns candidates. This is often a conventional web search index, so pages that rank well for those queries have a large head start.
- Pages are fetched and filtered. Pages that cannot be fetched (blocked crawlers, paywalls, heavy client-side rendering, errors) drop out here.
- Passages are selected. The system picks the passages that best match the question, usually a few paragraphs, not whole pages.
- The answer is written and attributed. The model composes the answer from the passages and attaches citations to the statements they support.
Each step is a filter. A page is cited only if it survives all five.
Which kinds of pages tend to get cited?
- Pages that answer the question directly. A paragraph that states "X costs €29 per month" is easier to lift and attribute than a price hidden in an image or a script.
- Pages that rank for the engine's searches. Because retrieval usually starts from a search index, search visibility and AI citations are linked.
- Entity and reference pages. Encyclopedia entries and official pages are used to confirm what something is.
- Comparison lists and reviews. For "best X" and "X vs Y" questions, articles that compare options match the question's shape.
- Discussions. Forum threads and community Q&A carry first-hand experience that engines use for "is it any good?" questions.
- Recent pages, where recency matters. For prices, news and fast-moving categories, fresher pages tend to win, and a given citation usually does not last long before newer pages replace it.
Each engine has its own habits. Some lean more on community discussions, others on reference pages or video platforms. These habits also change: an engine can sharply reduce how often it cites a whole platform after an update. Any fixed list of "domains AI loves" goes stale.
Why is my site not cited?
- It does not rank for the searches the engine makes for that question.
- The answer to the question is not stated plainly in the text.
- The crawlers used by AI search features are blocked in robots.txt, or the content only appears after JavaScript runs.
- Independent pages say the same thing more credibly, so the engine cites them instead.
- The page is outdated, and a newer page contradicts it.
Why are citations not the whole story?
If you only analyse cited pages, you miss two things. First, the pages the engine read but did not cite: every retrieved candidate influenced the selection, even if only three got a footnote. Second, the pages that shaped the model's training-time view of your brand: discussions, reviews and articles that may never be cited but set the tone.
A useful way to think about this is a source graph in three tiers:
- Cited pages: explicitly cited in answers. Certain.
- Formative pages: the pages engines are likely to draw on for this subject, such as the search results for the questions people ask, entity pages, and discussions and reviews about the subject. Estimated.
- Owned pages: your own site and profiles, which you control.
Formative pages have no certain weight, so they need an estimate: how often the engine cites that page or that kind of source, where the page ranks for the relevant searches, and how recent it is. The analysis is aggregate by nature: studies of AI answers suggest that the tone of any single cited page says little about the tone of the answer, while the overall balance of many sources, and specific claims repeated across them, matter more.
What does this mean in practice?
- Write owned pages that answer the questions people ask, in plain sentences, with dates on anything that changes.
- Make sure those pages can be fetched without JavaScript and are not blocked for AI search crawlers you want to reach.
- Look at which third-party pages are cited for your category, and at the discussions and comparisons that rank for it. That is where your reputation is being formed.
- Track citations per engine and over time, because habits differ between engines and change.
Tonecast does this analysis automatically for each campaign: it records every page ChatGPT, Claude, Gemini and Perplexity cite, discovers the formative pages behind the answers, gives each an influence weight per engine with the reason it was included, and re-estimates the weights every week. The What shaped it view shows the result.