How do AI assistants decide which sources to cite?
Most assistants search first, read a handful of results, then generate an answer from the passages that best match the question. Selection turns on three things: whether the page sits in the search index the assistant queries, whether a passage on it directly answers the phrasing used, and whether the claim is corroborated elsewhere. Domain authority matters, but passage relevance usually matters more.
Retrieve, read, write, then attach citations
The sequence inside most assistants looks like this. Your question is rewritten into one or more search queries. Those queries hit a search index. A set of results is fetched and chunked. A re-ranking step scores the chunks against the question. The model writes an answer from the surviving chunks. Citations are then attached to sentences, sometimes after the fact.
That last detail explains a lot of odd behavior. A citation is a claim about where a sentence came from, generated by a system that is estimating. It is usually right and occasionally attached to the wrong source entirely.
Two practical implications follow. First, if the search index the assistant queries does not contain your page, nothing downstream can save you. Second, because chunks are re-ranked against the question rather than against a general notion of quality, a modest page with one excellent passage can beat a strong page whose relevant material is buried.
Query fan-out is the part most people miss
One typed question frequently becomes several different searches. Someone asking whether their furnace is worth repairing may trigger searches about furnace lifespan, repair versus replace thresholds, and heat exchanger failure symptoms.
You are therefore competing for phrasings the user never typed. A page that covers one question superbly and ignores the three obvious follow-ups gets retrieved for one of four searches. Coverage of the adjacent sub-questions is a retrieval strategy, not padding. That is the argument for clusters of individual answers rather than one long article.
Corroboration beats confidence
Models lean toward claims that appear consistently across independent sources. A number only your site states is less likely to be repeated than one that three credible sources state the same way.
This is uncomfortable if your positioning depends on a contrarian claim. The workaround is not to shout louder. It is to state the contrarian point plainly, explain the reasoning, and be the clearest available source on the surrounding facts everyone agrees on.
Freshness works similarly. Assistants tend to prefer recently updated sources for anything that could plausibly have changed - regulations, equipment, availability, cost drivers. A page dated four years ago competes at a disadvantage even when the content is still correct, which is an argument for genuine review cycles rather than for touching the date field.
What honest practitioners admit they cannot see
No provider publishes its selection weights. There is no API that returns why you were or were not cited. Everything above is inferred from observed behavior and published documentation, and it changes without notice.
Treat anyone offering a precise formula, a guaranteed placement, or a claimed relationship with an AI provider as a warning sign. The defensible approach is to make the inputs good and measure the outputs by sampling. That is how we approach AI search work.
Topics: citations · retrieval · answer engines · how it works
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.