How do AI assistants actually choose which sources to cite in an answer?

Two separate systems are at work: the model's training data (slow to change, favours brands widely written about) and live retrieval, where the assistant runs web searches (ChatGPT leans on Bing's infrastructure) and cites pages it can extract cleanly. Citation studies show each engine has its own diet: Reddit and Wikipedia are heavy in ChatGPT, Perplexity rewards fresh structured pages, and Google's AI features pull from its normal index. The founding GEO research found that adding statistics, quotations, and clear structure lifts your odds of being cited by 30-40%.

A quick orientation. The real value is below: resources worth your time, from people who've actually done it.