Grounded AI assistants choose sources by running a retrieval query against a search index, ranking the returned pages using traditional relevance and authority signals, and then having the model select the passages it can quote with the highest attribution confidence.
The pipeline is documented at a high level by each provider. Perplexity's public FAQ describes retrieving from the open web and citing the pages it used. Google's AI-features documentation confirms that AI Overviews are generated on top of the same web index that powers classic search results, with the same crawl and quality guidance applying. OpenAI's crawler documentation lists OAI-SearchBot as the fetch layer for ChatGPT Search results.
Three signals show up across all of them: retrievability (the URL must be in the index and reachable), relevance (the passage must directly answer the user's question), and identity (the source must be attributable to a named entity). This is why QAPage markup with an author Person node and sameAs links, plus a self-canonical URL, disproportionately helps: it collapses the identity check.
The engines also prefer diversity — most answers stitch together two to five sources. Being one of those sources is achievable; being the only source is rare and usually reserved for primary documents.
Nothing in the public documentation supports the folk claim that a specific "AEO tool" or paid submission gets a page cited. The mechanics are boring: be crawlable, be authoritative, mark up the answer.