AI Overviews are Google's AI-written answer, sitting above every blue link

The process runs in stages rather than a single ranking step. Google retrieves a wide pool of candidate pages for the query, filters that pool for authority and clarity, re-ranks the surviving passages for how directly they answer the question, and only then composes a summary with inline citations to the sources it used.

Retrieval is often broader than the query typed into the box. Google can issue related searches behind the scenes, sometimes called query fan-out, to pull in supporting pages for subtopics the main query implies, then draw the final summary from across that wider set rather than from one search alone.

That is a meaningfully different pipeline from classic search ranking, and it is why a page's position in the AI Overview does not track its position in the ordinary results list one for one.

If nearly a third of citations come from pages nowhere near the first page of results, ranking position alone is not what earns the citation. Clear structure, named entities, and schema that states facts plainly give the synthesis step something safe to lift, which is the same ground AI-readable design covers.

It means a page with real disadvantages in domain authority or backlink count is not automatically shut out of the citation, provided it answers the specific question more directly and legibly than the pages that outrank it.

Practically, that means a single page answering one question clearly, with the answer stated in the first sentence a reader would meet, competes on more even terms in an AI Overview than it ever did in the ranked list, where age, backlink volume, and domain history carry more weight.

Once a page is a retrieval candidate, the deciding factor is usually how cleanly it isolates the answer. A page with a clear heading matching the question, a direct one or two sentence answer near the top, and specific supporting detail underneath gives the synthesis step a passage it can lift with confidence. A page that buries the answer inside a long, unstructured block of marketing copy gives it nothing safe to extract.

Schema that names the actual entities involved, the organisation, the person, the place, adds another layer the system can cross-check against on top of the plain-language answer, not instead of it.