Query fan-out is what happens in the moment between someone asking a question and a machine answering it. The question is not searched as it was typed. It is broken into several smaller and more specific searches, run at once, and the passages that come back from all of them are assembled into a single written answer.
Ask an assistant whether a particular kind of extension needs planning permission, and it may quietly run searches for the permitted-development rules, for the exceptions that apply in conservation areas, for typical costs, for how long approval takes, and for what happens if the step is skipped. Five searches, five sets of passages, one answer — and the person who asked sees none of the machinery.
This dismantles the habit of aiming a page at a phrase. A page is no longer competing for the question that was asked; it is competing to be retrieved for one of its parts, and its author never finds out which part. A page that ranks well for its target phrase but says nothing about cost, timescale or exceptions is simply not present for the four sub-searches that also fed the answer.
The response is coverage rather than targeting. A page that answers the main question does better when it also answers the questions travelling alongside it: what it costs, how long it takes, what is excluded, who it is not suitable for, what the alternative is, what usually goes wrong. Each of those becomes a passage that can be retrieved on its own, which is why this pairs so closely with writing passages that stand up alone in the first place.
The plain version: it used to be one answer to one question, with the question audible first. Now the preparation is for a whole conversation nobody will be present at, in which any part of what a page knows might turn out to be the part that gets quoted.
