This is placeholder copy for a guide about how assistants pick sources. It exists so the layout can be reviewed with something the length of a real article in it. Swap it for the real thing before this page goes anywhere near a reader.
What the assistant is doing
An answer is a reading exercise, not a lookup. Before it answers, the assistant rewrites the question into its own searches, runs them, and reads a handful of pages. The answer is assembled from what it read, which is why the page matters more than the prompt.
The practical version of that: you are not optimising for a person typing a question. You are optimising for the search the assistant typed on their behalf.
Where people go wrong
Most of the advice in this area is confident and untested. The three habits worth dropping first:
- Writing for a keyword that no assistant ever searches for.
- Treating a mention in an answer as the same thing as a recommendation.
- Changing a page and never checking whether the change did anything.
What to do instead
Start from the answer. Find the query the assistant actually ran, read the pages it cited, and ask what those pages have that yours does not. Usually it is something dull: a clearer heading, a number with a source on it, a date that is real.
Then change one thing, leave a control alone, and look again in a month. That last step is the one everyone skips.
The short version
how assistants pick sources is not mysterious. It is a retrieval problem wearing a new coat, and the work is mostly making a page easy to quote and easy to trust.