When an AI analyst answers a market question, it can do so in two very different ways. It can reason over what it was handed in the prompt, or it can go and gather more first. The second takes longer, costs more and is worth asking for when the question deserves it.
What gets gathered
A researched answer typically pulls the current computed state relevant to the question, then searches for public reporting around it, then reads its own archives for comparable past cases. Each of those is a separate source with a separate reliability, and a good answer keeps them distinct rather than blending them into one confident paragraph.
Why citation matters more than eloquence
The single most useful property of a researched answer is that it names what it read. Not because attribution is polite, but because it is checkable. An answer that cites five sources can be verified in five clicks; an answer with no sources can only be believed or not.
This also constrains the model usefully. A system told to cite what it used is far less inclined to assert a number it does not have, because there is nowhere to attribute it to.
What it still cannot do
Gathering more does not turn an explanation into a prediction. A deep read can tell you what the structure is, what has followed structures like it, and what the wire is reporting. It cannot tell you what happens next, and an answer that starts doing so has left the evidence behind.
It is also bounded by what was actually gathered. If a source failed, the honest version says so rather than quietly answering from memory.