Because nobody can tell it is wrong. A cited answer that is wrong points at a fixable article; an uncited one that is wrong is indistinguishable from a right one, so every previous answer becomes suspect too.
Make citation mandatory rather than encouraged. An answer the assistant cannot attribute to a source is not sent — that is a rule enforced outside the model, not a prompt asking nicely.
This means it will refuse things. That is the feature: a refusal costs a handover, and a wrong answer costs trust.
Upward if wrong answers appear. Lowering it hides them rather than fixing them.
Yes, by topic — it is your writing queue.
Refunds, credits, dates and exceptions — by rule.
Half an hour with your own data usually saves reading three of these. The guides will still be here afterwards.