The same effort moves a system differently depending on where you push. In content engineering the pushes stack from weak to strong.

Parameters are the weakest. Swapping synonyms, raising a word count, or nudging a title barely changes which facts exist or when they get cited. Parameters are for the last polish, not for GEO itself.

Information flow is stronger. Putting a source after the fact, giving a comparison an explicit dimension, and writing steps in order change what an answer engine can read, not only the wording. The annotation demo checks that flow: are the sentences the model grabbed the ones you thought mattered?

Feedback rules are stronger still. What must be sent back, which signal changes the next piece, and how long you wait before reading citations decide whether the system corrects itself. Without that loop, a correct information flow does not stay correct.

Goals are the strongest push, and the one teams touch least. If the goal is still "rank in search results," every other change gets pulled back into ranking. Write the goal as accurate, checkable, and restatable, and parameters fall behind on their own.

Before the next piece, write which layer you intend to move: parameter, information, feedback, or goal. If three pieces in a row only move parameters, stop and change the standard.