A team can believe GEO is done while answer engines still do not cite them, because the picture of the system and the system differ.

A wrong boundary is the usual cause. If the boundary is "this article", conflicting sources, the glossary, and neighboring pages sit outside it. An answer engine reads those pages together. Definitions that fight each other hurt more than plain wording. Draw the boundary around every citable page about the same subject.

The model is not the world. Scorecards, sample prompts, and dashboard charts are models. They help you see, and they leave things out. If the sample questions stay fixed while real questions move, the chart still looks stable. Resample with new questions, and compare them with the sentences the annotation demo shows a model actually grabbed.

Nonlinear means a small change suddenly matters, or a large change does not. Adding one decisive source may move citations. Three more polish passes may move nothing. Evaluate the kind of evidence that is missing, not the number of edits.

After each review, write three lines: where the boundary was drawn, which chart may be lying, and which kind of edit has already proved almost useless.