What is a baseline?

Before you can say a citation rate improved, you need a number it improved from. That starting number is the baseline, and everything downstream, whether a change worked, whether a decline is real, depends on how solid that first number was. A baseline measured once, from a small prompt basket, on a single day, inherits all the sampling noise that any single measurement carries.

The failure mode is setting a baseline during an unusually good or unusually bad stretch and not knowing it. If your baseline happened to land during a model update that temporarily boosted your citations, everything after it will look like a decline, when really the baseline itself was the outlier. This is why a baseline built by averaging a few measurements over a short stretch before the change is stronger than one built from a single reading, since averaging smooths out the chance that day one happened to be unusual.

A baseline isn't a fact about your brand, it's a starting estimate with its own margin of error, and any comparison against it inherits that uncertainty.