Reading changes and variance

AI answers vary between runs even for identical questions. Citeroot is built to show that rather than hide it.

#Intervals

Every rate carries a Wilson 95% interval that depends on the number of answers behind it. Few answers means a wide interval.

#"Within normal variation"

The Overview compares the last 28 days with the 28 days before. A change is called up or down only when the two 95% intervals do not overlap. Otherwise it's labelled within normal variation — it's not a finding. With fewer than five answers in either window the comparison says there isn't enough data.

#What isn't compared

• Citeroot compares across the whole panel and all engines. Use the per-engine and per-stage tables to look at one slice, and remember those have fewer answers each.
• It doesn't automatically mark when a provider changes the model behind an engine. The model name is stored on every answer, so you can check it on the answer page.
• It can't tell you why a number moved.

#Reading a drop

Before assuming a cause, check:

1. Did the number of answers change?
2. Did you change the prompts or competitors?
3. Did the engines or country in Project settings change?
4. Did competitors' mentions rise?
5. Did a crawler-access setting on your site change?

See also the blog post Why your AI visibility score moved.

#Causality

Citeroot records what the answers said and when. It doesn't claim that anything you did caused a movement; that judgement stays with you.