Drift: when a score moves more than its own history explains
We compare each pillar against its own history, and only against runs that are genuinely comparable to it, so an alert means real movement rather than a number crossing an arbitrary threshold.
What it is
After every analysis we compare each measured pillar against its own trailing history for your site, and record a finding when the movement is more than that history explains. There are two cases, because they support different claims. When your past runs varied, we can estimate a normal range and say a drop fell outside it. When your past runs were IDENTICAL, there is no variation to estimate, so we require a much larger drop and report it as a large change rather than as a statistical outlier.
When to use it
It runs automatically. Read drift findings first when reviewing an analysis. They flag what changed, as opposed to what has always been low.
Step by step
- 1Run Analyze a few times so there is a baseline to compare against.
- 2Open Analyze and read the findings list. A drift finding names the pillar, the size of the drop, and either the normal range it fell out of or the fact that a range could not be estimated.
- 3Cross-check What Changed on the same page. Where we can attribute the movement to a measured cause, it says so.
What you will see
One of four findings. “Technical Health dropped 12 points below its normal range”, with the trailing average, the range and how many prior runs informed it, when your history varied enough to compute one. “AI Visibility fell 47 points from a previously steady value” when every prior run read the same number, in which case we say plainly that normal variation could not be estimated. “AI Visibility is now measured, and is no longer comparable to its history” when a pillar switches from an estimate to a direct measurement. And “AI Visibility is measured over a different set of prompts now” when you edit your prompt set. The last two carry zero impact on purpose: the instrument changed, not your site.
Measured, never manufactured
Your first run raises nothing. A first observation is a baseline, never a regression. Beyond that, drift will only compare two runs when four things about them match: the scoring methodology, whether each number was measured or estimated, the sample behind it, and the set of prompts it was computed over. Any of those changing produces a different number without anything on your site moving, so we say which one changed instead of calling it a decline. We would rather miss an alert than send you one that blames you for our own instrument.
Related guides
Next
What Changed, and whyA measured diff between your last two runs, with the cause named whenever we measured it on both sides.