Proving a change worked
Also known as: baseline, attribution window
Establishing a baseline before you change anything, and knowing how long to wait — because in a noisy channel, an unmeasured improvement is indistinguishable from a lucky week.
Last revised 2026-09-07. Also available as markdown — request this URL with Accept: text/markdown.
In plain English
The most common way AI-visibility work goes wrong is not the work; it is the absence of a before. Someone rewrites twenty pages, the number goes up, and nobody can say whether the pages or the variance did it.
A baseline costs one extra cycle of sampling and settles that argument permanently.
How to fix / set it up
- Sample two cycles before touching anything, and write down the range you saw.
- Change one class of thing, and record the date.
- Wait at least four weeks before reading the result.
- Compare against the recorded range, not against the single most recent cycle.
The technical detail
Sample the frozen prompt set for at least two cycles before changing anything. Two cycles show you the size of the noise, which is the number you need in order to recognise a real move later.
Change one class of thing at a time — passage structure, or entity signals, or coverage — otherwise you learn that something worked without learning what.
Expect a lag. Retrieval indexes refresh on their own schedule; four to eight weeks before a content change is reliably visible is normal, and longer where the change is to how you are described elsewhere.
Compare like with like: the same prompts, the same engines, the same country. A prompt added mid-experiment invalidates the comparison for every prompt.
Keep a dated log of what changed. Six weeks later, nobody remembers which Tuesday the rewrite shipped, and the chart alone cannot tell you.
FAQ
How big a move counts as real?
Bigger than the spread you recorded in your baseline cycles. That is the whole reason for measuring the baseline rather than assuming it.
Can I speed up the lag?
Somewhat: updated sitemaps and re-indexing requests help the search-backed engines. Nothing makes a training-shaped signal move quickly.