Your AI Visibility Score Dropped. Now What?
Is the drop real or is it noise?
Most single week drops are variance. AI answers are probabilistic, so a score moving a few points without any cause is normal behaviour rather than a signal. Before you investigate, check whether the move is larger than your usual spread. If it is not, there is nothing to explain.
The first question is always the confidence band
A drop matters when it falls outside the range your score normally occupies. Your confidence band is that range, calculated from how much the measurement varies across runs.
A score of 34 moving to 29 means nothing if your band is 28 to 40. The same move means a great deal if your band is 33 to 36. The number alone cannot tell you which situation you are in.
This is why a single point score is a poor alert trigger. The useful trigger is a move that leaves the band and stays out for a second measurement.
Why two consecutive readings matter more than one
One reading outside the band is a candidate. Two in a row is a finding. Variance produces outliers, so a single extreme value is exactly what randomness looks like.
Waiting for the second reading costs you a week and saves you from most false alarms. In a weekly cadence that is an acceptable trade. The alternative is a team that chases every outlier and loses trust in the measurement.
If the second reading returns inside the band, log it and move on. Do not write a story about the week it dipped.
The order to check causes
When a drop holds for two readings, work from the cheapest check to the most expensive.
Did the question set change? Any edit to your prompts changes what you are measuring. This is the most common cause of a drop that has no other explanation, and it takes thirty seconds to rule out.
Did a competitor move? Look at whether your loss matches someone else's gain on the same questions. If one rival picked up exactly what you lost, the cause is competitive rather than technical.
Did the sources change? Check whether the pages that used to be cited for your questions are still cited. A listicle that dropped you, or a review page that got rewritten, moves visibility without touching your site. AI citation analysis covers how to read this.
Can the crawlers still reach you? Verify that AI crawlers are not blocked and that your pages render without JavaScript. This is rarer but it is the one cause that can take visibility to zero. The technical checklist has the steps.
Did the engine itself change? Model updates shift answer patterns across a whole category. If every brand you track dropped in the same week, the cause is upstream and there is nothing to fix on your side.
Why an engine level view finds the cause faster
A total score hides where the loss happened. Open the per engine view before anything else: a drop concentrated in one engine points somewhere different than a drop spread evenly.
One engine falling while the others hold usually means a source or index change specific to that engine. All engines falling together usually means something on your side, or a category wide shift.
Our methodology page explains how the engine scores roll up into the total, which is what makes this split readable.
What not to do
Do not rewrite content in the week of a drop. You will not be able to tell whether the recovery came from your change or from the number returning to its normal range.
Do not change the question set to make the number look better. That is not a recovery, it is a different measurement.
Do not report a single week move to leadership as a trend. Report the band, the direction over four weeks, and what you are testing.
Frequently asked questions
How big does a drop have to be before I act?
Bigger than your confidence band, and sustained for two consecutive measurements. A move inside the band is not evidence of anything, regardless of how large it looks in isolation.
My score dropped on one engine only. Is that a problem?
It is a narrower problem. Single engine drops usually trace to the sources that engine reads rather than to your site. Compare which pages were cited before and after on that engine.
Everything dropped at once across all competitors. What happened?
That pattern points to the engine, not to you. Model updates and answer format changes move a whole category at the same time, and there is no action on your side that reverses it.
Can a score recover on its own?
Yes, and it often does when the original move was variance. A real decline caused by lost citations or a crawler block does not recover without the underlying cause being fixed.