Measurement

Trend or wobble? Reading week-to-week movement

Your visibility fell from 40% to 30%. Panic, or shrug? A simple rule based on how many answers sit behind the number, with a worked example and a chart you can keep beside your dashboard.

AI answers vary from run to run, even when nothing in the world has changed. So every visibility percentage wobbles a little. The question every report needs to answer is whether this week’s movement is bigger than the wobble. Fortunately there is a simple way to tell, and it depends mostly on one number: how many answers sit behind the percentage.

A worked example

Made-up numbers, but a very common situation. A brand tracks 10 prompts on 3 platforms, once a week. That is 30 answers per week.

12 of 30answers mentioned the brand last week: 40%
9 of 30answers mentioned it this week: 30%
±25 points95% margin for a change between two 30-answer weeks
Worked example. The margins are explained below.

A ten-point drop sounds alarming. But with only 30 answers each week, two weeks can differ by more than that through chance alone. This drop is inside the wobble. It is worth watching, not worth an emergency meeting.

How big is the wobble?

For a percentage built from yes-or-no results (mentioned or not), the typical margin shrinks as the sample grows. Using the standard approximation for a 95% margin at a rate near 50%:

  • 10 answers±31 points
  • 30 answers±18 points
  • 100 answers±10 points
  • 300 answers±6 points
  • 1,000 answers±3 points
95% margin for one week’s rate near 50%: 1.96 × √(0.25 ÷ answers). Rates near 0% or 100% wobble less. It assumes answers are independent; real answers often move together, so treat these as the minimum.

Quadruple the answers and the wobble halves. That is why the same tool can feel jumpy on a small prompt set and steady on a large one.

The rule of thumbWhen you compare two weeks, the margin is about 1.4 times the figure in the chart (both weeks wobble). Treat a change as real when it is bigger than that and it holds for two runs in a row. Anything else is worth noting, not reporting.

Getting more answers without paying more

Your numbers are too jumpy to read. What can you change?
Group promptsReport on a group of ten related prompts, not one prompt at a time.
Combine weeksA four-week rolling view has four times the answers of one week.
Combine platforms carefullyMore answers, but only if the platforms behave similarly.
Look at the countsPrefer tools that show “12 of 30”, not just “40%”.

When a small move is still worth checking

Statistics tells you whether a change could be chance. It doesn’t tell you whether something real happened. So always open the saved answers when:

  • the brand disappears completely from a question where it was always present;
  • a new rival or a new cited page appears in several answers at once;
  • the change lines up with something you did, such as a launch, a page update or press coverage.

Those patterns can be real even when the headline percentage barely moves. The weekly routine builds this check into Monday’s triage.

Sources

  1. Scrunch: Metrics guide (sample sizes shown with rates) viewed 11 Nov 2026
  2. NIST/SEMATECH e-Handbook of Statistical Methods: confidence intervals for proportions viewed 11 Nov 2026

Directory figures come from Hyperfind’s profiles of 34 tools, not the whole market. Prices follow each provider’s published offer on the date shown in its profile.

Tools in this story

Price your workload ↗Compare tools ↗

Keep reading

Measurement

What an AI visibility tool actually measures

Strip away the dashboards and every tool is counting the same few things in an AI answer. Here is the anatomy, and the questions to ask about each number.

ChartWorked example
Measurement

Why the same question gets a different answer

Ask an AI assistant the same thing twice and the brands can change. That is not a bug in your tool. Here are the six things that move an answer, and what each means for measurement.

Decision tree
Measurement

Daily or weekly? What the extra runs actually buy

Daily tracking collects seven times the answers of weekly tracking, and usually costs more. Here is when that extra data earns its keep, and when a weekly rhythm tells you just as much.

ChartDecision treeWorked example