Key point. A rise from 20 to 32 mentions on 100 prompts is not yet evidence of change: the 95% range on the change runs from about 0 to +24 points. The test reads that range, not the headline number, and says PASS, FAIL or UNDERPOWERED.
Did my AI visibility really change?
Free calculator · runs in your browser
Did my AI visibility really change?
Enter two readings of the same prompt set: how many answers mentioned you, out of how many prompts. The test says PASS, FAIL or UNDERPOWERED, and how many prompts would settle it.
How the verdict is decided
The test computes a 95% range for the change (Newcombe's method for the difference between two rates, each from a Wilson score interval) and compares the whole range, not the point, with your threshold:
- PASS only if the lower end of the range is above the threshold.
- FAIL only if the upper end of the range is below the threshold.
- UNDERPOWERED otherwise: the range still includes the threshold, so this data cannot tell. That is a statement about the sample size, not about your visibility. The fix is more prompts, not a conclusion.
It treats the two readings as independent. If you used exactly the same prompts both times, the true range can be somewhat wider or narrower. Count prompts, not runs: repeat runs of one prompt are not independent answers.
Built by Elementera. Lenz tracks your mention rate over time.
How do I know if a change in AI visibility is real?
Enter two readings of the same prompt set: how many answers mentioned you, out of how many prompts, before and after. The test computes a 95% range for the change and compares the whole range with the smallest change that matters to you.
- PASS only if the lower end of the range is above your threshold.
- FAIL only if the upper end is below it.
- UNDERPOWERED otherwise. The data cannot tell yet, and the answer is more prompts, not a conclusion.
Why not just compare the two percentages?
AI answers vary from one run to the next and from one prompt to the next. A month-on-month move of several points can appear with nothing changed. Reading the range instead of the point protects you from reporting noise as a win, and from dropping work that may be helping.
Why is UNDERPOWERED not a failure?
An UNDERPOWERED result is a statement about the sample, not about your visibility. The rule is symmetric: a range that includes the threshold cannot be read as a pass either. The test also estimates how many prompts per reading would settle it if the rates held.
FAQ
What threshold should I use?
Zero asks whether there was any real change. If only a rise of 5 points would justify the effort, enter 5.
Can I test a drop?
Yes. Choose "Did it go down?" and the test reads the change in that direction.
Which method does it use?
Newcombe's method for the difference between two rates, each from a Wilson score interval, at 95%. It treats the two readings as independent samples.
Work with us
Want it done for you? Book a call about managed AEO services. Want to see where you stand first? Get your free AI visibility report from Lenz.