Gemini app prompt volatility tracking
Gemini's volatility has two components: the usual generative variance, plus whether grounding fired. We report both, so you can see which prompts are unstable in wording and which are unstable in retrieval.
How it renders
Derived across the sample set, with grounding_rate reported alongside.
Sampling
Volatility is what tells you the right n for every other field. We recommend using it to allocate credits: sample noisy prompts deeply, stable prompts lightly.
Gotcha
What the API returns
A single collection call against Gemini app for the prompt volatility field, sampled n times and returned as one object per prompt.
{
"engine": "gemini",
"field": "prompt_volatility",
"prompt": "best ai visibility tracking tool",
"country": "GB",
"language": "en",
"n": 30,
"volatility": {
"jaccard": 0.42,
"mention_variance": 0.11,
"recommended_n": 75,
"band": "moderate"
}
}Response fields
| Key | Type | Notes |
|---|---|---|
| jaccard | float | Mean pairwise citation-set overlap across samples. |
| mention_variance | float | Variance in brand presence across samples. |
| recommended_n | integer | Samples needed for a ±10 point interval on this prompt. |
| band | string | stable | moderate | volatile. |
Gemini app prompt volatility FAQ
- What drives Gemini volatility?
- Mostly grounding inconsistency rather than model temperature. The split makes that visible.
- Which prompts need more samples?
- Those with grounding rates between roughly 30% and 70%. We flag them automatically.
Find the Gemini prompts worth trusting
We will run a volatility pass and tell you which of your prompts are stable enough to report on.
Other Gemini app data
Across every engine
Compare how prompt volatility behaves on every engine we track.
Prompt volatility across all engines →