LLMentions

Methodology

AI answers are non-deterministic: the same prompt asked twice can name different brands. A single answer is an anecdote, not a measurement. LLMentions is built around that fact.

Every rate is a proportion with an interval

A mention rate is k mentions in n sampled answers. We always report it with a 95% Wilson score interval, which behaves well at small n and near 0% or 100%. For example:

Mentioned in 3/10 answers: 30% (95% CI 11–60%).

The wide interval is the point: ten samples cannot tell 30% apart from 50%. Use samples of 5 or more per prompt and engine, and more prompts, to tighten it. stats_mention_rate computes the interval for any k/n.

Changes are tested, not eyeballed

Run-over-run changes use a two-proportion z-test. Reports include change.delta and change.significant, and you should treat a change as real only when significant is true. Overlapping intervals mean there is no evidence of change.

Every observation is labelled with its source

source Meaning
api The provider's official API with its own native web search or grounding (OpenAI web_search, Gemini Google Search grounding, Claude web search, Perplexity, xAI)
ui What a real user sees in the consumer app (ChatGPT, Google AI Mode), via a licensed data provider
parametric API with search disabled: what the model "knows" from training

API and app answers differ, sometimes a lot, so they are never pooled silently. We call each provider's native search rather than a normalizing gateway, because each engine's own citation behaviour is the signal.

Mentions and citations

Share of voice is your mentions divided by all tracked-brand mentions.