# Measuring AI Visibility: Prompts, Citations and Share of Voice


# Measuring AI Visibility: Prompts, Citations and Share of Voice

## What this helps you with

“Are we visible in AI?” is not answerable with one screenshot. You need a repeatable set of questions, run across models, scored the same way every week.

## Three steps

1. **Define 10–30 buyer questions** (e.g. “best silicone dildo for beginners”, “which store ships to Germany”).
2. **Run them across models** (ChatGPT, Gemini, Claude, Perplexity, Copilot, Grok, DeepSeek, Qwen, Doubao, ERNIE, Kimi and GLM) and record: mentioned? cited? which URL? sentiment?
3. **Track weekly**: mention rate, citation share, competitor share, and which pages AI cites.

## The four numbers that matter

- **Mention rate**: how often your brand appears;
- **Citation share**: how often your URL is the source;
- **Share of voice**: your mentions vs competitors;
- **Source mix**: which pages AI actually cites (product, FAQ, blog, third-party).

> One run is an anecdote. A fixed prompt set, run weekly, is a measurement.

**Want the method and pitfalls?** Read the expert version: [[ai-visibility.expert]]


---


# AI Visibility (Expert): Prompt Sets, Citation Share and Reproducibility

## Chapter 1: Measurement design

### 1.1 Fixed prompt set

Define prompts by intent, not by brand:

- category discovery (“best X for Y”);
- comparison (“A vs B”);
- policy questions (“does store ship to Z”);
- branded checks (“is brand X legit”).

Version the prompt set and hash it. Changing prompts invalidates trend comparisons.

### 1.2 What to record per run

| Field | Why |
|---|---|
| prompt hash | reproducibility |
| engine/model + version | model drift |
| answer text (truncated) + hash | audit |
| brand mentioned | visibility |
| cited URLs | citation share |
| rank position | share of voice |
| sentiment | brand perception |

### 1.3 Metrics

- mention rate = mentioned runs / total runs;
- citation share = runs citing your URL / total runs;
- share of voice = your mentions / (your + competitor mentions);
- source mix = cited URL types (product/FAQ/blog/third-party).

Always publish sample size and time window. A rate without N is noise.

## Chapter 2: Pitfalls and operations

### 2.1 Model drift and sampling

Models change and answers vary. Run each prompt 2–3 times, record all runs, and report medians/ratios rather than single answers.

### 2.2 Don’t overclaim

- visible in ChatGPT ≠ visible in Perplexity;
- citation ≠ recommendation;
- AI answers are not rankings, and correlation is not causation.

### 2.3 Operational loop

1. weekly run → 2. detect changes → 3. inspect cited pages → 4. fix content/schema → 5. re-run and compare.

Keep raw answers (hash + capped text) for at least 90 days.

**Back to beginner version:** [[ai-visibility.novice]]