How to measure AI visibility

“It feels like AI mentions us more” is not a metric. Generative engines can change wording, sources, and competitor order on every reply. To run GEO, freeze the questions first, then count what the answers do.

Freeze three things

1. A prompt set. Use sentences sales and support actually hear, not internal jargon. Keep a few in each bucket: category, scenario, comparison, purchase.
2. An engine set. In China that is usually DeepSeek, Doubao, Qwen, and Yuanbao; add ChatGPT and Perplexity for overseas. Watch the same set every cycle.
3. A brand and competitor list. Include aliases. Otherwise “mentioned” and “who was mentioned” will not match.

If those three stay still, weeks become comparable.

The numbers that matter

• Mention rate. Of this prompt set, how many answers name you.
• Citation rate. Beyond the name, is your page used as a source.
• Share of voice. In the same answers, how often you appear versus named competitors.
• Role. Are you the recommendation, a comparison, or a passing counterexample.
• Gap type. If you are missing, is it un-citable copy, a fuzzy entity, or a competitor page that extracts better.

Traffic and keyword rank still help. They do not explain why this wording left you out.

How not to fool yourself

• Retest the same prompts. After you ship a page, ask the original sentences again. Do not swap in “better” prompts that hide the miss.
• Keep the raw answers. Store the text and source links. Disputes go back to the record, not to memory.
• Read weeks, not hours. Indexing and sampling lag. Intra-day swings are usually noise.
• Tie each edit to a prompt. Write down which question the page is supposed to win, then write the page.

CitePath binds the prompt set, the engines, and the retest into one watch: Overview for share, Issues for which content to add, then the same prompts again. Once visibility is a number, GEO stops being a slogan and becomes a weekly close.