Measuring brand mentions in Gemini, and why it is not Google Search

The most common reporting mistake on Google’s AI products is treating them as one thing. Gemini, Google AI Mode, and Google AI Overviews come from one company and behave differently enough that a single “Google AI visibility” number is close to meaningless.

Three products, not one

Gemini is a standalone assistant. Someone opens it deliberately, the way they open ChatGPT, and asks a question in full sentences. There is no search results page underneath.

Google AI Mode is a conversational search experience inside Google, returning an extended generated answer with sources.

Google AI Overviews is a feature of the ordinary results page. It appears above the blue links on some queries and not others.

The user is in a different posture in each. Someone in Gemini is in a conversation. Someone looking at an AI Overview typed a short query and is deciding whether to scroll. Visibility in one does not predict visibility in another, which is exactly why they should be tracked as separate surfaces and never averaged.

What to measure here

Gemini answers tend toward completeness, which affects how mentions behave. Being named is a lower bar on a surface that lists several options, so the useful signals shift:

• First-mention position matters more than on sparser surfaces, because being named eighth in a thorough answer is not the same as being named first.
• Share of voice is more informative than raw mention rate, because a long answer that names everyone makes mention rate approach 100% for the whole category and stop discriminating.
• Sentiment carries real weight, because there is enough text about each named brand for the characterization to be meaningful.

If your Gemini mention rate is high and your share of voice is unremarkable, you are being listed rather than recommended. Those need different responses.

Phrasing sensitivity

Conversational surfaces are sensitive to how a question is asked in a way that short search queries are not. “Best X for Y” and “what should I use for Y” can return different companies.

This is a prompt-set design problem, not a measurement problem. Write the questions the way buyers actually type them into an assistant, in full sentences, and hold them constant. A prompt set of three-word search queries measures something real, but it is not what people ask Gemini.

The measurement caution

Every surface here is non-deterministic, and comprehensive answers add a specific wrinkle: a long answer that names ten companies produces a high mention rate for all ten, which makes the metric look stable while telling you very little.

Watch position and share of voice on this surface, use mention rate as a floor rather than a headline, and compare completed runs rather than single answers. The general rule holds here as everywhere: one answer is an observation, not a ranking.