How to track AI share of voice correctly (most dashboards are lying to you)

Most GEO dashboards report a single share of voice number. That number is almost always misleading, and acting on it leads to the wrong investments. Here is how to measure correctly.

What share of voice actually is

Share of voice in an AI context is: of all the brand mentions a model produces when answering prompts in your competitive set, what percentage are yours.

That definition contains three variables you must define explicitly: which prompts, which competitors, which engine. Average those three away and you get a number that moves for reasons you cannot diagnose.

The four dimensions you must track separately

Engine. ChatGPT, Gemini, Perplexity, Claude, Copilot, and AI Overviews behave differently enough that averaging them obscures everything. We typically see 30-50 point spreads for the same brand on the same prompt set across engines.

Topic cluster. A B2B software brand might win in "team collaboration tools" and lose in "project management software." Reporting one company-wide number hides both.

Prompt intent. "What is X" prompts behave differently from "best X for Y" prompts which behave differently from "X vs Y" prompts. The optimization tactics for each are distinct.

Position. Being mentioned first, with a positive descriptor, in the introductory sentence is not the same as being mentioned fifth as a footnote. Treat them as different events.

The minimum viable dashboard

For each tracked topic cluster, you want a weekly grid that looks like this:

• Rows: engines
• Columns: prompt intents (informational, comparative, transactional)
• Cells: share of voice this week, change versus last week, position-weighted

Twelve numbers per cluster, refreshed weekly. That is the minimum. Anything less and you cannot diagnose. Anything more and your team will not look at it.

The math that matters

Position-weighted share of voice is the metric we recommend optimizing. The formula we use:

Weighted SoV = sum over mentions of (1 / position) divided by sum over all citations of (1 / position), expressed as a percentage.

This rewards being cited early and gives diminishing credit to being cited late. It correlates with actual brand consideration in customer surveys at r=0.71 in our panel, compared to r=0.43 for raw mention count.

The trap to avoid

Do not optimize the average. Optimize the weakest cell that matters. If you are at 12% in Perplexity comparative prompts in your most strategic cluster, that cell will move your business more than going from 58% to 64% in ChatGPT informational prompts where you already dominate.

The right metric, segmented the right way, makes the right action obvious.