In practice
The value of the metric is comparative. An absolute count of rankings or citations tells you little, because the denominator keeps moving as query sets change. Share against a fixed set of competitors and a fixed set of queries is what makes movement interpretable.
In generative measurement the panel has to be fixed before work starts and changed rarely. A panel adjusted mid-engagement produces improvements indistinguishable from a redefined test, which is how a great deal of AI visibility reporting becomes unfalsifiable without anyone intending it.
Rates also need larger samples than positions do. A single prompt returning a different answer between runs is normal variance, not a change in standing, which is why panels run to hundreds of prompts and why weekly movement should not be over-interpreted.
Not to be confused with
- Visibility score
- Vendor visibility scores are proprietary weightings of ranking data. Share of voice is a plain proportion of a defined set, which is easier to audit.
Questions
How large should a prompt panel be?
Typically one to two hundred commercial prompts. Below that, normal run-to-run variance swamps the signal you are trying to trend, and weekly movement becomes noise rather than information.
Can the panel be changed?
Rarely, and never mid-engagement. Quarterly review for genuine category shifts is reasonable. Adjusting it because results disappoint destroys the only thing that made the measurement credible.
Is share of voice better than rank tracking?
It answers a different question. Rank tracking gives position; share of voice gives proportion of a set. For generative engines, where no positional index exists, share is the only workable form.