Direct questions · No. 08 / 31
8. How do we measure share of voice in ChatGPT without overstating it?
Last updated
Define the denominator before calculating the percentage, and describe the result as a share of your disclosed test set. Do not call it US market share or the share of all ChatGPT searches.
We recommend publishing the following measures separately:
| Metric | Proposed calculation |
|---|---|
| Mention frequency | Valid observations naming your brand ÷ valid observations |
| Recommendation frequency | Valid observations recommending your brand ÷ valid observations |
| Owned-site citation frequency | Valid observations linking to your domain ÷ valid observations |
| Prompt coverage | Successfully tested prompts with at least one recommendation ÷ successfully tested prompts |
| Competitive recommendation share | Your brand–answer recommendation occurrences ÷ equivalent occurrences for all brands in a fixed comparison set |
Count a brand once per answer for the last measure, even if its name appears repeatedly. Disclose the comparison set and how “other” brands are handled; an empty recommendation denominator produces an undefined share, not zero market presence.
Illustrative example: a brand recommended in 12 of 30 valid answers has a 40% observed recommendation frequency. If those recommendations occur across four of ten prompts, its prompt coverage is also 40%—but the two figures describe different things.
Keep the prompt mix stable. If collection failures give some prompts more observations than others, also show equal-weighted prompt-level rates rather than allowing collection success to determine the weighting.
For AI Overviews, publish both overall exposure across successful searches and inclusion conditional on an Overview appearing. A successful search with no Overview is not a collection failure.