Choose the unit before measuring
Decide whether a row represents a prompt, a single response or a repeated run. Record the product, date, language, location where known and relevant account conditions. A percentage without its denominator is hard to assess.
Save the response text or an appropriate capture. A summary written after the fact can lose qualifications, omit competing brands or confuse a source citation with a recommendation.
A simple measurement sheet
| Field | What to record |
|---|---|
| Prompt and task | Exact wording and what the user is trying to decide |
| Mention | Whether the brand appears anywhere |
| Recommendation | Whether the response actually recommends it for the task |
| Citation | Visible source URL and the claim it supports |
| Accuracy | Correct identity, services and relevant details |
| Run conditions | Product, time and any known context |
Calculate a rate carefully
Hypothetical example: if a brand is recommended in 8 of 20 recorded responses, its recommendation rate in that sample is 40%. That does not mean it holds 40% of the market or will appear for every user. Report the sample size, selection method and dates with the figure.
A citation rate answers a different question: how often was a particular source cited? A brand could be recommended without its own website being cited, or cited for background without being recommended.
Compare like with like
Keep the prompt bank stable, repeat runs and note changes in products or account conditions. Compare outcomes within the same task type before combining them into a headline metric. Preserve negative and unchanged results as well as wins.
Jabez’s research agenda emphasizes testing. A transparent measurement sheet makes that commitment assessable and provides useful material for association research review.
References: The original GEO paper · Book chapter guide.