Measurement guide
How to measure AI visibility
A repeatable study fixes the prompt panel and collection rules before reading the result. Always show the numerator, eligible denominator, engines, period and failure rate.
Quick answer
Quick answer
Define the exact measurement question, version a representative prompt panel, fix the engine and locale conditions, capture each raw answer, calculate transparent mention and citation rates, repeat a sentinel subset to expose variability, and report limitations. Never treat failed calls as absent brands or silently combine API and consumer-interface results.
1. State the measurement question
Name the entity, audience, market, languages, engines, observation period and decision the study should support. This scope determines which prompts and responses are eligible and stops the metric from expanding after the results are visible.
2. Build and version the prompt panel
Cover distinct intents such as category discovery, problem-to-solution, features or workflows, comparisons, local fit, education, branded truth checks and negative controls. Allocate counts before collection and keep discovery prompts free of the measured brand name.
Assign each prompt a stable key, locale, intent, research-mode flag and panel version. Corrections create a new version; they do not rewrite historic observations.
3. Fix the engine, locale and run conditions
Use fresh sessions where possible and record provider, model identifier, retrieval mode, location or locale, timestamp and API request ID. Official grounded APIs are suitable for repeatable collection, but they do not reproduce every consumer interface. Report those modes separately.
4. Capture each answer as an observation
Store the raw prompt, response, citation metadata, completion status and error class before deriving a score. Resolve brand aliases against an approved entity list and preserve uncertain matches for review. Do not replace a failed response with demonstration data.
5. Calculate mention and citation metrics
Mention rate equals valid responses containing the resolved entity divided by valid responses. Citation rate equals valid responses containing at least one qualifying citation divided by valid responses. Owned-domain citation rate uses the same denominator and requires an approved domain list.
If ordered recommendation position is measured, publish the extraction rule. Mean reciprocal rank can summarize valid ranked appearances, while share of voice compares resolved brand mentions across a defined competitor set. Neither should hide the underlying counts.
6. Repeat runs to expose variability
Run the full panel on a stable weekly schedule and repeat a smaller sentinel subset during the week. Use a rolling window for directional reporting and require adequate coverage. Variance is a property to show, not noise to remove after seeing an inconvenient answer.
7. Report findings with denominators and limits
Every chart should state the prompt version, engines, dates, eligible observations and failure rate. Separate research API results from manual consumer-interface checks. Describe a change as an observation; reserve causal claims for a design capable of supporting them.
Primary sources and further reading
- Overview of OpenAI crawlers — OpenAI
- Perplexity crawlers — Perplexity
- Creating helpful, reliable, people-first content — Google Search Central