Measurement basics
How Peekaboo measures AI visibility
What a prompt run records, how brand mentions and citations are detected, and why AI answers must be treated as observations rather than fixed rankings.
7 minute read · Last reviewed: July 20, 2026
The measurement unit: a prompt run
Peekaboo starts with a defined prompt and runs it on a named AI model or search surface. The resulting prompt run is the smallest unit of evidence in the product. It connects the prompt and collection time to the returned answer and its analysis.
A run can record:
- the prompt, model or surface, collection time, language, and market context;
- the answer available to the collection process;
- explicitly detected brand or entity appearances and their position;
- the per-run visibility score for each tracked entity; and
- source URLs and domains when the returned answer exposes citations.
The run count, date window, prompt set, and filters are part of every aggregate. Without that context, a score is incomplete.
From answer collection to analysis
- Define the test. Select the prompt, model or surface, language, location, and collection time.
- Collect the observable answer. AI providers can expose different answer and citation formats. Peekaboo preserves the context available to its collection process.
- Identify explicit entities. The analysis is designed to retain brands that actually appear in the answer, resolve known entity variants, and avoid inventing a mention that is not present.
- Score the run. Mentioned entities receive a position-based score. A tracked entity that is not mentioned receives zero for that run.
- Aggregate comparable runs. Eligible runs are averaged for the selected window and filters, including non-mentions.
See the complete Visibility Score explanation for the formula and worked example.
Generated and synthetic prompts
A tracked prompt is a controlled question used to inspect an AI answer. Prompts may be supplied by your team, adapted from customer language, derived from Search Console evidence, or generated as suggestions. These origins are not equivalent.
A strong prompt set documents why each prompt matters, covers multiple stages of the decision journey, and stays stable long enough to compare windows. When the set changes, annotate the change because the aggregate now represents a different test.
Why repeated observations matter
AI output is probabilistic. The same prompt can produce a different answer on a later run even when your site has not changed. Models, indexes, interfaces, citations, and personalization can also change outside your control.
Treat one run as diagnostic evidence. Use repeated, comparable runs to assess a pattern, and inspect the underlying answers before concluding that a score movement reflects a durable market change.
- Keep the prompt cohort and filters consistent across the comparison.
- Compare similar-length windows and examine the run count.
- Break aggregates down by prompt and model to find the source of movement.
- Record content, product, PR, and measurement changes on the same timeline.
Four signals that should not be collapsed
| Signal | What it establishes | What it does not establish |
|---|---|---|
| Mention | The answer referred to the entity. | That it was endorsed or clicked. |
| Recommendation | The answer presented the entity as an option. | That a user chose it. |
| Citation | The answer exposed a link to a source. | That the source caused the claim or received a visit. |
| Business outcome | Your first-party system recorded an event. | That AI visibility caused it without further design. |
Next step
Read Turn AI visibility measurements into decisions before using a trend to change content, and consult Peekaboo’s public measurement methodology for the concise standard.
Frequently asked questions
Is an AI answer a fixed ranking like a traditional search result?
No. Generated answers can change with the model, prompt wording, market, language, time, account context, and normal output variation. Peekaboo records an observation under known conditions; it does not claim that one answer represents every user experience.
Does a tracked prompt prove that real customers use that exact wording?
No. A tracked prompt may come from customer research, Search Console, a team member, or a generated suggestion. Unless it is tied to observed demand data, treat it as a synthetic test query used to measure a scenario—not as an estimate of search volume or audience demand.
Does a mention mean the AI recommended the brand?
Not necessarily. A mention records that the answer referred to the tracked entity. Recommendation language, sentiment, ranking, and citation context are separate signals and should be reviewed independently.
Can Peekaboo prove that an AI answer produced a sale?
No. A prompt run measures the answer, not whether a person saw it or acted on it. Use session-level analytics and qualified business outcomes to evaluate downstream impact, and use controlled tests when a causal conclusion matters.