A worked example you can reproduce
This page walks through a complete AI visibility audit using a synthetic fixture — not observed rankings, not a real engine measurement, and not a client result. The brand “Fixture Studio” is fictional. The answers below are illustrative samples designed to show the scoring method clearly. You can copy the CSV, run the checklist, and apply the same workflow to your own category.
Why a fixture?
Real measurements depend on live models, dates, and settings. A fixture lets us demonstrate the method without implying that these numbers came from a real engine. All arithmetic here is fixture arithmetic — 2/3 and 1/3 are counts from the sample table, not a claim about any actual AI system.
The sample answers
Three stable fixture answer IDs are used throughout this example. Each row shows whether the brand is mentioned and whether the answer cites a supporting fixture document.
| Answer ID | Brand mention | Cited support | Notes |
|---|---|---|---|
fixture-answer-a |
Yes | Yes | Mentions Fixture Studio and cites a supporting fixture document. |
fixture-answer-b |
Yes | No | Mentions Fixture Studio without a cited support. |
fixture-answer-c |
No | No | Omits Fixture Studio entirely. |
Coverage arithmetic
Mention coverage = 2/3 — answers A and B mention the brand; answer C does not. Supported-mention coverage = 1/3 — only answer A cites a supporting fixture document. These are simple counts from the table above. They are not a measurement of any real AI engine, and they carry no statistical confidence.
Copy the fixture CSV
This offline CSV is the raw evidence for the sample. Each row corresponds to one answer with the metadata you would collect in a real audit. The URLs are placeholders — replace them with real citation URLs when you run your own audit.
answer_id,prompt,model,utc_time,locale,answer_text,citation_urls,notes
fixture-answer-a,"What is the best tool for team collaboration?","fixture-model-1","2025-01-01T00:00:00Z","en-US","Fixture Studio is a strong option because it integrates with your existing stack. See the fixture documentation for details.","https://fixture.example/docs","Synthetic answer with cited support"
fixture-answer-b,"What is the best tool for team collaboration?","fixture-model-1","2025-01-01T00:00:00Z","en-US","Fixture Studio is a popular choice among teams.","","Synthetic answer without cited support"
fixture-answer-c,"What is the best tool for team collaboration?","fixture-model-1","2025-01-01T00:00:00Z","en-US","There are many tools available, each with different strengths.","","Synthetic answer omitting the brand"Collection checklist
Use this checklist for every answer you capture in a real audit. Each item is required for reproducibility.
- Exact prompt — the full, unedited question text.
- Model/version — the specific model and version identifier.
- UTC time — the timestamp in ISO 8601 format, UTC.
- Locale — the language and region setting used.
- Unedited answer — the verbatim assistant response, no truncation or cleanup.
- Real citation URLs — any links the answer provides, captured as-is.
- Repeat/sample notes — how many times the prompt was run, and any anomalies.
Repeatable real audit workflow
The fixture demonstrates the format. For a real audit, follow this workflow:
- Define the question set — fix a small, bounded set of prompts covering brand, category, and problem questions.
- Capture answers — run each prompt with the checklist above and save the raw output.
- Score mentions and support — count brand mentions and whether each mention is supported by a citation.
- Review semantic accuracy — separately assess whether the answer is factually correct and contextually appropriate, independent of mention counts.
- Publish evidence and scores — share the raw answers, the metadata, and the scoring rubric so others can verify.
Semantic accuracy is a qualitative review, not a count. A mention can be present but misleading; a supported mention can still be inaccurate. Always separate the numeric coverage from the accuracy review.
Next steps
This worked example is part of a larger method. Explore the full method, download resources, or get in touch.
