The AI visibility audit method
A step-by-step, reproducible process for measuring how a brand appears in AI-generated answers. Every step is documented so the audit can be rerun and independently checked.
The core principle
An audit is only as good as its reproducibility. Before you run anything, fix the questions, the capture settings, and the scoring rules in writing. Then run the capture, keep the raw evidence, and apply the rules exactly as written. If a result surprises you, change the rules deliberately and note the change — never quietly.
The six steps
-
Scope and objective
Define the single question the audit answers. Common objectives: (a) does my brand appear at all, (b) does it appear accurately, (c) how does it compare to peers, (d) how consistent is it over time. Write the objective, the category, and the geography (here: US, en-US).
-
Build the prompt set
Create a fixed list of prompts: brand-named, category-named, and problem-framed questions. Define inclusion/exclusion criteria for what counts as a "mention" (e.g., brand name spelled correctly in the answer, in any position, within the response body). Lock the list before capture.
-
Fix the capture settings
Record the assistant or model, its version/settings, temperature if exposed, date, and capture method. Use the same settings for every prompt in a run so observations are comparable. Note any UI or model updates between runs.
-
Capture verbatim answers
Run each prompt and store the full answer text verbatim with its snapshot metadata. Save the raw output exactly as returned; do not edit. This is the evidence that supports every score.
-
Score with an explicit rubric
Apply a fixed scoring scale for each answer: mention present/absent, accuracy (does the stated fact about the brand match the brand's own sources), stance (positive, neutral, negative, or n/a), and consistency across repeated captures. Record a score and a short justification for every prompt.
-
Report with evidence
Summarize counts and patterns, then link every claim back to a captured answer. Publish the prompt set, settings, raw responses, and scores so another person can reproduce the audit. Be explicit about what is a measurement versus an illustrative fixture.
Scoring rubric
A simple, explicit rubric keeps scoring consistent and auditable. The three dimensions below are a starting point; define your own weights before capture and state them.
| Dimension | What it measures | Example scale |
|---|---|---|
| Mention | Whether the brand name appears in the answer at all, per the inclusion criteria. | 0 = absent, 1 = present once, 2 = present in a meaningful position |
| Accuracy | Whether stated facts about the brand match the brand's own verifiable sources. | 0 = inaccurate/unsupported, 1 = partially accurate, 2 = accurate |
| Stance | The overall framing of the brand within the answer. | negative, neutral, positive, or n/a (no mention) |
| Consistency | Stability of the answer across repeated captures with the same settings. | low / medium / high |
Know the limits
AI assistants can change between runs, and answers reflect the provider's data and policy, not a ranking system. A single snapshot is a measurement, not a guarantee of future behaviour. Do not treat any one observation as a permanent fact about your brand.
From method to example
The method above is applied end-to-end in the worked example, with illustrative fixtures clearly marked. Follow it to see each step in practice.
