Method
How the audit works, and what it can't tell you
Said by AI answers one question: when someone asks an AI model to recommend options in your category, does the answer name you? Here is exactly how that is measured.
1. 20 questions written from your category
You give a category in plain words, such as “project management software” or “dentist in Austin”. The audit turns it into 20 questions phrased the way buyers ask: best, cheapest, most trusted, for startups, alternatives, comparisons. A few examples:
- What are the best {category} companies?
- Best {category} software for small business
- Compare popular {category} platforms
- Which {category} service is most trusted?
- Popular {category} alternatives
- Most recommended {category} on Reddit
The free check uses the first 5 of these questions.
2. Each question asked in three answer styles
Every question goes to OpenAI's GPT-4o mini through the API three times, each with a different instruction:
| Style | What it asks for |
|---|---|
| Direct answer | A conversational recommendation, the way a chat assistant answers from its own knowledge. |
| Ranked list | A numbered list of the best options, established and emerging. |
| Research-style list | A broader list with short descriptions, the format answer engines tend to use. |
That gives 60 answers for the full audit (15 for the free check).
3. Every answer read for mentions and competitors
A second model pass reads each answer and records whether your brand is named (including close variations), its position if the answer is a ranked list, the sentence around the mention, and up to five other brands recommended. Your visibility score is the share of answers that named you.
4. The report
The full audit is emailed as an HTML report, usually within a few minutes of purchase. It includes the score, results per answer style, up to 10 competitors with counts, quotes of your mentions and a prioritised action plan. See the sample report.
What it can't tell you
- It measures what the model says, not live ChatGPT, Gemini, Claude or Perplexity sessions with web search switched on. Those can cite fresh pages the model has not seen.
- Answers vary between runs. Treat the score as a range, and compare runs made the same way.
- It cannot see personalised answers shaped by a user's memory or location.
- A mention is not a recommendation you would be happy with. Read the quotes, not just the count.
If you need tracking across live AI platforms every week, a subscription tool fits better. If you need a baseline and a plan, start with the free check.