The short version: an AI Visibility Audit is not a vibe or a single score. It is a fixed set of real buyer questions, run in ChatGPT, Perplexity, Gemini and Claude, with every mention recorded, from which one honest number falls out: your share of answer. This page walks through a worked example so you can see the deliverable before you ever request one.
What prompts go into the audit?
Prompts are written the way buyers actually ask, in five families: best/top (“best dental clinic in [your city]”), near-me and location variants, comparisons (“X vs Y for implants”), price-led (“affordable…”), and problem-led (“my leads never reply, what system fixes that?”). A useful set is 50–100+ prompts. Fewer and the number is anecdote; more and you are measuring your own thoroughness rather than the market.
How is share of answer calculated?
| Measure | Definition in this audit |
|---|---|
| Mentioned | Brand named anywhere in the answer (yes/no, per prompt, per engine). |
| Share of answer | Mentions divided by total prompt-runs. 100 prompts × 4 engines = 400 runs; 30 mentions = 7.5%. |
| Favourable | Named with a positive or neutral recommendation, not as a cautionary example. |
| Cited | A source on your own domain appears in the engine's references, where references are shown. |
A worked example (illustrative composite)
The table below is a composite built from typical first-audit patterns for a mid-size local service brand. It is labelled illustrative because publishing a real client's prompt data without permission would end our client relationships, correctly. The structure, maths and fix categories are exactly what clients receive.
| Engine | Runs | Mentioned | Favourable | Share of answer |
|---|---|---|---|---|
| ChatGPT | 100 | 6 | 4 | 6% |
| Perplexity | 100 | 11 | 8 | 11% |
| Gemini | 100 | 9 | 7 | 9% |
| Claude | 100 | 4 | 3 | 4% |
| Total | 400 | 30 | 22 | 7.5% |
Meanwhile the leading competitor sits at 41%, named in 164 of 400 runs, usually first, usually with a reason attached. That gap, not the 7.5% alone, is the finding: the category has already been decided in AI answers; it simply wasn't decided by you.
What fix list falls out of numbers like these?
- Entity first. In the composite, the brand's name, address and service list differed across its site, Google profile and two directories, so engines hedged. One canonical description, aligned everywhere, is week one.
- Two answer-ready pages. The service pages opened with history and adjectives. Rewritten answer-first (what it is, for whom, where, what it costs, proof), they become quotable.
- Reviews with recency. 19 reviews, newest 14 months old, read to an engine like a closed business. A steady, honest review rhythm is a citation signal.
- Third-party presence. The competitor appeared in three “best of” round-ups the engines clearly trusted. Earning a place in sources AI already cites beats publishing another blog post.
- Re-run weekly. Same 100 prompts, same sheet. Share of answer either climbs or the work changes, no quarterly surprises.
Key takeaways
- Demand the prompt list with any audit. A score without its questions is marketing.
- Share of answer is a ratio you can recompute yourself, insist on being able to.
- The first fixes are usually unglamorous: consistency, quotable pages, fresh proof, borrowed authority.