Write down 15–20 real questions a prospective client would ask an AI assistant. Run each one in ChatGPT, Gemini, and Perplexity, in both English and Chinese if relevant to your market. Log whether you're mentioned, where you rank against named competitors, and whether any stated fact about you is wrong. Repeat monthly — the answers change without notice.
Step 1 — write the questions your prospects actually ask
Avoid generic prompts like "best clinic Hong Kong." Real buyer questions are more specific and closer to how people actually type:
- "Best aesthetic clinic in [district] for [treatment]?"
- "Is [your practice name] trustworthy? What do reviews say?"
- "[Your practice] vs [named competitor] — which is better for [service]?"
- "How much does [service] cost in Hong Kong?"
- "Do I need a lawyer for [scenario]?" (for legal practices)
Fifteen to twenty questions, covering discovery ("who's best"), trust ("tell me about X"), price, and comparison, is enough for a first pass.
Step 2 — run each question across assistants and languages
Test the same question in ChatGPT, Gemini, and Perplexity — the three most commonly used by consumers as of 2026 — and, if your clients search in more than one language, run the same set again in Chinese. It is common to find a practice performs very differently by language; this gap is frequently the single most valuable finding.
Step 3 — log four things per answer
| Field | What to record |
|---|---|
| Mentioned? | Yes / no |
| Position | Named first, named among others, or not named |
| Sentiment | Positive, neutral, or cautionary |
| Accuracy | Any fact stated about you that is wrong or outdated |
A simple spreadsheet is sufficient for a first pass. The pattern that matters is not any single answer, but the aggregate: what fraction of your 15–20 questions name you, and who is named instead.
Step 4 — repeat monthly
Model providers update retrieval and generation behaviour without changelog. A result from today is a snapshot, not a permanent state. Running the same question set on a fixed cadence — monthly is a reasonable default — is what turns a one-off curiosity into a usable trend line, and is what any credible AI-visibility engagement should be reporting back to you.
What this manual method won't tell you
It surfaces the pattern; it does not tell you why a competitor is named instead of you, or which specific web sources the model is drawing from. That requires examining the citations an assistant provides (where available) and cross-referencing which directories, reviews, and articles a competitor appears in that you do not — the "citation neighbourhood" that actually drives the answer. That deeper analysis is the basis of a full audit.