The short answer: you can start in five minutes โ open an AI assistant, ask a question your buyers actually ask, and see whether you appear. Between "asking the right question" and "turning one curious check into a trackable metric" sit four common traps. This guide covers both.
The 5-minute manual test
- Don't ask "what do you think of brand X" โ ask your buyer's question. "Which X vendors are worth shortlisting", "what tools do people use for X" โ those are the questions that move deals. Ask about your own brand by name and the engine will usually oblige with a polite profile; that measures courtesy, not visibility.
- Ask in both languages that matter to you. Same engine, same question, different language โ often a different set of brands in the answer. Test at least the side your market lives on.
- Test several engines, from both sides. Pick two or three your overseas buyers use and two or three your Chinese-market buyers use (our platform currently has 8 engines live: ChatGPT, Kimi, DeepSeek, Gemini, Doubao, Qwen, Perplexity, Grok; pick any of them for a manual test). Single-engine conclusions are not representative.
- Record the raw answer and the date. Screenshot or save the text, tag it with the engine and date. Answers drift โ without records, next month you will be reduced to "it feels better, I think".
Those four steps give you a small grid: engines by questions, each cell "named / not named". That is your first AI-visibility snapshot.
The four traps of a one-off manual test
- Jitter. Ask the same question tomorrow and the shortlist may change. One absence does not mean you are out; one appearance does not mean you are safe. Only repeated measurement supports a conclusion.
- The courtesy mirage of self-referential prompts. As above: ask about yourself by name and you will usually get a decent write-up. The real signal lives in category questions that never mention you.
- Memory and retrieval, blended. Some answers come from training data (possibly a year-old snapshot of the internet), some from live retrieval. The former you cannot edit โ only wait out; the latter you can influence with content today. Mixing them up leads to the wrong prescription โ see tracking AI referral traffic.
- No denominator. "ChatGPT mentioned us" is not a metric; "named in 3 of 20 buyer questions" is. Good news without a denominator is exactly as untrustworthy as bad news without one.
When to graduate from manual tests to continuous monitoring
A manual test answers "am I on the list today". It cannot answer:
- Did last month's content earn us more mentions?
- Is our competitor's mention rate rising or falling?
- An engine suddenly stopped naming us โ did the engine change, or did we?
Answering those takes a fixed question set, daily sampling, multiple engines, versioned scoring rules (so trends stay comparable when the methodology evolves), and honest flags when samples run thin. That is what continuous monitoring tools are for โ our own approach is published in full in how we score and on the methodology page, and you could absolutely build one yourself from it.
If you want a snapshot of your brand right now, the free checkup produces a cross-engine snapshot in a few minutes โ per-engine hit tiers and the methodology notes are in the report.
One honest closing note: we ran this exact method on ourselves and the first measurement came back zero โ published as-is in scoreboard episode 0. The method is not afraid of ugly numbers. It is afraid of not measuring.