By KYN AI Advisory Team — AI implementation specialists, Singapore
To measure AI-search visibility, write about 50 prompts that real buyers would ask, group them by intent, ask each one of an AI assistant, and log which brands it names and which sites it cites. Repeat on a schedule and compare. Fifty is enough to see a pattern and small enough to review by hand. Our AI search check costs about S$0.05 per check, and we have run 22 checks.
What is an AI-search visibility check?
It answers one question: when a buyer asks an AI assistant for help in your category, does your company appear, and which pages does the assistant rely on?
Each run produces two lists per prompt:
- Mentions. Every company or product named in the answer, in order.
- Citations. Every URL the answer links or lists as a source.
Counted over 50 prompts, those lists show your share of mentions, your competitors' share, and the sites that feed the answers. The cited sites tell you where to publish, correct or earn a mention. For the wider topic, see how to appear in AI search answers.
How do you write 50 good prompts?
Start from buyer questions, not keywords. A keyword is a fragment. A prompt is a full request, the way a person types it to an assistant.
Use six intent groups. Aim for roughly the counts below, then adjust to your market.
| Intent group | What the buyer wants | Example shape | Count | |---|---|---|---| | Category discovery | Who offers this | "Which companies build custom [system] for SMEs in Singapore?" | 10 | | Comparison | A choice between options | "Custom CRM or [named tool] for a 20-person firm?" | 10 | | Cost | A price range | "How much does [service] cost for a small business?" | 8 | | How-to | A method | "How do I automate [task] in [country]?" | 10 | | Risk and trust | A safety answer | "Is it safe to let an AI agent do [task]?" | 6 | | Local and regulatory | A country rule | "What does [country] require for [practice]?" | 6 |
Then add structure to each prompt:
- Country. Run variants for Singapore, Hong Kong, Indonesia and Malaysia if you sell there. Assistants answer differently by place.
- Company size. "For a 10-person company" changes the answer.
- Language. If buyers ask in Bahasa Indonesia or Chinese, include those prompts.
- Role. A finance lead and an operations manager ask the same need in different words.
Where do the prompts come from?
Five sources, in order of value.
- Sales calls and enquiries. The exact words buyers use. This is the strongest source and the cheapest.
- Search Console queries. Queries you already show for, rewritten as full questions.
- Competitor pages. The questions their FAQs and comparison pages answer.
- People-also-ask style questions from search results for your core terms.
- A model. Ask an assistant to propose 100 prompts per intent group from a description of your buyer. Then a person cuts the list to 50, removing duplicates and any prompt no buyer would type.
A model can widen the list. A person must choose, because a model tends to produce tidy, repetitive questions that real buyers do not ask.
How do you run the check?
- Fix the wording. Freeze the 50 prompts as version 1. Changing wording between runs breaks the comparison.
- Ask each prompt in a clean session. No history, no personalisation, and the same model and mode each time. Record which assistant and which version.
- Store the full answer. Save the text and the citations with the date.
- Extract mentions and citations. A script, or a model with a strict output format, lists the brands named and the URLs cited. A person spot-checks ten answers.
- Normalise names. "KYN", "KYN Technology" and "kyn.com.sg" count as one.
- Repeat. Answers vary between runs. Run each prompt more than once and report the share of runs in which a brand appears, not a yes or no.
What do you report?
Keep it to numbers a team can act on.
- Mention rate. The share of runs in which your brand appears, overall and by intent group.
- Competitor rate. The same for your three nearest competitors.
- Top cited sites. The 20 domains cited most often, with your own pages marked.
- Gaps. Prompts where competitors appear and you do not.
- Change. The difference from the last run.
The action list comes from the gaps. If an assistant cites a comparison site that omits you, that is a listing to earn. If it cites a competitor's guide for a how-to question, that is an article to write. If it cites your page but names a competitor, the page needs a clearer statement of what you do.
What does it cost?
On our system, an AI search check costs about S$0.05 per check, and we have run 22 checks. A full 50-prompt sweep multiplies by the number of prompts and repeats, so budget for it as a small monthly line, not a project. The larger cost is the person time to review the gaps and write the fixes.
Be careful with what the numbers mean. Assistants change with every model update, and answers vary between runs. A change of a few points between two runs is noise. A change that holds over several runs, or a gap that is the same in every run, is a signal.
What are the trade-offs?
- Sample bias. Fifty prompts represent your market only as well as you chose them. Refresh the set each quarter from new enquiries.
- Moving target. Assistants update. Keep a log of the assistant version with each run.
- Causation. A rise after you publish an article does not prove the article caused it. Treat the data as a guide to where to look.
Where KYN fits
KYN runs this check as part of its marketing work and builds it into clients' systems in their own cloud account. You can run a first look yourself on our AI search check page. Agents have their own restricted accounts and every action is logged with the actor. The first 14 days are a review of your stack and interviews with your team; a full build ships in under 90 days. Pricing is a one-off build fee and a retainer of at least twelve months, quoted after the review.
Frequently asked questions
Why 50 prompts and not 500?
Fifty covers six intent groups with enough repeats to see a pattern, and a person can review the results by hand. More prompts add cost and noise before the first set is understood.
How often should I run the check?
Monthly is enough for most businesses. Run it again within a week after a major content or site change, and run each prompt more than once because answers vary.
Does being cited mean being recommended?
No. A citation means the assistant used your page as a source. A mention means it named you. Track both, and read the answer text to see whether the framing helps you.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
