By KYN AI Advisory Team — AI implementation specialists, Singapore
To check whether AI assistants cite your site, write 20 questions your buyers ask, put each to ChatGPT, Perplexity and Gemini at least three times, and log whether your brand is named and whether your URL is cited. Count the results as a rate per assistant. One answer proves nothing, because responses vary from run to run. A fixed question set repeated monthly shows whether your work moves the number. A spreadsheet is enough to start.
Why not just ask once and look?
Assistants generate each answer fresh. Ask the same question twice and the sources can differ. Results also depend on location, account settings and whether the assistant used live search.
So treat it as sampling. A single "yes, they cited us" is an anecdote. "We were cited in 4 of 9 runs for this question, across three assistants" is a measurement.
Step 1: Which questions do you test?
Build a list of 20. Mix four types:
| Type | Example for a Singapore accounting software firm | |---|---| | Category | "What software handles GST for a small Singapore company?" | | Comparison | "Xero vs a custom finance system for a Singapore SME" | | Problem | "How do I reconcile bank statements faster?" | | Brand | "What does [your company] do?" |
Write each as a full sentence, the way a buyer would. Take wording from Search Console queries, sales call notes and WhatsApp enquiries. Fix the list. Do not change it month to month, or you cannot compare.
Step 2: How do you run the tests?
For each question and assistant:
- Open a fresh session. Log out of personal memory features where you can, or use a clean browser profile.
- Turn on web search or browsing if the assistant has the option, since citations usually come from live retrieval.
- Paste the question. Do not add context.
- Save the full answer and the list of cited links.
- Repeat three times, ideally on different days.
For 20 questions, three assistants and three runs, that is 180 answers. A person takes a few hours. An agent can do the same in minutes and write the log, which is the better use of the time. Check each service's terms before automating its consumer interface, and prefer its official API where one exists.
Step 3: What do you record?
One row per answer. Columns:
- date, assistant, question ID, run number;
- brand named in the text (yes/no);
- your URL cited (yes/no, and which URL);
- position of the mention (first, middle, last, none);
- competitors named;
- other domains cited;
- accuracy of what was said about you (correct, partly wrong, wrong);
- note.
The accuracy column is the one most people skip. An assistant that names you with the wrong price or wrong country harms you more than silence.
Step 4: How do you read the results?
Compute three numbers per assistant.
- Mention rate. Answers that name your brand, divided by total answers.
- Citation rate. Answers that link to your site, divided by total answers.
- Share of voice. Your mentions as a share of all brand mentions in the category.
Then read the patterns.
- Mentioned but not linked. The assistant knows you from somewhere, perhaps another site. Your own page may not be answering the question directly.
- Competitor linked, you absent. Open their cited page. What does it have that yours lacks: a direct answer, a table, a sourced number?
- Wrong facts. Find the page the assistant drew from. Fix the source or publish a clearer page.
- Not cited anywhere. Check access first, before content.
Step 5: What do you check if you are never cited?
Work through access before content.
- robots.txt. Does it block the AI search bots? OpenAI's documentation says OAI-SearchBot surfaces sites in ChatGPT search and recommends allowing it. Perplexity's documentation says PerplexityBot helps sites appear in its results and honours robots.txt.
- Firewall and CDN rules. Some security settings block bots by default. Both vendors publish IP ranges you can allow.
- Rendering. Does the content appear in the HTML, or only after scripts run?
- Indexing. For Gemini and Google's AI features, the page must be indexed and eligible to appear in Search with a snippet, according to Google's guidance. Use Search Console to check.
- The page itself. Does it open with the direct answer? See how to write an answer-first article.
Vendors state that changes to robots.txt can take around 24 hours to apply for search, so wait before you re-test.
How often should you re-run the check?
Monthly is enough for most SMEs. After you publish or change a page, wait two to four weeks before expecting a change. Weekly checks mostly record noise.
Keep a one-page summary: three rates per assistant, the three biggest gaps, the three actions taken. Review it in your marketing meeting.
What does the check not tell you?
- It does not tell you traffic. Citations may produce few clicks. Pair it with referral data from your analytics, which often shows visits from assistants.
- It does not tell you what every buyer sees. Personalised results differ.
- It does not prove cause. If the rate rises after you publish, that is a signal, not proof.
For context on the method as a whole, read what is generative engine optimisation, and see our earlier piece on getting a business website cited by ChatGPT and Google AI Overviews.
What does a filled log look like?
An illustrative extract for one question, "What software handles GST for a small Singapore company?", run three times in one assistant. These rows are invented to show the format.
| Run | Brand named | URL cited | Position | Competitors named | Accuracy | |---|---|---|---|---|---| | 1 | Yes | No | Middle | 3 | Correct | | 2 | No | No | None | 4 | Not applicable | | 3 | Yes | Yes | First | 2 | Partly wrong |
Mention rate for this question in this assistant: 2 of 3. Citation rate: 1 of 3. Run 3 needs action because the assistant named you with one wrong detail.
Three habits make the log useful:
- Save the raw answer text in a second sheet, so you can re-read it when the numbers look odd.
- Add a column for the date of your last page change. It lets you see whether a change preceded a shift.
- Mark any run where the assistant did not use live search. Those answers reflect older training data and should not be mixed with retrieval answers.
Agree one owner for the log. Without an owner, the second month's check does not happen.
Where KYN fits
KYN runs this check as an agent. It asks a fixed question set across assistants, stores every answer and citation, and reports mention rate, citation rate and share of voice, with the full log behind each figure. Our AI search check costs about S$0.05 per check, across 22 checks. It runs in your own cloud account. You can also try the AI search check directly.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. See marketing systems.
Frequently asked questions
Can I check this for free?
Yes. The free tiers of the assistants are enough for a baseline: 20 questions, three runs, a spreadsheet. The cost is time, not money.
Why does the assistant name my competitor but not me?
Often because a competitor's page answers the question directly and is easy to fetch. Open the cited page and compare it with yours. Check access first, then structure, then facts.
Should I pay for a tool that tracks AI citations?
Pay once the manual check proves the method and the volume makes it slow. Ask any tool for its question set, number of runs per question and a raw log you can export.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
