Run two different tests
“What does Atsell do?” checks how a tool describes a named company. “Which agencies can manage Shopee and TikTok Shop in Singapore?” checks whether the company appears during discovery.
Both questions are useful. Keep the results separate. A tool that describes your business accurately after you name it has not necessarily recommended you to someone who has never heard of you.
“What does [your company] do?”
“Who can help [buyer] with [need] in [market]?”
Our opinion: an attractive screenshot is worth less than a test you can repeat. Keep the original question beside the result.
Choose a small set of real questions
Start with five questions based on recent sales conversations. Include a straightforward service search, a particular customer situation, a comparison and a question about the work involved. Add the location when it affects the choice.
For a Singapore renovation firm, one question might be: “Who should I consider for a small HDB kitchen renovation?” Another could ask what to compare in renovation quotes. These are example prompts, not measured search-volume claims.
Our AI search prompt builder creates a starting set and a downloadable log. Edit the questions to reflect how your customers speak, then keep the chosen wording fixed for comparable runs.
Write down the conditions
Use a fresh conversation for each question. Record the date, platform, visible model or mode, whether web search is enabled, and whether you are signed in. Note the intended market and any relevant location setting.
If a service does not show its underlying model, record the service and mode you can see. Do not guess. Personalisation and prior conversation can change the context, so keep those conditions as consistent as possible.
A simple spreadsheet is enough. Give every answer its own row and save a screenshot or transcript. Keep unsuccessful tests too; otherwise the log becomes a collection of favourites.
Separate a mention from a recommendation
Read the answer closely. A brand can appear in a comparison without being recommended. An answer can recommend a business while linking to a third-party directory rather than its own website.
| Record | What it means |
|---|---|
| Mention | The answer names your business. |
| Recommendation | The answer suggests considering or choosing you. |
| Citation | The answer links to a source. Save the exact URL. |
| Enquiry | A prospect contacts you. Record their stated referral source. |
For example, if your business appears in 6 of 20 completed, comparable answers, the observed mention rate for that test set is 30%. It is not a market-wide visibility score. If you change the questions or platforms, start a new comparison group.
What our own test showed
On 13 October 2026, Atsell appeared in the unbranded ecommerce-agency shortlists returned by ChatGPT, Perplexity and Google AI Mode. We kept the screenshots and question. Google used a signed-in session; the other two services were signed out.
In a separate Perplexity test for Singapore AEO agencies, District10 was absent. One set of positive results for Atsell and one absence for District10 give us observations to investigate. They do not establish a stable ranking for either business.
We did not obtain answers from Claude or Gemini in that run. Record an unavailable test as unavailable, rather than counting it as a negative answer.
Follow the enquiry
Repeat your fixed questions on a schedule your team can maintain, such as weekly. Log website changes between runs and look for patterns across several dates.
Then check what happens after discovery. Review identifiable referral visits and ask prospects how they found you. A WhatsApp message saying “ChatGPT proposed your company name” tells you something a mention count cannot.
Use the record to pick the next page to improve. If answers repeatedly miss a service you genuinely offer, start with the page that explains that service.