Singapore AI Search: what engines recommend
Most articles about "who does GEO in Singapore" are written by an agency that puts itself on the list. This one is a record of what two answer engines said when we asked them — including the part where they did not say us. The raw answers, timestamps and conversation links are kept so you can check any number here.
What we asked
Two buyer questions, in English, on 2 November 2026, asked once each on ChatGPT (with web search) and Perplexity:
Which agencies in Singapore offer generative engine optimization (GEO) services? List them with their websites.
Our Singapore B2B company is not showing up when buyers ask ChatGPT or Perplexity for suppliers. Which agencies in Singapore can help fix this, and what do they cost?
On 4 November we asked Chinese versions of both questions, and asked the English version again the same day as a control. Six raw answers in total. One question, one answer per engine per run — no re-rolling until a nicer list appeared.
1. The two engines barely agreed
On 2 November, ChatGPT named 16 agencies and Perplexity named 7. Only 5 appeared on both lists: Hashmeta, Digitrio, Impossible Marketing, OOm and Rankpage.
A company that checks its visibility on one engine has, at best, seen half the picture it cares about.
2. Change the question, change the sources
For the first question, Perplexity cited hashmeta.com, digitrio.com.sg, bthrust.com, advergreen.com and an Advergreen blog post. For the second question it cited a largely different set: ismartcom.com, simaia.co, an Abnewswire release, stridec.com and aeoseo.sg.
Same engine, same day, same buyer intent phrased differently — almost entirely different sources. The sources an engine quotes are not the same thing as the agencies it names, and a company needs to show up in both.
In our ChatGPT captures the cited-source list was empty. That means no machine-readable citation links were present in those answers; it does not mean ChatGPT cites nothing at all.
3. Change the language, change the shortlist
On 4 November we held everything else constant and changed only the language of the first question. The named agencies changed on both engines:
- ChatGPT — English only: Aivancement, Plexial, SerpEyes, Canlah AI. Chinese only: AI Studio, Rankpage.
- Perplexity — English only: AEMorph, DigitalMarket, iSmartCom. Chinese only: Outrankco, Digitrio, MediaPlus, PurpleClick.
The Chinese answers also leaned toward price tables (from S$688/month to S$1.5k–3.5k/month) and local comparison articles; the English answers leaned toward company-plus-URL lists.
Read this carefully: each question was asked once, and answer engines are non-deterministic. This observation shows that language changes the result; it does not prove language is the only cause, and it does not say which language favours which kind of firm. The 2 November English baseline and the 4 November Chinese run are different days and must not be compared directly — that is why we ran a same-day English control.
4. We appeared in none of the six
All six records have appeared: 0. We sell the work of getting a company into these answers, and our own company was not in them. We are leaving that number in, because a report that quietly removes the inconvenient result is not a report.
What a company should take from this
- Measure on more than one engine. A single-engine check cannot see the disagreement.
- Use the buyer's actual wording, in every language they use. Brand-keyword checks and English-only checks both miss answers.
- Track mentions and citations separately. Being in the answer text and being a source the engine quotes are two different jobs.
Limits and raw data
These are six single-run answers on two days in November 2026. They are not market share, not a ranking, and not evidence that any agency is better than another. Answer engines change per question, per user and per run; a name absent today can appear next month with nothing having changed.
The verbatim answers, cited URLs and timestamps are in the project's public evidence folder: github.com/TIY-IT/tiy-geo-website / state/probes. The 4 November run and its same-day control are summarised in geo-probe-log-zh.csv.
We will re-run the same questions monthly and publish what comes back, including our own count.