Open dataset · CC BY 4.0 · DOI

Singapore GEO Citation Probes, September 2026

Name: Singapore GEO Citation Probes, September 2026 (ChatGPT and Google AI Mode)
Published: 2026-09-28
License: https://creativecommons.org/licenses/by/4.0/

What ChatGPT and Google AI Mode cited when asked 110 buyer questions from five Singapore industries. Every cited URL, its order in the answer, the ChatGPT answers, and our page-by-page notes on 141 cited pages.

Collected 2026-09-23 · published 2026-09-28 · v1.0 · 110 questions, 917 citation rows

Download citations.csv All files on GitHub What we built from it: the GEO Playbook →

Sample frame. 110 buyer questions from five Singapore industries (dental and aesthetics, family law, tuition, B2B SaaS, e-commerce), asked once each through APIs on 2026-09-23. Round 1: 60 questions on ChatGPT and Google AI Mode, 526 citations. Round 2: 50 questions on ChatGPT only, 241 citations.

Unit. One citation is one distinct URL within one answer. 526 citations are not 526 distinct URLs: across answers they are 474.

What it found

0 / 307

round-1 ChatGPT citations that went to forums or social sites

AI Mode: 30 / 219
19 / 507

cited URLs the two engines shared for the same question

3.7% · round 1, 60 questions
22 / 307

ChatGPT citations that were in the same question's Google top 10

7.2% · AI Mode: 89 / 219
53 / 222

URLs shared when the same question was asked twice

23.9% · 30 round-2 questions
38 / 50

dental ChatGPT citations to government and public institutions

76% · AI Mode: 7 / 52
How the Dental & Aesthetics Edition uses this →
33% → 57%

ChatGPT citations to government pages, round 1 → round 2

102 / 307 → 138 / 241 · different questions

The headline is not any single share. It is that the two engines answer the same question from different pages, and that ChatGPT mostly cites pages that are not in Google's top 10. Ranking well on Google and being cited by ChatGPT are two different jobs.

How it was measured

ChatGPT: OpenAI Responses API, gpt-5.5 with the web_search tool, approximate user location Singapore, no system prompt, the question sent as-is. Citations are the url_citation annotations in the answer.
Google AI Mode (round 1 only): a third-party SERP API with gl=sg, hl=en. Citations are the returned reference list. Round 2 ran on ChatGPT only because the SERP quota was used up.
Round 1 asked buying questions (price, who is best, A or B). Round 2 asked one question per family: definition, procedure, statistic, calculation, nearby, single-business review, complete guide, rules, checklist.
Site types (government, public, forum/social, commercial) are assigned by domain rules, not checked by hand. The rules are in the README.

Limits — read these before citing

Each question was sampled once in the official runs. When 30 questions were asked twice, only 23.9% of URLs repeated — treat any single answer as one draw, not a ranking.
Round 2 has one engine only. Samples are small and Singapore-only.
Results come from APIs, not the consumer apps; the two can differ.
Positions of copied sentences in page_annotations.csv were estimated by eye, not coded by character. The annotations were drafted by an AI multi-agent workflow and not re-checked line by line.
Not included, by the terms of the sources: Google AI Mode answer text, Google organic result lists (only aggregates are reported) and copies of third-party pages (only URLs and our notes).

Files

File	What is in it
queries.csv	110 questions: industry, round, language, our intent labels, time probed, which engines ran and how each answered.
citations.csv	917 rows: cited URL, host and registrable domain, order in the answer, engine, round, run, official flag, rule-based site type, forum/social flag; for dental, our page-type label.
chatgpt_answers.jsonl	141 ChatGPT answers with the model receipt. Official runs kept the first 600 characters; the supplementary re-runs are full text.
page_annotations.csv	Our annotations of 92 cited pages from round 1 and 49 from round 2: H1, tables, FAQ, byline, estimated position of the copied sentence.
page_types.csv	46 cited-page types with family, evidence tier and per-industry counts.
stats.md	The recount: every number with its denominator, and how each was counted.

CSV downloads from this site: citations.csv (917 rows) · queries.csv (110 rows) · page_annotations.csv (141 rows) · page_types.csv (46 rows) . The answers file, README and recount are on GitHub and Hugging Face.

Citing this

CC BY 4.0 — reuse it, including commercially, with attribution.

Pang, H. (2026). Singapore GEO Citation Probes, September 2026 (ChatGPT and Google AI Mode). Dataset v1.0. Canlah AI. Zenodo. https://doi.org/10.5281/zenodo.23005020

@dataset{pang_2026_sg_geo_citations,
  author    = {Pang, Haoyang},
  title     = {Singapore GEO Citation Probes, September 2026 (ChatGPT and Google AI Mode)},
  year      = {2026},
  version   = {1.0},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23005020},
  url       = {https://canlah.ai/data/ai-citations-sg-2026/}
}

Version: v1.0 — collected 2026-09-23. Later rounds become new versions under the same record; corrections get a dated entry here.
DOI: https://doi.org/10.5281/zenodo.23005020 (Zenodo)
Mirrors: GitHub · Hugging Face
Method book: GEO Playbook (full text in English and Chinese, plus a quick guide) · doi:10.5281/zenodo.23005022

Found an error? Write to admin@canlah.ai — corrections get a dated changelog entry on this page, not a silent edit.