How Repeat Runs Reshape ChatGPT Business Recommendations in Singapore: Evidence From 8 Industries
Asked 39 Singapore buyer questions three times each, GPT-5.5 with web search named the same business in all three answers for only 30.4% of 329 names.
Quick answer
Canlah AI asked OpenAI’s GPT-5.5 model with web search, set to Singapore, 39 buyer questions across eight service industries, three times each, on November 2, 2026. Only 30.4% of the 329 businesses it named appeared in all three answers, and the first business named was the same in all three runs for 16 of the 39 questions. Law firms were the most stable at 55.2%, while renovation and interior design firms were the least stable at 12.5%.
A single screenshot of a ChatGPT recommendation is one sample, not a ranking. The industries also differ in where the model looks: legal and renovation answers lean on third-party directories, while tuition, school and corporate-services answers cite the businesses’ own websites.
Disclosure: Canlah AI ran this study and sells AI visibility measurement and GEO services. No business named in the answers is named on this page, and no business paid to be included.
Key findings
- Most names change between runs: 100 of the 329 distinct businesses named across 117 answers, or 30.4%, appeared in all three answers to the same question.
- The top name holds less than half the time: the first business named was the same in all three runs for 16 of 39 questions, or 41.0%.
- Law firms are the most stable: 55.2% of the 58 law firms named appeared in all three runs, and 54.0% of the 124 links cited in legal answers pointed to legal directories such as Chambers and Partners, The Legal 500 and Doyle’s Guide.
- Renovation is the least stable: only 12.5% of 32 renovation and interior design firms appeared in all three runs, although directories and listing sites made up 54.9% of the 82 links cited.
- Own websites carry some industries: 91.7% of the 145 links in tuition answers and 76.5% of the 102 links in school answers pointed to business websites rather than directories or government pages.
- Every answer named businesses: all 117 answers recommended named firms, 5.2 per answer on average.
Study design
The 39 questions were written the way a Singapore buyer would ask them, such as “Best divorce lawyer in Singapore?” or “Best tuition centre in Singapore for PSLE maths?”. They covered eight industries: law firms, tuition centres, schools and preschools, renovation and interior design, corporate secretarial and accounting, home services, IT support and wedding services.
Each question was sent three times through the OpenAI API to GPT-5.5 with the web search tool on and the user location set to Singapore, between 03:13 and 03:37 Singapore time on November 2, 2026. Business names were extracted from each answer with Claude Haiku, then normalised by removing accents, punctuation and suffixes such as “Pte Ltd” before the runs were compared. A business counts as “in all three runs” when it appears in all three answers to the same question.
The same method was applied to five comparison sets run on the same day: 22 Singapore shopping questions, 12 Singapore SME software questions, 15 fintech buyer questions for Asia, 8 Singapore hotel questions and 12 Sydney service questions.
Results
Table 1. How stable GPT-5.5’s recommendations were, by Singapore industry, November 2, 2026
| Industry | Businesses named in all three runs | Same first name in all three runs | Main source of cited links |
|---|---|---|---|
| Law firms (7 questions) | 32 of 58 (55.2%) | 5 of 7 | Legal directories, 54.0% of 124 |
| Schools and preschools (6) | 16 of 50 (32.0%) | 3 of 6 | School websites, 76.5% of 102 |
| Wedding services (3) | 7 of 23 (30.4%) | 2 of 3 | Vendor websites, 66.7% of 63 |
| Home services (4) | 9 of 31 (29.0%) | 1 of 4 | Listing sites, 39.7% of 68 |
| Tuition centres (8) | 19 of 72 (26.4%) | 4 of 8 | Centre websites, 91.7% of 145 |
| Corporate secretarial and accounting (4) | 8 of 37 (21.6%) | 0 of 4 | Provider websites, 71.4% of 77 |
| IT support (3) | 5 of 26 (19.2%) | 1 of 3 | Provider websites, 73.2% of 56 |
| Renovation and interior design (4) | 4 of 32 (12.5%) | 0 of 4 | Directories and listings, 54.9% of 82 |
| All eight industries (39) | 100 of 329 (30.4%) | 16 of 39 | Directories 23.4%, government 18.1% of 717 |
Source: Canlah AI run of 39 buyer questions × 3 through the OpenAI API (GPT-5.5 with web search, location Singapore), November 2, 2026. Government includes .gov.sg pages, the Consumers Association of Singapore and the Law Society.
Table 2. The same measure in the five comparison sets
| Question set | Businesses or brands named in all three runs | Same first name in all three runs |
|---|---|---|
| Singapore shopping (22 questions) | 104 of 294 (35.4%) | 16 of 22 |
| Singapore SME software (12) | 43 of 139 (30.9%) | 8 of 12 |
| Fintech buyers in Asia (15) | 70 of 284 (24.6%) | 5 of 15 |
| Singapore hotels (8) | 14 of 68 (20.6%) | 3 of 8 |
| Sydney services (12) | 18 of 109 (16.5%) | 2 of 12 |
Source: same method and date as Table 1.
Shopping questions were the most stable comparison set: the top brand held in 16 of 22, and in 14 of 22 at least four brands appeared in all three answers. Local services in Sydney were the least stable, at 16.5%.
What the November 2 ChatGPT dataset does and does not show
The November 2 ChatGPT dataset shows how much one model’s named recommendations moved between three runs of the same question on one day. It does not show which business is best, and it does not rank the businesses that were named. The pattern between source type and stability is an association, not a cause.
The dataset measured the OpenAI API with web search, not the ChatGPT app, which can phrase answers and choose sources differently. Three runs per question is a small sample, so a business that appeared once may still appear often over more runs, and a stable name on November 2 may not stay stable.
The contrast between law and renovation suggests that the kind of directory matters more than whether there is one. Legal directories publish ranked bands with a few firms in each, while renovation listing sites host profiles for hundreds of firms, which leaves the model more candidates to rotate through.
Implications
For a Singapore business, the first implication is that a single ChatGPT screenshot, whether flattering or not, says little. A fair reading needs the same question asked several times, and in Canlah AI’s measurement work each buyer question is asked 5 to 10 ways and re-tested monthly.
The second is to be present in the sources the model reads in that industry. In law, that means the legal directories; in renovation, the listing platforms and the consumer association’s accreditation pages; in tuition, schools and corporate services, the business’s own website, with the facts a buyer asks about stated plainly.
The third is to track the stable core rather than a single position. A business that appears in all three runs of a question is part of the model’s shortlist for it; a business that appears once is a candidate.
Where Canlah AI fits
Canlah AI is a Singapore-based SEO + GEO agency that measures how often AI answers name a business and works to raise it, in English and Chinese. Standard tracking covers ChatGPT, Gemini, Google AI Overviews and Google AI Mode, and every buyer question is re-run rather than read once, because answers move between runs as Table 1 shows.
Its GEO service is priced per AI platform from $1,500 a month (pricing), and its free AI visibility report shows where a business stands in AI answers before any work begins.
Frequently asked questions
Does ChatGPT recommend the same businesses every time?
No. In Canlah AI’s November 2, 2026 test of 39 Singapore buyer questions asked three times each, only 30.4% of the 329 businesses GPT-5.5 named appeared in all three answers, and the first name repeated in all three runs for 16 of 39 questions. Law firms were the most stable at 55.2%.
Why are ChatGPT’s recommendations more stable in some industries?
In this dataset, the most stable industry, law, was the one where the model cited ranked legal directories for over half of its links. Renovation also leaned on directories but was the least stable, likely because its listing sites carry far more firms. The link between source type and stability is an association, not a proven cause.
How should a business check whether ChatGPT recommends it?
A business should ask the same buyer questions several times, in a fresh session or through the API with web search on, and record which names appear in every run. Canlah AI’s guide to checking ChatGPT recommendations sets out the steps, and its free AI visibility report runs the check across engines.
Methodology and limitations
The questions were run through the OpenAI Responses API with the GPT-5.5 model, snapshot gpt-5.5-2026-04-23, and the web search tool with the user location set to Singapore, three times each between 03:13 and 03:37 Singapore time on November 2, 2026. The Sydney set used a Sydney location. Answers came from the API, not the ChatGPT app.
Business names were extracted by Claude Haiku from each answer and compared after normalisation that removed accents, punctuation and common suffixes such as “Pte Ltd” and “Singapore”, with one name matching another when its words were a subset of the other’s. Exact matching without normalisation gives lower stability figures; for the shopping set, it gives 12 of 22 questions with at least four brands in all three runs instead of 14.
Cited links are the URL citations attached to each answer’s text, classified by host into directories and listing sites, government and official bodies, and all other sites, which were mostly the businesses’ own websites. The study covers 39 questions in eight industries and three runs per question, so the per-industry figures rest on small samples and should be read as directional.
- Canlah AI ChatGPT recommendation run, November 2, 2026: 39 Singapore service questions and five comparison sets, raw answers archived.
- OpenAI web search tool documentation. How the API’s web search and user location work.
- Canlah AI: How to check if AI search recommends your brand. Steps for a repeat-run check.
See whether ChatGPT names your business on repeat runs. Run the free AI visibility audit →
Related articles
FREE WHITEPAPER
Marketing in the Agent Era
All 13 chapters public — no email wall. Includes an original dataset on the agent-readiness of 50 cross-border DTC brands.
Read it free →