ChatGPT does not simply search the sentence a person types. It rewrites the question, searches the web and sometimes a local-business feed, then chooses a small number of sources to cite.
On 17 November 2026, I captured the network responses behind five ChatGPT conversations and one Google AI Mode conversation from Bangkok. The tests covered English, Burmese, and Thai questions about SEO, search market share, and local agencies in Yangon and Bangkok.
This is a small field study, not a ranking formula. The value is in seeing the intermediate data: the queries an AI system writes, the results it receives, the business records it uses, and the sources that survive into the answer.
What we tested
The test set compared five ChatGPT conversations in Auto mode with one Google AI Mode conversation. ChatGPT's browser response exposed JSON fields that could be inspected field by field. Google AI Mode returned streamed HTML, so that test focused on the visible answer and outbound links.
| Test | Prompt language | Web results | Local results | Main citation or result |
|---|---|---|---|---|
| ChatGPT 1 Investing in SEO in Myanmar | English | About 150 across 3 searches | — | DataReportal, StatCounter, World Bank |
| ChatGPT 2 Thailand search share this month | English | 3, all StatCounter | — | StatCounter |
| ChatGPT 3 SEO agencies and prices in Yangon | English | 12 from 8 domains | 5 from Google Maps | marketingmyanmar.com ×3 |
| ChatGPT 4 Same Yangon question | Burmese prompt, English query | 12 from 8 domains | 6 from Google Maps | marketingmyanmar.com ×3 |
| ChatGPT 5 SEO agencies and prices in Bangkok | Thai prompt, English query | 12 from 11 domains | 4 of 5 Maps listings | siit.co, imvisible.tech, seotal.com |
| Google AI Mode Best agency in Yangon and price rates | English | 9 domains cited | None shown | blink.com.mm, its own FAQ page |
Why look at network data?
Most advice about optimising for AI search is guesswork because the answer hides how it was built. The browser does not hide everything. In ChatGPT, the conversation response contains the exact queries the model wrote, each result returned, pages opened, and sources cited.
That matters in Southeast Asian markets. English-language advice often assumes a deep web of local sources. Myanmar has relatively few, while Thailand has a crowded market of agency listicles. Network data lets us ask who gets retrieved, who gets recommended, and what evidence appears to influence the final answer.
Retrieval is not citation. A page can appear in the results and still lose the answer because another page is more closely matched, more useful for the question, or easier for the system to describe.
How ChatGPT writes its searches
ChatGPT never searched with the user's sentence in these tests. It generated short English keyword queries, often added the current year, and sometimes named the source it expected to use.
Broad questions split into angles
“Should I invest in SEO in Myanmar?” became three searches covering market size, business climate, and digital-access risk. Narrow fact questions became one query.
Thai and Burmese became English queries
The Burmese and Thai prompts were translated into English searches. The answers were localised again, including Burmese or Thai wording and local currency.
A source can be named before retrieval
Asked for Thailand's search engine share, ChatGPT added “StatCounter” to its own query. All three results then came from StatCounter.
“This month” became a dated search
The system searched for the last complete month, October 2026, published in November. It opened the page instead of relying only on the snippet.
The broad-question result is especially useful for publishers: the three angles in a query fan-out are the topics a page or content cluster needs to cover if it is meant to be retrieved for the larger question.
How ChatGPT picks local businesses
For “best agency in [city]” questions, ChatGPT ran a web search and a local-business search at the same time. The local results carried Google Place IDs, Maps links, ratings, review counts, opening hours, and a provider ranking score.
In Myanmar, the list came from Google Maps
Web results for Yangon were mostly directories and blogs, but every recommended agency came from the business feed. The same five agencies appeared in both the English and Burmese tests, but the order changed.
| Agency | English position | Burmese position | Maps rating and reviews | Provider score |
|---|---|---|---|---|
| SEO Myanmar | 1 | 5 | 4.6 (8) | 0.99993 |
| Pixellion 360 | 2 | 2 | 4.8 (26) | 0.99998 |
| Digital Dots | 3 | 3 | 4.7 (26) | 0.99994 |
| BEYOND 360 | 4 | 4 | 5.0 (19) | 0.99994 |
| Amara Digital | 5 | 1 | 4.4 (17) | 0.99960, lowest |
| DigitalMarketing.com.mm | — | 6 | 4.8 (93) | 0.99999, highest |
The order followed neither rating, review count, nor the provider score. In the Burmese answer, the first three were sorted by founding year. In the English answer, the agency whose Maps listing was literally named “SEO” came first. Getting into the Maps results put a business on the candidate list; the model's description of the business influenced its position.
In Thailand, listicles and Maps were combined
Bangkok has a dense web of “Top 10 SEO agencies” pages, so the answer used both web mentions and Maps listings. IBEX appeared first from its own “Top 10 Best SEO Agencies in Bangkok & Thailand” page despite having no matching Maps record in the result set. UNEED Digital had a 5.0 rating and 41 reviews but appeared only as a map pin.
“Getting into the Maps results gets a business onto the list; how the model describes the business decides its position.”
A working interpretation from the Yangon testsWhat gets cited
ChatGPT retrieved about 12 results per agency search but cited only two or three. The clearest pattern was topic depth: marketingmyanmar.com had three to four closely matching pages in each Yangon search and won all three price citations in both languages. Larger directories were retrieved but not cited for prices.
The year in a title can beat the real publication date
| Page | Title says | Published | Cited? |
|---|---|---|---|
| marketingmyanmar.com pricing guide | “What to Expect in 2026” | August 5, 2025 | Yes, described as a 2026 guide |
| marketingmyanmar.com agency ranking | “Best SEO Agencies in Myanmar (2026)” | July 8, 2025 | Yes |
| kabarkyawmarketing.com pricing guide | “2026 Pricing Guide” | June 29, 2026 | No |
| sorank.com Bangkok list | “(2027)” | November 7, 2026 | No |
This is a signal, not a recommendation to mislabel old content. It does suggest that publishers should keep titles, visible dates, and structured data aligned so a retrieval system can understand the page's intended time period.
Self-published content can enter the results
An agency's own “Top 10” page put that agency at number one in Bangkok. An agency city page with a public price table was the only agency page shown by itself in the English Yangon answer. Useful, specific first-party pages can be retrieved, but they compete with listicles, directories, and independent mentions.
Links are approved against a public web index
Every link in the ChatGPT answers appeared in a safe_urls list. Most came from tool results, but ChatGPT also linked two agency domains that the Maps listings had not supplied. Those URLs were still approved because they were already known to the public web index.
This matches OpenAI's published explanation of URL safety: the exact URL is checked against an independent index of public addresses before automatic retrieval. A public, indexed URL can therefore be linked even when it was not opened in the current conversation.
Logo images are unreliable
In all three agency tests, image search returned the wrong company for several agencies. Clear brand-and-city filenames and useful alt text on a site's own logo can reduce the chance of an unrelated image being attached to the answer.
How Google AI Mode compares
Google AI Mode built the Yangon answer from websites alone and exposed no sub-queries in the browser. Because the prompt was vaguer than the ChatGPT agency questions, this is not a like-for-like comparison.
| Signal | ChatGPT agency test | Google AI Mode |
|---|---|---|
| Queries visible in the browser | Yes, in search_model_queries | No |
| Response format | JSON | Streamed HTML |
| Handling of unclear intent | Not tested in the same prompt | Split into marketing, recruitment, and travel agency intents |
| Top marketing source | Google Maps listing | blink.com.mm, the agency's own FAQ page |
| Social posts as sources | Not seen | A Facebook post and its blog article appeared |
The three Google intents matched autocomplete suggestions in the same session. Google has also described AI Mode as splitting questions into subtopics and searching them together. That makes Search Console a useful place to watch for unusually long queries that reveal the questions behind AI-generated visibility.
What marketers should do with these findings
The tests point to six practical actions for agencies and local businesses in Myanmar and Thailand:
- Complete your Google Business Profile. Use the exact business category, real brand name, service area, opening hours, and a steady stream of genuine reviews.
- Earn mentions on pages that get retrieved. In crowded markets, a Maps listing alone did not earn a written recommendation. Study the listicles, directories, and comparison pages appearing for your target questions.
- Answer buyer questions in English with local detail. Thai and Burmese prompts were searched in English, so an English page with local prices, examples, and context can be important even when the final answer is localised.
- Publish prices and dates clearly. State ranges plainly, show a visible update date, and keep the page title and dateModified value current.
- Build topic depth. Several useful pages on one topic can outperform a larger directory with only a single generic page.
- Keep the crawler in. Allow OAI-SearchBot if you want your public pages to be eligible for ChatGPT search, and check that your firewall or CDN allows the published bot traffic.
To track progress, rerun the same prompts each month in a Temporary Chat. Record whether your domain appears in the results, whether it is cited, and where your business sits in the list. The goal is a repeatable observation log, not a promise of a fixed position.
Limitations and open questions
These are six conversations, one run each. Treat the findings as strong signals rather than rules. Memory was on in the ChatGPT tests, and some closing questions echoed earlier chats, so personalisation may have shaped parts of the answers.
Page text was not fully exported, which means we cannot always tell whether a fact came from a retrieved page or from training data. We did not repeat identical prompts, so we do not yet know how much the list order changes between runs. The Google AI Mode prompt also differed from the ChatGPT prompts and was probably logged in.
Finally, ranking factors are not published. Every explanation of “why” is an inference from the captured data, not a confirmed statement from OpenAI or Google.
Frequently asked questions
How does ChatGPT decide which websites to cite?
In these tests it wrote its own keyword queries, retrieved about 12 results per agency search, and cited two or three. The cited pages matched the query closely, came from sites with several pages on the topic, and often carried the current year in their titles.
Does ChatGPT use Google Maps data?
Yes, for local questions in this study. The Yangon and Bangkok business results carried Google Place IDs, Maps links, ratings, and review counts. Every recommended Yangon agency came from that business data.
Does ChatGPT search in Thai or Burmese?
Not in these tests. Both prompts were translated into English queries and the answers were translated back. One Thai-language comparison page was still retrieved and cited.
What is query fan-out?
Query fan-out is when an AI search engine splits one question into several narrower searches and runs them together. ChatGPT exposes these queries in its network response; Google AI Mode did not expose them in the browser session we captured.
What is safe_urls?
It is the list of addresses the ChatGPT interface may link in an answer. URLs already known to OpenAI's public web index can pass automatically; others may trigger a warning.
Sources and citation
Primary data: six network responses captured on November 17, 2026 — five ChatGPT conversations and one Google AI Mode conversation.
- Keeping your data safe when an AI agent clicks a linkOpenAI, January 28, 2026.
- Overview of OpenAI CrawlersOpenAI documentation for OAI-SearchBot and GPTBot.
- Google Adds AI Mode Traffic To Search Console ReportsSearch Engine Journal, June 16, 2025.
Suggested citation: Wei Yan Min, “How ChatGPT and Google AI Mode Choose Sources: 6 Network-Level Tests in Myanmar and Thailand,” November 17, 2026, this article.
AI search is easier to improve when you can observe the path from question to query to citation.