Two prompts can express the same practical need and still lead toward different businesses because language changes names, geographic cues, categories, and the source environment surrounding the question.
An English prompt asked for information about a Bangkok wellness clinic using its most common transliterated name. The answer described a private hospital, cited the clinic’s treatment page, and added an emergency-care claim absent from that page. A matched Thai prompt returned the clinic as a specialised treatment provider, placed it in the correct district, and showed a map listing alongside its Thai-language website.
The clinic was Study Object A, a composite scenario assembled from recurring observations rather than a named real business. The Thai answer was not clean. It attached an old operating hour to the clinic and translated one treatment category awkwardly. Yet it appeared to identify a different kind of entity from the English answer, despite the shared discovery intent.
What counts as the same question across languages
A matched bilingual prompt is harder to construct than a literal translation suggests. Thai and English questions may differ in formality, geographic specificity, category language, and the way a business name is normally written. Preserving identical words can distort the practical request.
For the clinic scenario, the laboratory built prompts around the same discovery intent: identify the named provider, describe its services, and state where it was located. The Thai prompt used the ordinary Thai name and district wording. The English prompt used the transliteration found most often on English-facing listings. A second English formulation used another plausible transliteration. The remaining run conditions were preserved as closely as the observation procedure allowed.
A matched bilingual run is a comparison of prompts with equivalent discovery intent because literal wording alone cannot preserve how names and places function in each language. The equivalence is procedural and semantic, not character-for-character.
This distinction matters in Thailand. A district name may be sufficient in Thai because local context narrows the place. English listings may append Bangkok, a province, or a landmark. A Thai business category may describe the provider’s function precisely, while an English platform substitutes a broader category familiar to international users.
The laboratory records these differences rather than hiding them under the label of translation. Otherwise, a language comparison can become circular: the English prompt is judged against an unnatural translation, then its different answer is attributed to language.
The team also preserves the exact name form used. Transliteration is part of the experimental condition. One English spelling may retrieve the clinic’s own site, while another resembles the name of a larger medical facility. Treating both spellings as interchangeable would remove the very ambiguity under examination.
The entity can change even when the name survives
In the clinic observations, several English answers retained the expected name string. At first glance, this looked like successful identification. The surrounding attributes told a different story.
The generated category shifted from clinic to hospital. The location expanded from a Bangkok district to a broad metropolitan label used by the larger facility. Emergency services appeared, although the composite clinic offered scheduled treatments. One citation led to the clinic’s website, but another described the larger medical organisation whose English name resembled one transliteration.
A correct name string does not establish that the correct entity was identified. Entity identification concerns the apparent connection between the name and one particular business, branch, brand, or place. The surrounding category, location, services, ownership, and sources help show which entity the system appears to have assembled.
The Thai runs generally attached the clinic name to its treatment pages and local map listing. The English runs were more variable. Some identified the clinic correctly but inflated its category. Others produced a hybrid entity: the clinic’s name and treatments combined with the larger facility’s institutional description.
This hybrid state is easy to miss because readers tend to ask whether the answer named the business. The laboratory asks a stricter question: what set of attributes has been placed behind the name?
Entity identification fails quietly when the label stays familiar but the business underneath it acquires another organisation’s shape.
The source relationships varied within the same answer. The clinic’s treatment page directly supported several service claims. It offered stretched support for calling the provider a general medical centre. Material from the larger facility created borrowed identity when its emergency and inpatient functions entered the clinic description. The claim that the clinic was “internationally accredited” was an unsupported arrival in the visible record.
Language did not produce one single type of error. It changed the mixture.
Why Thai and English retrieval environments differ
The public record surrounding a Thai business is rarely duplicated evenly across languages. The Thai website may carry precise service names, branch descriptions, district references, and legal wording. English pages may be shorter, older, platform-generated, or written for visitors with different category expectations.
In the composite clinic case, the Thai service pages distinguished consultation, treatment, and follow-up care. An English directory compressed these into “medical services.” A booking platform used “wellness centre.” A map listing assigned a broad clinic category. The similarly named larger facility had extensive English pages using terms such as hospital, specialist care, and emergency department.
When the prompt was written in English, the visible source set more often included those broad or institutional descriptions. That observation is compatible with several explanations. English wording may have changed the search interpretation. The transliterated name may have increased similarity between the entities. English-language pages may have offered stronger lexical matches. The laboratory could not determine which internal route dominated.
Thai prompts also carried geographic cues that behaved differently. A district name written in Thai often remained attached to the clinic. In English, the same place was sometimes normalised to Bangkok or paired with a similarly named district elsewhere. One run located the provider correctly but cited a travel directory whose page described the neighbourhood rather than the business.
The difference should not be romanticised. Thai-language material is not inherently more accurate. It can be outdated, duplicated, sparse, or inconsistent across platforms. In the clinic scenario, an old Thai listing supplied the incorrect operating hour. Another Thai page used a category associated with traditional treatment, although the clinic had broadened its services.
The useful point is narrower: changing language alters the field of names, categories, place labels, and source pages available to the answer. The resulting observation deserves independent inspection.
Four recurring outcomes in bilingual comparisons
Across the composite clinic and restaurant objects, the laboratory used four descriptive patterns to distinguish the outcomes of matched Thai and English runs. These patterns do not replace the canonical Four Source Relationships typology.
The first is stable identification. Both languages point to the same entity, branch, category, and location, though wording and citations may differ. Source relationships still require claim-level checking. Two answers can identify the same clinic while one stretches a service description further than the other.
The second is category drift. The entity remains recognisable, but one language assigns a broader, narrower, or adjacent business category. The clinic becoming a hospital belongs here when the remaining attributes still concern the clinic itself.
The third is entity substitution. One language selects a different business, branch, or place. In Study Object B, the composite regional restaurant group, an English prompt using a shortened name sometimes returned the similarly named independent venue in a neighbouring province. The Thai prompt more often retained the group’s Bangkok branch.
The fourth is hybrid identification. Attributes from two entities are assembled behind one name. This occurred when the English answer named the clinic but borrowed the larger facility’s institutional functions, or when a restaurant answer used the intended branch name with another venue’s address and atmosphere.
These patterns describe what changed between observations. The Four Source Relationships classify how each claim relates to visible evidence: direct support, stretched support, borrowed identity, or unsupported arrival. The two schemes answer different questions. One maps bilingual outcome differences; the other evaluates evidential relationships inside each answer.
Hybrid identification is especially important because a simple “same business or different business” comparison misses it. The answer may preserve enough correct details to pass a casual check. It is partly the expected entity and partly something else.
The laboratory does not convert these patterns into scores. No threshold establishes that two answers are “seventy percent the same.” The comparison remains qualitative unless an actual preserved sample supports numerical reporting.
Repeating the run without demanding identical sentences
Generative answers vary. A repeated Thai prompt may reorder services, change its confidence, or display another citation while retaining the same entity. An English prompt may produce a different paragraph without changing the underlying category error.
For this reason, repeatability means preserving enough of the prompt conditions and procedure to conduct the inquiry again and compare the returning pattern. It does not require identical wording.
The laboratory records the prompt text, language, name form, model context, observation date, visible citations, and other relevant conditions. Renewed runs are then compared at the level of entity, branch, category, location, services, and claim-source relationships.
In the clinic scenario, the exact hospital wording did not return every time. The broader institutional identity did. One run added emergency care. Another called the clinic a private medical centre and omitted emergency services. A third used the correct clinic category but cited the larger facility. The surface sentences moved around a recurring ambiguity.
Thai runs varied too. The wrong operating hour returned more consistently than the awkward treatment translation. A map citation appeared in some observations and disappeared in others. These changes were recorded rather than averaged into one “Thai answer.”
Cross-model comparison added another layer. Several systems produced some form of category inflation in English. That agreement did not confirm that hospital was the correct category. Shared public sources or naming ambiguity could lead several systems toward the same error.
A bilingual comparison is therefore a set of observations, not a contest in which one language wins. In some cases Thai may preserve local entity distinctions. In others English may surface a current official page omitted from the Thai answer. The method asks where the representations diverge and what visible evidence accompanies each version.
What can be concluded from a language difference
A change between Thai and English answers can establish that entity identification or claim-source relationships were not stable under the preserved prompt conditions. Repeated runs may show that the difference returns and belongs to a broader pattern.
The method cannot reveal private retrieval systems, hidden ranking logic, undisclosed intermediate steps, or every source used internally. It cannot prove that the language itself caused the change. Name form, source availability, query interpretation, model context, and platform categorisation may all move together when the prompt language changes.
Literal causal claims are therefore premature. “The English prompt retrieved the wrong clinic because English sources are worse” would exceed the record. The evidence may instead show that English runs more often produced a broader category and mixed source set under the tested conditions.
The composite objects impose another limitation. They are built to expose recurring mechanisms while avoiding unsupported claims about named businesses. They illustrate plausible patterns from preserved work, but they are not a published census of Thai AI search behaviour.
Predictions must remain provisional. If the same clinic were to standardise its English naming, clarify branch pages, and align platform categories, the laboratory might expect less variation in later runs. That expectation could be weakened if hybrid identification continued under quoted names and tightly specified locations.
The durable conclusion is methodological. Thai and English answers should not be treated as translations of one underlying result. Each is an observation with its own entity identification, sources, categories, and uncertainties.
Sometimes the system retrieves a different version of the business. Sometimes it appears to retrieve a different business altogether. The name alone will not tell them apart.