A branch error can survive every quick check: the brand exists, the address looks plausible, and the cited page is relevant. The failure appears only when those pieces are tested as one location-specific identity.
In a composite scenario involving a Thai restaurant group, an English prompt asked whether the business had a branch in a neighbouring province. The generated answer said yes, supplied an address, and cited a map listing. The brand name was correct. So was the province. The address, however, belonged to an unrelated venue with a similar name, while the stated opening hours came from the group’s Bangkok branch.
A Thai-language run produced a quieter error. It identified the provincial branch correctly but described the group’s full menu and reservation policy as though both applied at that location. The branch page showed a shorter menu and no reservation option. Nothing in the answer sounded absurd. The system had found the organisation, then allowed the organisation to swallow the branch.
A parent brand is not a location
Branch networks are represented unevenly across the public web. A company site may have one page for the whole group and separate pages for some locations. Maps usually create a record for each physical place. Booking platforms may list only reservable branches. Social pages can represent the brand, a local branch, or both, sometimes without saying which. Directory entries frequently flatten the structure into one business with several addresses.
That patchwork produces a simple-looking but consequential question: what exactly has the answer identified? The parent organisation, one branch, a place carrying the same name, or a mixed record assembled from all three?
The composite restaurant group made this distinction difficult. Its Thai branch labels were fairly clear, but English versions varied. One platform used the district name, another the province, and a third omitted the location word from the page title. The group’s main website described shared brand history and cuisine, while map listings supplied branch-specific hours and addresses. An older directory page treated a former location as current.
The laboratory uses “branch identification” for the apparent connection of a generated claim to one physical or operational location. Branch-specific facts must remain attached to that location rather than being assumed from recognition of the parent brand. An answer can identify the correct company and still fail at the branch level.
The distinction matters whenever the user’s decision depends on place. Opening hours, services, stock, menus, parking, accessibility, booking, delivery range, and contact details can all differ within one organisation. A general brand description may be accurate and still be useless for the branch the reader intends to visit.
What the repeatable runs preserved
The laboratory built recurring inquiries around three common tasks: checking whether a branch existed, asking for its location, and requesting branch-specific operating information. Thai and English formulations were included where the branch label changed with language. The procedure preserved the prompt, generated answer, visible citations, model context, observation date, and relevant run conditions.
Repeatability did not mean demanding identical prose. The systems varied wording, list order, and confidence. The team looked instead for the return of a structural pattern. Did the same branch remain selected? Did its address stay stable? Were hours, services, and booking details attached to the same location? Did the cited pages continue to refer to that branch?
The researchers used several prompt forms without treating one as the canonical query. Some named the parent brand and province. Some used the local branch label. Others described the service and asked for the nearest location. This variation exposed different apparent retrieval paths. Exact branch wording tended to keep the name stable, while service-led prompts more readily widened toward the whole brand or a nearby look-alike.
Each output remained an observation. The team recorded what appeared before proposing why it appeared. When a provincial address was paired with Bangkok hours, the mismatch was visible. The explanation was less certain. The system may have retrieved a parent page for general information and a map listing for location. It may have merged records earlier in the process. It may have relied on another source that was not displayed.
A useful irregularity appeared in one run. The answer found the correct branch, gave the right hours, and cited the branch map profile, but used the parent group’s telephone number. That number still reached the business, so a casual user might never notice. Operationally, though, the answer had crossed levels within the organisation.
Four branch states in generated answers
The laboratory described the observed branch outcomes through four qualitative states. A resolved branch was one where the brand, branch label, address, operating details, and sources pointed coherently to the same location. A parent collapse occurred when group-level information was presented as branch-specific. A sibling transfer occurred when facts from one legitimate branch moved to another. A false branch occurred when an unrelated or former location was represented as part of the organisation.
These states are not scores, and they do not imply a progression from mild to severe. A parent collapse can be harmless when the shared fact is brand history, yet damaging when it concerns reservations or services. A sibling transfer may preserve the right company but send a customer across a city. A false branch may arise from an old listing rather than a similarly named competitor.
The branch-state classification describes which organisational level the answer appears to have selected. The laboratory’s Four Source Relationships typology answers a separate question: how well does a visible source support each claim? Keeping those two analyses apart prevents a relevant citation from hiding the wrong branch.
Consider the sentence, “The group’s provincial branch is open until 10 p.m. and accepts reservations.” The branch page may directly support the location. A Bangkok map listing may directly support 10 p.m. for a different branch, creating borrowed identity when applied to the provincial one. A parent website mentioning enquiries may provide only stretched support for reservations. If no displayed source supports booking at all, that clause becomes unsupported arrival.
This claim-by-claim reading is slower than checking whether the citation domain looks respectable. It is also closer to the actual failure. A source can be official, current, and still belong to the wrong branch.
Where branch signals lose their attachment
Branch labels are often weak metadata. A location word may appear in a page title but disappear from the body. A map profile may show the branch name while user photographs display another branch’s signage. A booking page can use the parent brand in its heading and reveal the location only in the address. When generated prose compresses those records, the location marker is easy to drop.
The composite group showed three recurring attachment problems. First, general descriptions travelled downward. Brand history, cuisine, and broad service language were assigned to every branch without checking local variation. Second, operational details travelled sideways. Hours, telephone numbers, and reservation policies moved between sibling branches. Third, stale records travelled forward. A closed or renamed location reappeared as an active branch because older directory pages remained available.
Language altered some of these signals. Thai prompts could preserve a locally familiar district or branch label that English prompts replaced with a broader province. English transliterations sometimes made two branch names appear more alike than their Thai forms. Yet Thai did not automatically solve the problem. A Thai-language answer could still collapse the group into one business record or repeat a platform category that ignored branch structure.
The team therefore treated location hierarchy as evidence, not decoration. Province, city, district, subdistrict, landmark, and mall name can each narrow an entity differently. Two sources may seem to agree because both mention Bangkok while referring to locations far apart. Conversely, a branch can be known locally by a mall or neighbourhood name that does not match the formal district listed elsewhere.
The apparent retrieval-path interpretation becomes more credible when several branch signals move together. A changed address accompanied by changed hours, telephone number, and citation set is compatible with a different location becoming central. One changed detail may instead be stale data or ordinary generation variation. The evidence is thinner than the finished sentence makes it look.
What branch errors imply for visibility work
A brand-level mention can conceal a location-level failure. This is especially important for organisations whose revenue, availability, or customer experience varies by branch. A national or regional company may appear prominently in generated answers while individual locations are missing, merged, or represented with the wrong operating facts.
The first diagnostic question is not simply “Did the business appear?” It is “At what organisational level did it appear?” The answer may refer to the parent company, a specific branch, several branches treated as one, or a place that only resembles the brand. That distinction should be made before evaluating sentiment or recommendation position.
Branch verification also changes how source consistency is understood. Perfectly identical descriptions across every platform are neither realistic nor always desirable. Branches genuinely differ. The aim is coherent differentiation: stable names, explicit location markers, accurate local details, and clear relationships to the parent organisation.
A group page can help by stating which facts apply everywhere and linking to location pages for what varies. Branch pages can identify the parent brand while preserving their own addresses, services, hours, and contact routes. Maps, directories, and booking platforms may still diverge, but clearer source boundaries can reduce ambiguity in the public record.
The laboratory does not treat these steps as guaranteed control over AI outputs. Public records are copied, cached, reformatted, and interpreted through systems whose internal processes remain partly hidden. The point is diagnostic. When a branch error occurs, the organisation needs to know whether the problem resembles parent collapse, sibling transfer, a false branch, or a claim-source mismatch within an otherwise correct identification.
Limits and unresolved explanations
The repeatable runs could not reveal private retrieval infrastructure, hidden ranking logic, undisclosed intermediate steps, or every source used by a model. Visible citations were examined as evidence available in the observation, not assumed to be a complete account of generation. The laboratory therefore reconstructed apparent retrieval paths and kept that qualifier intact.
The composite restaurant group was selected because its branch structure exposed meaningful contrasts. It included active and former locations, Thai and English labels, platform differences, and a similarly named venue. The study cannot establish how often branch confusion occurs across Thai businesses or which sector is most affected. No rates or thresholds were inferred from the case.
Some outputs supported more than one explanation. Parent-level wording beside a correct branch address could reflect retrieval from the group site, ordinary summarisation, or a source not shown. When the record did not distinguish those routes, the laboratory left the cause open. A fluent answer does not justify a precise causal story.
Cross-model agreement remained observational. Several systems repeated the same former branch, apparently compatible with a widely copied directory record. That repetition showed persistence across the preserved runs. It did not convert the old location into a current one.
A provisional prediction follows from the material: branch errors should be more likely where location markers are inconsistent and operating facts are stored mainly on parent-level or third-party pages. The prediction would weaken if repeated inquiries preserved branch identity despite those conditions, or if errors appeared just as often in clearly separated records.
The durable conclusion is narrower. A correct parent brand does not establish a correct branch. Branch identity becomes credible only when the location label, address, local operating details, and supporting sources remain attached to the same place.