Recommendation errors rarely arrive wearing an obviously wrong name. More often, the answer selects a plausible venue and quietly furnishes it with another business’s address, menu, branch history, or reputation.
A composite restaurant scenario used by Khlong Trace Laboratory began with a recommendation for a family dinner outside central Bangkok. The answer named a real restaurant from a regional group, described a riverside setting, and cited a booking page with photographs that appeared to fit. The problem surfaced in the travel time. The stated province belonged to another branch, while the photographs and signature dish came from a similarly named independent venue.
A renewed run produced a different mixture. The correct branch address returned, but the answer now called the restaurant part of a hotel and praised a view that existed only at the independent venue. One detail remained oddly accurate in both outputs: the group’s closing time on weekdays. The answer was neither wholly wrong nor cleanly attributable to one source. It had assembled a convincing place from several nearby identities.
Why recommendations are especially vulnerable
A direct lookup usually starts with a name. A recommendation starts with a bundle of attributes: cuisine, neighbourhood, atmosphere, price, suitability, opening hours, perhaps parking or a view. The system must find entities and then decide which one best fits the requested combination. Shared or similar names create trouble because the evidence can be distributed across several records that all look relevant to the same request.
The composite restaurant group used a common English rendering of its Thai name. Branch pages added location words inconsistently. One booking platform placed the district before the brand; a map listing placed it after. Social pages shortened the name. The independent venue used almost the same English words, served a related cuisine, and appeared in the same broad discovery category. To a reader scanning quickly, the businesses looked connected even though they were not.
Recommendation prose makes the blend harder to detect. Lists often compress several claims into one polished sentence: “A popular riverside branch known for seafood, relaxed family dining, and sunset views.” Each element may come from a different source relationship. The name may identify the group, the address may identify one branch, the riverside setting may belong to another venue, and “popular” may have no visible support at all.
The laboratory uses “mismatched recommendation” for an observed answer that names one entity while drawing part of its recommendation rationale from attributes belonging to another. The term describes the final claim pattern. It does not establish where in the system the identities were joined or whether one particular retrieval step caused the mismatch.
The laboratory is careful not to infer intention from this pattern. The answer does not “decide” to merge businesses in the human sense. What can be observed is a final description whose parts fail to remain attached to one entity.
Reconstructing the recommendation one claim at a time
The team preserved the complete recommendation answer rather than extracting only the suspicious sentence. Context mattered. A venue recommended for “quiet riverside dining” could be selected differently from the same venue requested by exact name. The prompt, language, model context, visible citations, observation date, and run conditions formed the observation.
Each recommendation was then broken into claims. The researchers separated identity, branch, address, province, cuisine, setting, operating details, ownership, and evaluative language. This felt fussy at first. It was also the only way to see that some clauses belonged to the named restaurant while others did not.
The visible sources were read in the same granular way. A booking page could support photographs and a reservation option without establishing ownership. A map profile could support an address while using a broad category generated by the platform. A social page could show a current menu but leave branch status ambiguous. The laboratory did not allow one relevant page to certify the whole recommendation.
The team also ran matched variations of the discovery request. Some prompts emphasised location, others atmosphere, cuisine, family suitability, or comparison. The purpose was not to discover the “best” prompt. It was to observe whether different request wording pulled the answer toward different members of the shared-name cluster.
A rough pattern emerged. Location-heavy prompts more often stabilised the address but could still inherit reputation language from elsewhere. Atmosphere-heavy prompts were more likely to attract photographs and descriptive phrases associated with the independent venue. Comparison prompts sometimes widened the identity further by treating branches and unrelated venues as alternatives within one brand family. These were descriptive observations from the composite case, not measured frequencies.
One renewed inquiry resisted the pattern. It selected the correct branch, cited the group’s own page, and kept the setting accurate, yet listed a menu item that had been removed. That error appeared to come from stale information rather than shared-name confusion. The laboratory kept it in the record because a case becomes misleading when every imperfection is bent toward the main explanation.
Four ways a source can sit beside the wrong recommendation
The Four Source Relationships typology gave the analysis a stable vocabulary. Direct support occurred when a source supported the claim as stated for the named entity. The restaurant group’s current branch page, for example, directly supported the branch address and weekday closing time.
Stretched support appeared when the source supported a narrower claim than the answer made. A booking page describing “outdoor seating” did not, by itself, support “panoramic riverside sunset views.” The source was relevant, but the recommendation enlarged it. This expansion matters because atmosphere claims often carry the persuasive weight of a recommendation.
Borrowed identity described the central failure in the composite scenario. Photographs, location features, or reputation claims belonging to the independent venue were carried into the named group branch. The evidence itself was not fabricated. It had crossed an entity boundary.
Unsupported arrival covered claims for which no visible source in the observation supplied support. “One of the area’s most beloved restaurants” and “a top choice for families” appeared in this category when the displayed pages did not establish them. Such phrases may arise from hidden material, general language patterns, or other undisclosed inputs. The classification records only the visible relationship.
These four relationships are qualitative classifications, not confidence scores. A sentence can contain more than one relationship. “This popular riverside branch closes at 10 p.m.” may combine unsupported arrival for “popular,” borrowed identity for “riverside,” and direct support for the closing time. A single citation marker cannot express that structure.
The typology also prevents a common analytical mistake: declaring the entire recommendation unsupported because one clause fails. That response is too blunt. The useful finding is where support stops, where another identity enters, and which part of the answer remains sound.
How shared names become shared reputations
Names are only the first bridge between entities. Platforms add more. Similar categories, nearby geography, overlapping cuisine terms, copied descriptions, and photographs without clear branch labels can turn a name resemblance into a shared informational neighbourhood. Once records cluster together, reputation can travel along the same route as an address or menu item.
The restaurant group and independent venue both appeared under broad dining categories. Their English names differed by a small word that some listings omitted. User-contributed photographs were not always labelled by branch. A directory description referred to the group without naming a location, while a booking page used the independent venue’s full name but shortened it in page metadata. None of these records alone established the cause of the mismatch. Together they made the blended description compatible with the visible source environment.
Reputation language is particularly mobile because it is often weakly anchored. “Popular,” “well-known,” “family-friendly,” and “worth the trip” can be generated from reviews, directory summaries, travel prose, or other undisclosed material. When a shared-name cluster exists, the adjective may attach to whichever entity the answer has already selected.
The laboratory therefore treats recommendation identity as a chain rather than a label. The named entity must connect coherently to the location, branch, category, descriptive attributes, and sources. A break anywhere in that chain can change the decision a reader makes. Someone may travel to the wrong province, expect a riverside table that does not exist, or assume two unrelated businesses share ownership.
For business owners, this is why mention tracking has limited diagnostic value. The brand may appear in the answer and even receive praise, yet the praise can belong to a neighbour. Visibility looks positive while identity has already split.
What can be inferred, and what cannot
The material supports a modest conclusion: shared names and overlapping discovery attributes are compatible with mismatched recommendations. Repeated changes in the cited pages, branch location, and descriptive features strengthened the interpretation that the answer was moving among several entities. They did not reveal the model’s complete route.
Private retrieval infrastructure, hidden ranking logic, intermediate queries, and undisclosed sources remained inaccessible. A visible citation may reflect evidence used during generation or may be attached through a later process. The laboratory therefore reconstructed an apparent retrieval path from the record rather than claiming access to the system’s internals.
The composite scenario also concentrated several risk factors at once: similar names, related cuisine, inconsistent branch labels, nearby provinces, and platform-generated categories. Many businesses will have only one or two of these conditions. The case exposes a pattern for further testing; it does not establish how common that pattern is across Thailand.
Cross-model agreement offered no shortcut. When several systems repeated the same riverside error, the observation showed that the mismatch was reproducible across those runs. It did not confirm the riverside claim. Shared public sources can make several systems confidently wrong in the same direction.
A provisional prediction follows: recommendation prompts that emphasise attributes shared by similarly named entities are more likely to produce blended rationales. The prediction would weaken if matched runs repeatedly kept all claims attached to the correct business despite the shared-name source cluster.
The strongest conclusion remains claim-sized. A recommendation should be trusted only as far as its identity and source relationships hold together. The correct-looking name is the beginning of that check, not the end.
What a reader can verify before acting
The first check is geographical. Does the cited page belong to the stated district, city, or province? The second is organisational. Is the page for the named business, a branch, a parent group, a hotel tenant, or an unrelated venue with a similar name? Only then does it make sense to assess the recommendation language.
Descriptive claims deserve the same discipline. Photographs should be tied to the correct location. Menu items should be current and branch-specific where branches differ. Ownership language should come from a source that actually establishes the relationship. Praise should be treated as a claim, not as decorative prose that requires no evidence.
This verification does not produce a single visibility score. It produces a more useful record: which parts of the recommendation directly support the named business, which parts stretch available evidence, which parts borrow another identity, and which arrive without visible support.
The composite case looked favourable at a glance. The restaurant was recommended, described warmly, and accompanied by citations. Once the claims were separated, the answer no longer described one place. It described a stitched venue, convincing at sentence level and unstable at the level where a customer would actually choose where to go.