An outdated listing does more than preserve an old fact. It preserves an earlier version of the business, ready to be mistaken for the current one when dates, branch labels, and later corrections are weak.
In a composite scenario, the answer sent diners to a restaurant branch that had already moved. It gave the former street, described an outdoor seating area shown in old photographs, and cited a directory page that still marked the venue as open. The telephone number belonged to the current branch. The business identity had survived the move; the location record had not moved with it.
The scenario draws on recurring patterns around the regional restaurant group in the laboratory’s study objects. The mismatch is not perfectly clean. One social profile still mentioned the former neighbourhood in an old image caption, while a booking page showed the new address but retained the previous branch name in its page title. The generated answer had several routes back to the past.
Old information is not always visibly old
A listing can be outdated without looking abandoned. Its page may load normally, carry recent platform branding, display current navigation, and appear beside fresh search results. The stale part may be one field buried inside an otherwise active site.
Addresses are the obvious example. Former services, operating hours, names, branch relationships, ownership details, and business categories can persist in the same way. A restaurant may stop offering breakfast while old booking descriptions continue to advertise it. A clinic may change its English name, yet directory pages retain the earlier transliteration. A branch may close while reviews and photographs keep the page active.
An outdated business claim is a formerly accurate or once-published detail presented as current because the answer does not preserve the time boundary attached to its source. The definition includes two possibilities. The detail may genuinely have been correct in an earlier period, or it may merely have appeared in an older listing without ever being reliable. The preserved record often cannot establish which.
Date visibility matters. Some pages show a publication date but no update history. Map listings may change individual fields without exposing when each change occurred. Directory records can be copied between sites, causing one old description to appear under several current-looking domains. Social pages mix years of posts in one searchable record.
The result is a peculiar form of freshness. The page is available now, the business still exists, and much of the information remains correct. Only one piece belongs to a previous version of the entity. That is enough for a generated answer to build a description that feels current.
The laboratory avoids treating every disagreement as evidence that the source is old. The business itself may have inconsistent records, recently reversed a change, or maintained multiple locations. A later source is not automatically the more accurate one. The task is to preserve the disagreement and identify what each page actually supports.
How the past stays attached to the entity
Old listings remain useful to retrieval because they often contain strong identity signals. They may use the exact business name found in the prompt, provide a full address, include photographs, and repeat the relevant service category several times. A newer first-party page can be more accurate while offering less text for a system to connect.
The composite restaurant group had changed one branch’s address and shortened its English name. The old directory page retained the longer name used by customers in earlier reviews. A prompt using that familiar name could therefore align more easily with the outdated page than with the current branch page.
Branch changes are particularly vulnerable. Platforms may treat a moved branch as the same entity, a new entity, or a closed listing followed by a replacement. The business may see continuity where the platforms see two records. Alternatively, a platform may merge two branches that the company regards as distinct.
An old listing can also borrow freshness from surrounding material. New reviews may appear on a page whose address has not been corrected. A directory may update its footer and page design while leaving the business description untouched. Search snippets can surface recent wording from the site beside an old address extracted from the record.
Generated prose removes these seams. It does not normally say that the phone number came from one page, the address from another, and the service description from a third. The pieces are rendered as one present-tense account.
The answer may even use explicit recency language without visible evidence. Phrases such as “currently located,” “now offers,” or “is open daily” can be added during composition. These words are claims of temporal status. They require support just as the underlying address or service does.
A page retrieved in the present is not necessarily evidence about the present state of the business. That distinction is the centre of the old-listing problem.
Classifying the old claim against the visible source
The Four Source Relationships typology helps separate several patterns that can otherwise be collapsed into “outdated information.”
Direct support applies when the visible source supports the claim and its current status. A maintained branch page that clearly gives the present address can directly support a current-location statement. The investigator still records the observation date because the page may change later.
Stretched support occurs when the source contains the detail but does not establish that it remains current. An undated directory page may list an address, while the answer states that the business is “currently located” there. The source supports the existence of the address in the record. It does not support the added temporal claim.
Borrowed identity occurs when an old detail belongs to another branch or entity. The restaurant group’s former province branch might remain online and supply its opening hours to an answer about the Bangkok branch. The information can be current for one location while outdated or irrelevant for the one described.
Unsupported arrival occurs when no visible source supports the old detail. The answer may state a former service or address that appears nowhere in the preserved citations. An unseen internal source may exist, but the visible observation cannot carry the claim.
These classifications can shift as the claim is phrased. “A directory lists the restaurant at the former address” may have direct support. “The restaurant is currently located at the former address” may have only stretched support. The underlying page is identical; the evidential demand changes with the sentence.
This is why the laboratory reads time words closely. “Previously,” “formerly,” “since,” “now,” and “currently” are not decorative. They determine whether a source has been represented faithfully. Removing a date from an old statement can change historical support into a false present-tense claim.
The same source can therefore contain useful history and poor current guidance. Calling the whole page reliable or unreliable would lose the distinction.
Repeated runs reveal persistence, not age
One generated answer containing an old address is an observation. Renewed runs can show whether the old version returns, disappears, or competes with the current one. They cannot, by themselves, reveal why the source was selected.
The laboratory preserves the prompt, generated wording, visible citations, language, model context, observation date, and relevant run conditions. A repeatable run keeps enough of that procedure stable to compare the pattern. The wording may vary. One answer might state the former address directly, another might describe the old neighbourhood, and a third may recommend the correct branch while citing the obsolete listing.
These are related but distinct observations. The team looks for the underlying temporal relationship. Does the system continue to present the old record as current? Does it recognise the move but retain an old service description? Does the entity split into two apparent businesses?
Language can alter which version surfaces. The earlier branch name may remain common in English reviews, while the current Thai page uses a revised name. An English prompt can therefore retrieve the historical identity more readily. The opposite can happen when an old Thai listing persists after an English rebrand.
Cross-model agreement remains weak evidence of current truth. Several systems may reproduce the same old address because the listing is widely copied or strongly linked. Shared error can reveal the durability of the public record without validating the claim.
The observation date is essential here. A generated answer captured in 2025 cannot establish what a system would return in 2026. Likewise, a page viewed during later analysis may have changed since the answer was preserved. The team records the page state available during investigation and notes when the original state cannot be recovered.
Archived screenshots supplied by a business can help, as can saved answer captures and dated page exports. They still require interpretation. A screenshot proves what was visible at one moment. It does not show how broadly that version was indexed or whether the model retrieved it.
What correction work should target
Businesses often respond to an outdated answer by updating the most obvious page. That is useful where the old claim comes from a page they control. The harder cases involve copies, abandoned profiles, platform records, and listings that cannot be fully removed.
The first task is to identify which version of the business the answer found. A former address may be attached to the current entity, or the answer may have selected a dormant branch record from the beginning. Those situations call for different correction work.
Where identity continuity is clear, current pages should state the change plainly. A branch page can name the former location, the current location, and the date or sequence of the move where appropriate. This gives retrieval systems a relationship rather than two competing addresses.
Where the old and new records appear as separate entities, stable naming becomes more important. Branch labels, locality terms, contact details, and first-party links should make the relationship visible. A vague current page may lose to a richly described old listing even when the current page is accurate.
Businesses should also inspect descriptive claims, not only contact fields. Old menu items, service categories, photographs, and operating details can shape recommendations after the address has been corrected. A venue may be recommended for a feature that belonged to its former premises.
No correction guarantees a changed generated answer. The laboratory does not control indexing schedules, model retrieval, or ranking systems. A better result after an edit is a later observation under changed public conditions. It is not proof that the edit alone caused the improvement.
The research value lies in narrowing the mismatch. If the cited directory page still gives the old address, the source relationship is visible. If all visible sources show the new address while the old one continues to appear, the investigation moves toward unsupported arrival or an undisclosed retrieval path. The uncertainty becomes more specific.
Limits of temporal reconstruction
The laboratory cannot inspect every version of every page considered by an AI system. A source visible during analysis may differ from the version indexed when the answer was produced. Cached pages, snippets, structured fields, and platform databases can update at different times.
Private retrieval infrastructure and hidden ranking logic remain unavailable. The team can reconstruct an apparent path from visible sources and disagreements, but it cannot describe the complete internal route. Even a citation to an outdated page does not prove that the page caused the old claim. It shows a compatible source relationship.
The age of a listing may also be unclear. A page can lack dates, or its displayed update date may refer to the template rather than the business record. When no reliable temporal marker exists, the laboratory describes the page as conflicting with better-supported current material rather than assigning an invented age.
Composite scenarios expose mechanisms without measuring prevalence. The restaurant case shows how an old address, current telephone number, and revised branch name can merge. It does not establish how often that combination occurs across Thai businesses.
A provisional prediction is possible when conditions remain stable. If the obsolete page continues to rank visibly for the old branch name while current pages use a different transliteration, the old identity may continue to surface in renewed answers. That expectation would weaken if later runs consistently retrieve the current branch and preserve the move accurately.
The final boundary is simple. Retrieval in the present does not convert an old record into current evidence. The time attached to the claim must survive the journey from source to answer; otherwise, the system may describe a business that still exists only in fragments.