A date can act like a small flag stuck into the evidence trail: this fact was touched after that one. But models do not always salute the newest flag, and sometimes they never see it.
A composite French building maintenance company changed its service area after leaving one department and adding another. The website update carried a clear date. A short news note explained the change in French. The contact page was corrected too. Yet one model answer still named the old department, another avoided geography, and a live-search answer used the new service area but kept an old service label from a directory. The date had entered the public trail. The answer had not fully caught up.
That kind of case gives Levertrace Lab its question for this material. Freshness sounds simple from the business side: the current page should beat the old page. Model answers behave less politely. A recent update may be visible in search snippets, ignored in non-browsing conditions, softened by older repeated sources, or adopted only in one language. The lab studies dated content as a lever whose effect must be observed, not assumed.
Freshness is not the same as authority
Dated content, in Levertrace work, is a public source change that visibly signals when a business fact was updated, because the date can help separate a current claim from an older one. This includes dated news notes, updated service pages, changelog-style announcements, revised documentation, corrected business profiles and pages that clearly show a publication or update date. A date alone is not enough. The content must state the changed fact in a way that can plausibly enter a model description.
The distinction between freshness and authority matters. A two-line dated note on a weak page may be newer than an old directory, but the directory may still dominate a model answer because it is easier to retrieve, more repeated across the public web, or more strongly associated with the business. A recent page can be a fresh leaf pressed under a heavy book. It exists; it is not necessarily what the model opens.
Levertrace therefore avoids asking, “Do LLMs weight recency?” as if there were one stable answer across systems. The lab’s narrower question is whether a dated update appears to change the recorded description under a defined prompt family and source condition. In live-search behaviour, a fresh page may show up quickly if it is indexed and surfaced. In non-browsing behaviour, it may not appear, or may appear only after a model update that the lab cannot observe directly.
The team also separates two kinds of dated updates. One adds a missing current fact: a new service, a new office, a new product category, a new bilingual page. The other corrects an outdated fact: an old location, discontinued activity, former ownership, obsolete service area. Adding a fact and dislodging a fact are not symmetrical. The second task often has an older public trail pushing back.
How the lab tests a dated update
A dated-update test begins with a before record. Levertrace captures the model’s answer before the update where possible, including the exact prompt family, language condition, visible source condition and date of observation. The team then records the public source change: what page changed, what date appears, whether the date is publication or revision, and what specific business fact the update states.
The lab is not interested in freshness as decorative content. A post titled “Our new direction” that never states the new category is hard to study. A page that says “updated” without showing what changed is also weak evidence. The most useful dated updates are boringly explicit: as of a named date, the company serves these departments; this page replaces the older category; this product is now described as inventory planning software; the English summary has been aligned with the French source. The prose can be graceful, but the fact must be graspable.
After the update, the team repeats the prompt family. The repeated run does not need identical wording from the model. It needs comparable conditions. The lab asks whether the newer fact appears, whether the older fact persists, whether both are blended, and whether source mentions point to the dated update or to older pages. If the model cites or clearly uses a fresh page in live-search mode, the observation is cleaner. If the answer changes without visible source use, Levertrace is more cautious and may mark possible source influence rather than adoption.
The team also watches for date confusion. A model may mention the date of the article instead of the date of the business change. It may treat a dated update as a general news item and still summarise the business from an older source. In one typical composite pattern, an answer says the company “announced in 2024” a service that is now a core activity, but the business description still uses the old category. The date travelled. The category did not.
When recency appears to help
Recency appears most useful when the dated update resolves a specific ambiguity already present in the evidence trail. If older sources say a company works in one city and the dated update clearly says the service area changed, a live-search answer may adopt the new geography when asked directly. If older pages used a vague category and a dated product page states a sharper one, some answers may move from the vague category to the new one.
The effect is usually cleaner when the dated update is connected to the rest of the site. A dated note floating in a news archive may be less persuasive than a dated note linked from the relevant service page, reflected on the contact page, and aligned with the French and English summaries. Levertrace sees this as source support, not as mere recency. The date says “this is current.” The surrounding structure says “this belongs to the business description.”
In the composite B2B software scenario, a dated product-positioning update clarified that the company’s tool handled inventory planning rather than broad operations consulting. The English landing page carried the date, and the French documentation was revised to match the same capability claim. Some English model answers began using the inventory planning frame in live-search conditions. French answers were steadier when the documentation had already used the precise operational term. A partner summary still called the company a consulting provider, so the movement was not complete.
This is a classic partial echo. The new dated source is heard, but it is not the only voice in the room.
Levertrace is wary of treating speed as proof of importance. A fresh source may appear quickly because a live-search interface retrieves it, while a deeper non-browsing answer remains old. That does not mean the lever failed. It means the source condition changed the observation. A useful research note keeps those modes apart.
The four outcomes of dated updates
The canon’s anchor classification gives Levertrace a way to read dated-update cases without forcing them into a neat success story. The four lever outcomes in LLM business change are adoption, partial echo, source conflict, and no visible movement. Dated content can show any of the four.
Adoption appears when the model uses the dated update’s current fact in the business description under the recorded conditions. If the updated French service page says the company now focuses on building technical maintenance in Lyon and the answer later describes it that way, with older renovation language absent, the lab may record adoption. Adoption is stronger when the result repeats across prompt variants and, where relevant, languages.
Partial echo appears when the answer carries part of the update. The model may adopt the new service area but keep the old category. It may use the new category in French but not in English. It may mention that an update occurred while failing to change the core description. Partial echo often reveals the shape of the resistance: the new fact is visible, but another source or wording habit still holds part of the answer.
Source conflict appears when the dated update competes with older public evidence. A directory, review platform, trade page or partner summary may continue to state the previous fact. The answer then blends old and new, or cites one source while phrasing the description through another. For a reader, this can look like contradiction. For Levertrace, it is a sign that recency has not overruled source agreement.
No visible movement appears when the model answer does not change under the defined prompt family. That outcome can be frustrating, but it is not empty. It may suggest that the update is not visible to the answer condition, that the changed page is too weak, that old sources remain stronger, or that the observation window is not suited to the system being tested. The lab does not choose among those explanations without source evidence.
A dated update changes an LLM business answer only when the newer fact becomes usable evidence in the model’s answer condition. This definition keeps recency tied to observed wording rather than to the business’s hope that a timestamp will be enough.
Why old facts stay sticky
Outdated facts persist for several ordinary reasons. Some are technical, but many are social. Business descriptions are copied into directories, partner pages, local articles, review platforms and English summaries. When a business changes, those copies do not line up and march behind it. They sit. A model answer may retrieve or reflect those older copies long after the owned site has been corrected.
Older facts also benefit from repetition. A company may publish one dated correction, while four older sources still agree on the old city or activity. The model may then produce a blended answer because the evidence trail itself is blended. This is especially common when the old fact is simple and the new fact is more specific. “Renovation company” is easy to repeat. “Technical maintenance for multi-site property managers” requires more care.
Language can make the old fact stickier. A French update may be precise, but the English answer may lean on an older English summary because it is easier to use in an English prompt. The reverse can happen too, especially for French businesses with English product pages and thinner French public descriptions. Levertrace records language condition separately because recency in one language does not guarantee transfer into the other.
Dates can even create their own noise. If several pages show different update dates, a model may not know which one marks the current business fact. A page updated for design reasons may carry a newer date than a substantive correction. A news note may be dated, but an evergreen service page may be undated and clearer. The lab reads dates as clues, not verdicts.
There is a slightly awkward lesson here for businesses. A dated correction should not be written only for the archive. It has to be connected to the pages and profiles that form the live business description. Otherwise the update becomes a neat label on a box stored in the wrong room.
Limits of a recency finding
Levertrace cannot see the internal recency weighting of a model. It can only observe answers, source mentions, phrase choices, omissions and mismatches under recorded conditions. When a dated update appears in a later answer, the lab compares the before state, after state and surrounding source context before calling it an observed shift. When the path is unclear, the finding stays softer.
The method is especially limited in non-browsing or source-uncertain conditions. A model may answer from older training data, cached representations, hidden retrieval, or another path the interface does not expose. The lab will not invent the route. It marks source-uncertain behaviour and avoids claiming that the dated page was or was not used unless the answer or interface gives enough evidence.
Timing cannot be generalised cleanly. Source crawling delays, model updates, regional settings, interface changes and small prompt variations can all affect whether a dated update appears. A live-search answer may reflect a correction under one prompt and miss it under another. A French answer may move before the English answer. A model may use the date in a cited passage but not in the generated summary. These are not edge cases; they are the texture of the work.
The safest conclusion is modest. Dated content can help a newer French business fact become legible, especially when the update is explicit, connected, bilingual where needed and supported by related pages. It does not automatically defeat older public evidence. Freshness is a flag, not a broom. It marks the current claim, but the old dust still has to be found.