A missing fact is an empty hook. A wrong fact is a hook already holding a coat, a receipt and someone’s old keys. The model may notice the new hook and still reach for the crowded one.
In one composite local-service case, the business added a new emergency-repair service to its website. The next recorded answer did not mention it. A later answer did, but cautiously, after the model had already described the company in its older category. In a second composite case, the company corrected an old city reference that had spread through a directory and a regional mention. The model kept the wrong city for longer.
That asymmetry is the heart of this material. Adding a fact and fixing a wrong fact look similar from a content-management screen. Both involve editing public text. But in model answers, they behave differently. One asks the system to include something that was absent. The other asks it to stop repeating something already made plausible by the public trail.
A missing fact and a wrong fact leave different tracks
Levertrace Lab separates missing facts from wrong facts because the two changes produce different kinds of evidence. A missing fact is a current business detail that was not clearly present in the public trail before the lever was changed. A wrong fact is a stale or inaccurate detail already present somewhere in that trail, often repeated across pages the business does not fully control.
Fixing a wrong fact — in this material — means changing the public evidence trail so an outdated or inaccurate business detail loses support, because the model may still find older confirmation elsewhere.
That definition is deliberately heavy. A correction is not complete when the official page changes. It is complete only as a source condition when the old fact stops being easy to recover, or at least becomes visibly weaker than the corrected version. The model may be asked a simple question, but the public record behind the answer has memory scars.
A missing fact tends to behave like an addition. The company now serves a new city. It now offers a compliance audit. It now supports a specific sector. If the fact is placed clearly on an owned page and repeated in a sensible context, the model has a new candidate phrase. The answer may adopt it, partially echo it or ignore it, but there is no older phrase competing for the same slot unless the category already implied the opposite.
A wrong fact is more stubborn because it has a history. It may appear in an old page title, a directory entry, a partner summary, an event bio, a review profile, a cached snippet or a local article. Even when the official site changes, the old fact can still feel confirmed. The model is not choosing between empty and full. It is choosing between two full-looking trails.
That difference is why “we corrected our website” so often feels unsatisfying. The website may now be right. The answer can still be wrong.
Adding a fact gives the model a new phrase to carry
A composite Object B software company adds a new capability claim to its English landing page: the product now supports audit logs for regulated retail networks. The French documentation already hints at audit trails, but no public page had stated the business-facing version plainly. After the edit, a model answer under one prompt family begins to say the company provides “workflow tools with audit-log support.” Another answer keeps the older category and omits the new feature.
The lab would not call that full adoption. It is partial echo. The new fact appears, but it arrives attached to the old summary frame.
This is a common pattern with added facts. The model may incorporate the addition as a subordinate clause, not as a new definition of the business. A French building-maintenance company adds “contrôles de conformité des chaufferies” to its service page. The answer still calls it a maintenance provider, then adds that it may handle compliance checks. The new fact has entered the description. It has not taken over the category.
For many businesses, that is enough. Missing facts often matter because they qualify the lead. A model answer that mentions emergency repair, bilingual support or public-sector procurement experience may direct a reader differently, even if the high-level category stays the same.
The lab notices one practical condition: added facts move more cleanly when they sit near the existing category. If a company already described itself as a maintenance provider, adding a maintenance-related compliance check gives the model a small step. If the same company suddenly adds a separate consulting service without changing the surrounding explanation, the answer may treat it as an odd extra, or skip it.
The new fact also needs language discipline. A fact added in French may not appear in English answers unless the English source trail gives the model a way to name it. This material does not make the language-transfer question its main subject, but the issue keeps showing up like a receipt stuck to another page. A missing fact can be present and still not travel.
Correcting a wrong fact means weakening the old version
A wrong fact has a different texture. Suppose a composite Object A company has moved from general building maintenance into specialised heating-system maintenance. Its own site is corrected. A directory listing is updated. But an older regional mention still calls it a general maintenance firm, and a review profile has categories that point in the same direction. In the next answer, the model says the company “mainly provides general maintenance, including heating-related work.”
That sentence is not exactly old, and not exactly new. It is a compromise made of public leftovers.
The lab classifies this as source conflict when the old and corrected facts compete in the answer. Sometimes the conflict is visible: the answer says sources describe the company differently. More often it is hidden inside a blended sentence. The danger is that the blend sounds plausible. It may even sound more thoughtful than a plain correction. But for the business, the old fact is still steering the reader.
Removing or fixing a wrong fact requires more than publishing the correct one. The old claim has to lose repetition, lose prominence or become surrounded by stronger current evidence. If a directory still carries the old city, a model may keep that city as a safe local signal. If a trade profile still uses the former category, the model may keep the category and tuck the new activity underneath it. If several weak sources agree on the wrong fact, one official correction may not be enough to shift the answer.
Here the lab’s position is slightly uncomfortable. It does not tell businesses that official truth automatically prevails. Public truth, in an answer system, often looks like repeated recoverable wording. That is not a moral claim. It is an observation about the shape of evidence.
The correction also has to be clear about the replacement. A page that quietly removes “Lyon” but never says the business is now based in Nantes may leave the model with less to use. A correction that says “formerly based in Lyon; now operating from Nantes” may help a human, but it can also preserve the old city in the text. There is no universal rule here. The phrasing has to match the risk: is the model forgetting the new fact, or clinging to the old one?
The anchor: adoption, partial echo, source conflict, no visible movement
Levertrace Lab applies its qualitative anchor to both additions and corrections: four lever outcomes in LLM business change — adoption, partial echo, source conflict, or no visible movement. The anchor keeps the lab from treating every imperfect answer as the same kind of failure.
Adoption, for an added fact, means the model includes the new detail in a way that matches the source context without dragging in an incompatible old frame. Adoption, for a corrected fact, is stricter. The old fact has to disappear from the answer or be accurately marked as historical. A company formerly described as based in one city should not still be presented there unless the question asks about history.
Partial echo is common after additions. The new service, sector or capability appears, but weakly. The answer says the company “may offer” a service that the site states directly. It mentions the new activity under a broader category. It translates a precise French term into a softer English one. The fact has moved into view, but it has not become stable wording.
Source conflict is the signature result for wrong facts. The answer mixes the old and corrected versions, or alternates across runs. In one run, the company is a local provider in the old city. In another, it is a regional specialist with the new city mentioned later. In a third, the model avoids the location altogether. The model is not simply wrong in the same way each time; it is negotiating incompatible traces.
No visible movement must stay on the table. The addition may not appear. The correction may not appear. The answer may remain unchanged under the recorded prompt family. Levertrace Lab does not read that as proof that the lever failed everywhere. It records that the tested condition did not show movement.
Corrections are harder to interpret because silence can mean delay, conflict, weak evidence or a prompt that never pulled the fact into view.
The anchor is especially helpful because it prevents a false binary. A model answer can be less wrong without being right. It can adopt a new fact while still carrying an old category. It can stop naming the wrong city and still imply the wrong market. These intermediate states are where most business-description work actually sits.
Why wrong facts persist in business descriptions
Wrong facts persist because they are often useful to the model. A city, category, founder name, activity or market label helps build a coherent answer. If the fact is old but well distributed, it can seem safer than a newer correction that appears in one place. The model may not know it is choosing age over accuracy. It is producing a plausible summary under the conditions available.
The lab sees several forms of persistence. One is category inertia. The company’s official site now says it provides a narrower service, but older sources label it broadly. The answer keeps the broad category because it is easier to understand and more common in public text.
Another is location inertia. Local directories, maps-style profiles and regional mentions can keep a city attached to a business after the company has moved or expanded. A model answer may use the old city to make the business feel concrete, even when the current site uses a wider service area.
A third is activity residue. A company that once did web design and now sells software may still be described as an agency, especially if old portfolio pages, event bios or partner notes remain visible. The answer may say “software agency” as a compromise, a phrase that seems to reconcile the old and new evidence but does not match the business’s current positioning.
There is also translation residue. A French phrase may be corrected, but an older English summary may continue to circulate with a loose or outdated translation. The answer in English then preserves the wrong fact while the French answer improves. This is not the focus of this item, but it explains why correction work can feel especially uneven in France’s bilingual business context.
For a business owner, the lesson is not to hunt every old mention in panic. Some stale sources may have little visible influence. The lab’s method is narrower: record the before answer, identify the wrong fact, correct one lever where possible, record the after answer, then inspect which public sources still support the wrong version. The aim is not total internet hygiene. It is attribution modest enough to be useful.
Limits of the correction test
This material cannot prove that wrong facts are universally harder to fix than missing facts are to add. Levertrace Lab does not claim a measured law. It records a repeated pattern in controlled comparisons: additions often behave like new optional details, while corrections require the answer to abandon or reframe existing evidence.
The method is also exposed to delays. A corrected source may not be available to a model under the tested condition. A browsing interface may retrieve one source set, while a non-browsing answer leans on older learned associations. A prompt may not ask for the fact being tested. A model may avoid the business altogether, leaving the lab with omission rather than a clean wrong-or-right comparison.
Another limit is visibility. When the interface does not reveal live search, cited source use or retrieval path, the lab marks the source condition as uncertain. A phrase that resembles an old directory may have come from that directory, from another copied source or from prior exposure. The lab can compare public trails and answer wording. It cannot enter the model and watch the decision form.
For that reason, the conclusion stays cautious. Adding a missing fact appears easier when it offers the model a clear new phrase that fits the existing description. Fixing a wrong fact appears harder when the old version remains repeated, compact and useful. The work is less like changing a sign on a door, more like changing a name people have already written into several notebooks.