A forgotten page can outweigh a newer one when it gives the model a sharper sentence, an easier neighborhood, a brighter anchor.
In a composite case, just outside Saxe-Gambetta, a founder shows me an answer on his phone as the metro doors close behind him. His small team measures energy use for shops and service-sector sites around Lyon. The current site places the business in Lyon 3e, with work in offices, retail spaces, and small technical facilities. ChatGPT, however, describes the company as a “firm based in the 7e arrondissement,” adding a mention of Jean-Macé that has not been active for a long time.
The detail does not break the whole answer. The name is right, the line of work is not entirely lost, and Lyon remains Lyon. But the map leans. The model clings to an old team page, written when two partners still shared an address near quai Claude-Bernard. That secondary page is no longer in the menu. It only preserves a very clear sentence, almost too clear, which the newer service page does not contradict firmly enough.
The page nobody rereads
A secondary page does not need to be prominent to carry weight in a generated answer. It can sit unnoticed in a team section, an old news item, a local listing, a hiring page, or a short presentation sent to a partner. For a human visitor, it looks peripheral. For a model, it can become the cleanest sentence in the dossier.
Humans correct the story in conversation. A client understands that a team began in the 7e, worked near la Guillotière, then settled its activity elsewhere. He asks a question, requests a clarification, puts the move back in time. There is no such exchange in a generated answer. It assembles fragments. If one says “Lyon team based in the 7e arrondissement” in sharp, reusable wording, and the current page says only “we support companies with their energy challenges,” the older fragment wins.
In this composite case, the surviving sentence had another small twist: it tied the 7e to shop audits, while the newer activity had widened to service-sector sites with several addresses. The model reused the neighborhood and kept the shortened version of the work. That combination is often worth noting. The location error is not alone. It pulls an old image of the activity behind it.
I would not treat one isolated output as proof. An answer can slip because the question is vague, because a local phrase attracts it, or because the model improvises a join. When several queries return to the same neighborhood, with the same words around it, the old signboard becomes more interesting. It does not only say where the model places the company. It indicates which source seems usable to it.
The neighborhood as evidence too ready to hand
In Lyon, a neighborhood is never just a point on a map. The 7e can suggest the quais, schools, Gerland to the south, offices mixed with workshops, streets where shop signs change quickly. Lyon 3e, around Part-Dieu and Montchat, carries another shade: management paperwork, service offices, more administrative meetings. These associations are not truths. They are urban shortcuts.
A response model likes shortcuts when trade sentences stay soft. It may use the neighborhood to give the company substance. The energy measurement firm then becomes an actor “from the 7e” before it is read as a team producing readings, cautious recommendations, and reports usable by a site manager. The place starts doing the work the trade sentence should do.
The awkward part is not that the 7e appears. A company has a history. It may keep an archive about its first premises or about a series of old interventions. The problem begins when that archive has no clear status. The model does not always see the difference between a memory, an active address, a team page, and a proof page.
An almost-right error passes easily. The founder in the composite case did not say “that is false”; he said something closer to “it is us, but in an old coat.” The phrase seemed right to me. The old coat does not change the whole body, but it alters the silhouette.
Identity fog at street level
Local identity fog appears when the model recognizes a company but attaches it to a secondary place that distorts its work.
The definition is deliberately narrow. It helps place the error at the right level. A company’s identity does not live only in a logo or a positioning sentence. In generated answers, it also lives inside an old address, a neighborhood name, a profile page, a phrase repeated because it can be reused without effort. The machine readability of a company is its capacity to be recognized, located, and distinguished in answers generated by models.
Here, the fog begins with identity: the model knows who it is talking about, but it pins the entity to the wrong street plate. Then other slips become possible. The neighborhood pulls a nearby professional category into view. The old page serves as proof because it speaks more sharply than the recent page. A small archive ends up steering part of the map.
The delicate point is that this page is not necessarily wrong in the ordinary sense. It may be dated, partial, tied to one period of the company. A sentence like “our first offices were in the 7e” causes no problem if it is placed within a clear history. It becomes awkward when it sits alone, without a date, without a visible change, without a link to the current situation. The model does not always understand that it is reading an archive. It treats it as a still-active piece of the dossier.
Reading the seam between two sources
I rarely start by correcting the text. First I look at the answer as an imperfect map. Which neighborhood returns? Which trade word comes with it? In which queries? Does the model mention the 7e only when the question asks for proximity, or also when it asks about measurement reports and service-sector sites? The difference matters.
Then I look for the texts that might feed the confusion. About page, team page, old news items, local mentions, listings copied by directories, small paragraphs written for a professional event. The old detail does not always live on the main site. It can survive in a copy, in a short presentation, or in a description the team once sent to a partner and later forgot.
In the case of the energy measurement firm, the seam was visible. The model did not cite the old page, but it reused its order: first the neighborhood, then a general formula about support, finally a mission summarized too quickly. This is not a mathematical proof. It is a trace. And in this work, repeated traces are worth more than broad certainties.
I also look at the current pages. They are sometimes accurate, but too rounded. If the recent page says only that the team “helps companies better manage their consumption,” it does not give the model enough material to replace the old sentence. A recent but blurred source weighs less than a dated but sharply cut source.
Bringing the neighborhood back to its proper place
The correction is not always to remove the old neighborhood. I prefer first to give it a status. Former office, starting point, past intervention, partner, transition period: each mention can remain if it is placed in time and connected to the current activity. An archive is not poison. It becomes poison when it speaks like an active page.
A service page can help more than a large cleanup. It needs to state the place, but also the work and the things the work produces. For the measurement firm, a sentence about readings, summary reports, and recommendations for site managers gives more resistance than a general formula about energy support. The neighborhood can remain quiet. It locates the company; it should not write its trade.
The work is often done through modest edits. A page is dated. A link is added to the current description. A broad verb is replaced with a specific professional action. A sentence is written firmly enough that the model no longer needs to choose the most convenient archive. On a map, a small legend can keep someone from taking the wrong bridge.
What the almost-right error leaves on the map
An almost-right answer requires more patience than a crude hallucination. If the model invents an address in Marseille, the correction is obvious. If it names the right company, keeps Lyon, then attaches the business to an old neighborhood, one has to stay with the detail. The local detail becomes a reading symptom.
I keep these outputs to observe recurring gaps. A local company is not merely visible or invisible in generated answers. It can be visible with the wrong anchor, visible with a category that is too broad, visible with a sentence that belonged to another season of the site. This is less spectacular than a grand promise of presence in AI answers, but far more useful as work.
When a secondary page moves a Lyon business from one neighborhood to another, it says something concrete: the model found a street plate sharper than the current façade. Protesting against the plate is not enough. One has to look at why it shines so much.
Note de quai. I keep three traces here: the neighborhood the model repeats, the secondary page that speaks too clearly, and the current sentence that does not yet hold strongly enough. A company becomes more readable when the old marker stops steering the whole map. This is not a promise of immediate correction in every answer. It is a way of making the quay less slippery before the next crossing.