How a model glues a branch to an old category

Old listings do not always look old inside an AI-search answer. Sometimes they become a quiet layer between the company name and a description that no longer matches the business’s local work.

In one Field run for the query local company old listing, the lab examined a composite scenario assembled from several observations around a small bookkeeping firm in a Midwestern suburb. The firm worked with home-service contractors: HVAC, plumbing, roofing, cleaning. In the answer, the model kept the new name and even added a fresh suffix from the current storefront, but placed an old detail beside it: a former office near an industrial road and the phrase tax preparation, which had long stopped being the main service. The source in the citation trace confirmed only the address listing. It did not show that the firm now handled bookkeeping for contractors.

A related review produced another composite scenario, again assembled from several observations: a regional commercial roofing contractor in North Texas with a branch history across several counties. In one AI-search answer, the company was named almost correctly, but the description drifted toward general contractors, and the local anchor stretched back to an old suburban office. The crooked detail was small, almost clerical: the model gave the current phone number, then inserted a district where the team no longer kept an active office. The lab did not tie this record to a specific named company. For them, it was a field card about how an old record can outlive operational reality.

Where the old listing starts speaking for the company

An old listing rarely enters the answer as a character of its own. It appears more quietly: in an adjective, in a district note, in a neighboring service, in a small phrase after the name. The reader sees a familiar name and reads the answer as acceptable. But once the AI-search answer is taken apart, a seam appears. The name came from one layer, the regional label from another, the category from a third, while the source supports only the safest piece: the company’s existence, or an address that once appeared somewhere.

In the lab’s field notes, this sits closest to citation-description split. The source sits next to the description and creates the feeling of confirmation, although it actually holds only part of the answer. When the old listing comes from a broad directory, directory dependency enters the record: the model reaches for a more available or more familiar directory phrase, even when the company’s current site describes the work differently. If a former service is mixed into the answer, category drift appears. If an old branch pulls the company into a neighboring county or suburb, regional substitution sits nearby.

These classes are not fighting each other. They mark different points along the same crooked stitch. A working record can carry several labels, but the center of gravity remains the mismatch between source and description. The old listing becomes a sticky tag: someone tried to peel it off, but a gray rectangle of glue stayed on the box, and the model read it as part of the new label.

What the Field run preserved

The lab did not begin with the question of which model “sees” a local business better. The field card preserved more mundane details: query wording, Regional frame, model, answer mode, named companies, the description of each company, and the citation trace if one appeared. In the bookkeeping-firm series, the team watched whether the tie held between the name, the suburban frame, and the work for home-service contractors. In the commercial roofing series in North Texas, they checked whether the description slipped into general contractor language after an old branch detail appeared.

The old source behaved like a small magnet under a sheet of paper. On the surface was the current map of the business, but the line bent slightly where the old record lay beneath it. In one answer, the model held the service but added the former district. In another, it kept the district but widened the category. In a third, the source confirmed that the company existed, while the description stepped sideways. These differences did not become a numerical index. They gave the lab material for qualitative reading: where, exactly, the coupling weakened.

For the lab, this kind of review does not require a dramatic error. The model does not have to invent the company or move it to another state. It simply glues a current fragment to a stale one. That makes the answer look tidier than it is. The owner recognizes the name, the reviewer sees a source, and the category crack remains as thin as a hairline in glass.

The team also marked where the old detail enters the text. Sometimes it appears in the opening descriptive sentence, immediately after the name. Sometimes it hides at the end as a note about the service area. Sometimes the source itself contains no obvious error, but its language is too old for the current category. In that kind of record, the lab does not argue with every line of the source. It watches how the source begins to steer the description more than an address trace should be allowed to.

Subtypes inside citation-description split

An old listing shift is a form of citation-description split: the source supports existence or address, but smuggles a past branch, district, or service into the current description of a local business. The lab describes it as a set of subtypes inside citation-description split, not as a separate branch of the canon.

In the address shadow, the model names the company correctly, but the Regional frame leans on an old office or former suburban point. To a reader, this can look like a harmless clarification, although in a local service category the district changes the meaning of the answer. A category tail appears when the source stores an old service and the model mixes it into the current specialization. For a bookkeeping firm, that might be tax preparation or payroll software; for a commercial roofing contractor, it might be the broader language of general contractors. Branch glue gathers the old point, the old service, and the current brand into one smooth phrase. A directory backing appears when a broad directory listing sets the tone for the whole description; there, citation-description split already touches directory dependency.

The typology is not ornamental. It forces the researcher to ask what the source actually confirmed. Did it confirm the current service? The current office? The service region? Or only that the company once existed under a similar name? Without that question, the citation trace becomes a decorative paperclip, pinned to the text for reassurance.

Why an old branch pulls the category with it

In this series of runs, the lab did not make technical claims about the internal mechanics of models. The more careful interpretation is this: an AI-search answer can be assembled from traces of different freshness. Some traces sit closer to the company’s current site. Others live in directories, old listings, industry descriptions, and short third-party profiles. When the query is local, the model tries to hold several things at once: name, service, city frame, source. If one trace looks convenient enough, it can drag an extra detail along with it.

This is especially visible with companies that have branch history. A branch closes, the service area changes, the service is renamed, and the outside listing remains. It does not necessarily lie in a blunt way. It simply lags. A person can see that lag in details: an old address, an odd service phrase, a mismatched county. To the model, the listing may look like usable support because it contains a name, a region, and a category label. The former structure of the business then becomes part of the current answer.

There is another reason this shift is hard to catch on a first reading. The old listing is often right at the grain level and wrong once milled into flour. It confirms that the company was tied to this city, that the name is not random, that an industry tag sits nearby. The error happens in the milling: those grains are turned into a description of current work, even though the source cannot bear that load. That is why the lab checks the source not only for presence, but for granularity. An address trace supports the existence of the company; it does not yet support the category. A former branch does not define the current Regional frame either.

The composite scenario with commercial roofing in North Texas showed another uneven edge. When a company serves several counties, an old branch record can narrow it to one suburb or widen it into a neighboring city area. Category drift and regional substitution then start moving together. The model seems to be speaking about the same company, but its local geometry changes. The outline of the building is the same; the sign by the entrance has shifted.

What becomes a finding

One crooked AI-search answer is not enough for a finding. A record enters the corpus as an observation when the conditions of the Field run are preserved. It becomes material for a finding only after comparison with related records where a similar seam repeats. Repetition does not mean a perfect copy. In one field card, the old record pulls the address; in another, the service; in a third, the branch description. The shared mechanism remains: the citation trace supports a weak part of the answer and nudges the model toward a stale local picture.

For a business owner, this failure mode is unpleasant because it is so plausible. A completely false answer is easier to spot. A glued answer looks almost familiar. The name is recognizable, the city is nearby, the source is present. Yet the category is already foreign, the branch is out of date, and the directory sounds as if it has seen the company better than the company’s own current pages. In research reading, this is not a reason to jump straight to poor visibility. It is a reason to open the record and ask which fragment of the answer sits on which trace.

The lab therefore does not assign a score to the region or rank models by their resistance to old listings. The team marks the class of regional citability gap and its manifestation. In this batch of observations, the old branch became a useful magnifying glass: through it, the lab could see how a source may be formally adjacent to an answer and still pull the description back into a past operational reality.

Where the boundary of the conclusion lies

The method does not show the final truth about a company, branch, or region. Model answers change, source access can affect the citation trace, and local companies update sites and directories unevenly. Even an old listing does not always mean an error: sometimes a business really does keep a service in a separate form, work from an old address for some customers, or use an old name in legal documents. Without checking the current business structure, the conclusion has to stay soft.

That is why the lab phrases these records narrowly. In a specific Field run, under a specific Regional frame, the model tied a local name to a stale listing and produced a description where a former branch or former category blended into the current business. If that type of answer persists in related runs, it becomes a finding inside regional citability gap. If it does not, the record remains a field note: useful, uneven, and too thin for a broad conclusion.

This caution has practical value for the corpus itself. A stale-listing record does not get mixed up with a missing-company case, and it does not turn into a general discussion of weak business representation. It stays pinned to its place: old category, old branch, source beside description. The map of errors becomes uneven, but readable. The lab sees fine breaks instead of one large crack, and those breaks show where the AI-search answer lost its local coupling. In the field folder, this record is valuable precisely because it is small. It does not promise to explain the whole market. It points to the spot where an old paper label is still smearing fresh text, keeping the local card from drying cleanly before the next run and the next check of the same connection.

Last Local Pass

Last local pass: the tag on the sample reads Midwestern suburb, bookkeeping, citation-description split. The old listing still has glue on it: the firm’s current name sits on top, while the former address and tax preparation language show through underneath. The team keeps the seam visible because the citation trace supports the old shelf, not the current contractor bookkeeping claim. The next repeat run keeps the query wording unchanged and checks whether the stale branch still stains the description.