Where the citation hides beside the model answer

A source in an AI-search answer can look like proof while holding only the edge of a phrase. The lab reads these links as seams: where they are stitched into the description, and where they hang loose beside it.

The link stood right after the contractor’s name, like a neatly sewn tag. In a model comparison for the query AI answer cites directory, the lab used a composite scenario assembled from several observations around a commercial roofing contractor in North Texas. ChatGPT named the company inside the right regional frame, but the description widened into general contractors, while the source resembled a broad directory listing with an address and phone number. Perplexity, in a related Field run, produced a denser citation trace, yet a phrase about construction services still sat beside it without a clear roofing specialization. Gemini kept the industry label cleaner, but the attached link confirmed the business’s existence more than the specific roofing service.

A second field card in the same series brought in another composite scenario: a small bookkeeping firm in a Midwestern suburb, working with home-service contractors: HVAC, plumbing, roofing, cleaning. One AI-search answer held onto an office listing, while the nearby text spoke about payroll software. In another answer, a broad directory became the closest source for a phrase about tax preparation, although the business’s current emphasis was bookkeeping for contractors. The crooked detail: the company name appeared in its newer form, with an updated word at the end, while the category came from older language. The lab recorded not the presence of a link by itself, but its distance from the particular claim.

The citation that cannot carry the weight

Citation trace often calms the eye. The answer contains a source, so the description feels checked. But in the lab’s field notes, the source sometimes works like a thin board under one leg of a table: the table stands until someone presses down. It may confirm that the company exists, sits in a region, or appears in a directory. That is still not enough to confirm the service, specialization, service area, or connection to the exact category in the query.

That is how citation-description split appears. The description says more than the source can hold. With directory dependency, the weakness becomes easier to see because a broad directory gives the model a convenient form: name, address, phone number, and sometimes a rough category label. For a local business, that form is like an old luggage tag. It helps identify the owner, but it does not say where the suitcase is going now.

The lab therefore looks at the source’s load-bearing work; the fact of a link remains only the first check. Which part of the AI-search answer does it hold? Does it support the name? The region? The category? Or does it merely stand near the text as a sign that the model is not speaking from empty air? These questions involve plenty of dull manual work. The description has to be read beside the source, not merely scanned for the presence of a link.

How the models attached the source differently

In this series of runs, the comparison of ChatGPT, Perplexity, and Gemini did not become a tournament. The team watched the attachment pattern. One type of answer can name the correct company and provide a source that holds only the address. Another type gives a richer citation trace but uses it as background for a neighboring category. There is also the more cautious version, where the phrase sounds softer, yet the reader still cannot tell what the source confirmed.

For commercial roofing in North Texas, this appeared when the specialization widened. When the answer spoke about roofing, the source sometimes remained weak: it showed a contractor listing but did not support the commercial specialization. When the answer spoke about general contractors, the source looked more compatible with the text, but the category itself became too broad for the query. The model seemed to be choosing between industry precision and source convenience. The lab does not claim that this is how the model’s internal decision works; it is only the readable effect inside the AI-search answer.

In the bookkeeping scenario for home-service contractors, the weak citation trace led in another direction. A directory could store the firm under tax preparation, while the answer nearby spoke about contractors, payroll software, or a general financial service. The source did not disappear. It hid beside the answer: close enough to create trust, far enough away to avoid carrying the main phrase.

The team also watched the order in which the model displayed the source. Sometimes the link appears immediately after the name, and the disputed category arrives in the next sentence. Sometimes the reverse happens: the description first sounds confident, and the source appears at the end of the block as a general trace. These differences matter because readers often attach the whole paragraph to the nearest link. The lab reads the answer more slowly: does the source belong to the specific phrase, or only to the company listing? In local categories, that pause in the eye changes the whole reading.

Subtypes of side citation trace inside the canon

Side citation trace is a form of citation-description split in which the source sits beside a claim but supports only a weak part of the answer: business existence, address, broad region, or broad directory category. In the lab’s canon, it remains a subtype inside citation-description split; when the source is drawn almost entirely from a broad directory, the record also touches directory dependency.

Inside this subtype, the lab distinguishes several forms, described in words rather than points or scales. There is the existence trace: the source confirms that the company is listed somewhere, but does not support the category. There is the regional seam: the source helps hold the city or county, while the service description comes from another layer. There is the category hollow: the source gives a name and address, but does not explain why the company is named in commercial roofing, bookkeeping, or HVAC. There is the directory backing: a broad directory sets the tone for the whole answer, and directory dependency begins to sound louder than the company’s current local page.

This typology is useful because it makes the question more exact. The question shifts from a simple “is there a link?” to a narrower one: “which fragment of the answer can this link bear?” That change of angle removes some of the ceremony around citation. The source stops being a stamp on a document and becomes a thread that can be pulled.

Inside one record, that thread can lead to several shelves. Different shelves carry the company name, city, and industry label. When the model folds them into a smooth paragraph, the shelves disappear from view. The side citation trace subtype puts them back in place. It shows that support for an answer may be dispersed, while the confidence of the sentence is assembled from pieces of uneven density.

Why the directory feels strong

A broad directory is convenient for the model because it packages a local company into a short form. Name, address, phone number, industry label, and neighboring companies often sit side by side. To a person, this looks like a raw listing that needs checking. For an AI-search answer, it can become a ready-made piece. Especially when the query asks for a list or comparison, the model looks for compact traces that can be placed next to one another.

The problem begins when compactness substitutes for support. For a commercial roofing contractor, a directory may say “contractor” and therefore avoid a blunt falsehood, but for a commercial roofing query the label is too coarse. For a bookkeeping firm, the listing may store tax preparation because that was once a visible description, while the current local demand is tied to bookkeeping for contractors. The answer still does not look chaotic. It can even sound neat: the company is named, the region resembles the right one, the source is in place. The main tie between category and source is simply thin.

Here the difference appears between a source as a trace and a source as support. A trace says: “a company like this appears somewhere.” Support says: “this company belongs to this category inside this regional frame.” The lab tries not to confuse these two registers in its field cards. Otherwise, a broad directory begins to work like a heavy folder on a light sheet of paper: it pins the text to the desk, but it does not make the writing more accurate.

In the lab’s field notes, these cases resemble a shelf of archive folders. The correct company name is written on the spine, but an old version of its map sits inside. The model does not pull out a folder that is entirely wrong. It pulls out a folder where some pages should have been replaced. That is why the error rarely looks like a full collapse. It looks more like an outdated paragraph in an otherwise usable dossier.

How the lab separates a useful trace

A useful citation trace does not have to be perfect. Sources for local businesses are often uneven: the site is updated, the directory lags, the industry profile writes too broadly, and the regional page speaks only about the address. The lab does not require every source to confirm everything. But it marks the moment when the source does not hold the part of the answer it seems to be attached to.

In the working card, the team reads the source against a specific phrase. If the answer says that the company specializes in commercial roofing for a defined Regional frame, the source should at least support that specialization or show a clear connection to it. If the answer speaks about bookkeeping for home-service contractors, a source with one office listing and old tax preparation language looks weak. It may be a useful trace of the name, but it is a poor trace of the category.

That is why a small research note appears beside each source in the card: what it actually carries. Sometimes it is the name. Sometimes the address. Sometimes an old category. Sometimes a regional hint without the service. The note looks almost bookkeeping-like, but without it, model comparison becomes reading by covers. Different citation trace chains can look equally calm, although one holds the main category while another merely confirms that a similar company exists somewhere. In model comparisons this is especially visible: different answers may look equally sourced, but one source is stitched to the claim and another clings to the fabric’s edge. So the lab does not record citation trace as a simple “present” or “absent.” It looks at the work this citation trace performs inside the sentence. That reading is slower, but it prevents a broad directory from posing as proof of a narrow local specialization.

This kind of review disciplines the finding. The lab can describe the form of regional citability gap more precisely: the source is attached to the name, the description draws the category from a side channel, the directory holds the region, and the citation trace creates trust where support is partial. For model comparison, the length of the link is not the main issue. What matters is how tightly it is stitched to the claim.

Limits of the comparison

The method does not show universal behavior for ChatGPT, Perplexity, or Gemini. Answers change, source access mode changes the picture, and local businesses update pages, listings, and directories unevenly. One Field run does not make a model a permanent offender. One weak source also does not prove that the company is poorly represented across all AI-search systems.

There is a subtler limit as well: sometimes a source seems weak only because the researcher does not see the full context. A company may have a separate page, an old service may remain in a narrow segment, and a regional listing may be acceptable for part of the market. The lab therefore keeps its phrasing at the level of observation. In a specific series of runs, the source sat beside the main phrase. If that seam repeats under comparable conditions, it becomes a finding inside citation-description split or directory dependency. If not, the record remains a carefully labeled specimen that cannot yet be presented as a map of the whole district. This restraint makes the conclusion quieter, but preserves the main sense of the observation: a source has to be read at its point of attachment, not by the mere fact of its presence.

Last Local Pass

Last local pass: the tag on the sample reads North Texas, commercial roofing, citation-description split. The thread sits off to the side: the link hugs the contractor’s name but carries address and directory residue more than a narrow commercial roofing role. The team follows it back to the model sentence and marks where construction language grew wider than the source. The next repeat run keeps the Regional frame unchanged and checks whether the citation still hangs beside the claim.