Why a General Directory Sounds Stronger Than the Company Site
A local company may keep a careful, updated site, while AI-search still hears it through the old metal horn of a directory. The voice is recognizable; the accent belongs to someone else.
The team had two descriptions of the same local company on the table. The company site spoke narrowly and calmly: a local service, a specific region, several specializations. The card in a general directory sliced the same business into a broader shape — broad category, neighboring service, city-level wording that left out the smaller service area. In the record for AI search local directories, the AI-search answer chose that broader cut. The company site either appeared as a secondary source or did not shape the language of the answer.
This is a composite scenario drawn from several lab observations. It does not point to one specific organization. In these scenes, the error rarely looks like an invention. It looks more like an old label on a crate: the tools inside have already been rearranged, but the outside still says “general services” in large print. The model sees the label, repeats it with confidence, and sometimes sounds more persuasive than the company itself.
Where the directory becomes the business’s voice
Directory dependency is a class of regional citability gap in which an AI-search answer relies on a general directory more strongly than on a local or company-owned source. This does not necessarily happen because the directory is “better.” It may be easier to read, more stable in structure, more common across similar cards, and more likely to provide a ready-made line: name, city, category, phone number, sometimes a short description. For a machine answer, that form is almost ideal. For a local business, it can be too blunt.
In the lab’s field notes, a general directory often sounded like the company’s official interpreter, though the company had not appointed it to that role. A site might describe commercial roofing for certain kinds of properties. A directory might give the line “roofing contractor,” or an even broader category. A bookkeeping firm’s site might describe work with home-service contractors, while the card says tax services or financial services. In the finished answer, the model takes the voice that is easiest to assemble, even when it is not the freshest or most precise.
This dependency is especially visible in local categories where company sites are written unevenly. Small businesses sometimes update the homepage but forget old cards. Sometimes a new specialization lives on a services page while the directory preserves an earlier category. Sometimes the site is written in human language, with long paragraphs and no ready-made short definition; the directory gives a dry but convenient line. AI-search likes lines it can place quickly into an answer.
Why the company site does not always win
The company’s own site seems like the natural source of truth. But for AI-search, truth in the answer is assembled from available and comparable text traces, without the legal source hierarchy a business owner might have in mind. The lab phrases this cautiously as an observation, not as knowledge of the models’ internal machinery. In a Field run, the answer may pull in a general directory because it sits closer to the query’s words, because it has an explicit category, because it provides a short card, or because the company site does not contain the needed phrase in a convenient place.
A site can also be too specific. For the owner, that is a virtue: it describes real services, boundaries, areas, and customer types. For a model building a quick answer, that specificity can become loose material. The directory cuts the business into a ready form. It resembles a school stamp: crude, but even. When the user’s query is broad as well — for example, local directories, local business, contractors near region — the model may prefer the stamp to the finer-grained company description.
In one observation, this was almost comic. The company site spoke at greater length, with more life and more precision, but without a short phrase that could be easily lifted. The directory, in contrast, gave a dry category and city. The model used that dry line as the spine of the answer, while details from the site, if they appeared at all, became soft additions. The broad source does not necessarily win because it is right; it wins because of its form.
This is where directory dependency crosses citation-description split, though the two patterns are not the same. In citation-description split, the main question is whether the source supports the description. In directory dependency, the question is why the general directory became the main voice even when a more local source sits nearby. The classes can appear together. A directory may dominate the answer and weakly support a specific service at the same time. In the lab’s record, however, it is useful to separate dependence on the directory from the break between link and description.
Manifestations inside directory dependency
Within directory dependency, the team describes several manifestations. These are working subtypes within one canonical class; the material does not introduce a parallel classification of regional citability gap. One subtype is directory wording: the answer nearly borrows the directory’s broad language and smooths away the company’s specialization. Another is directory geography: the card sets the city or broad area, and the model carries that frame into the answer even when the company site draws a more precise regional frame.
There is also directory category: the directory assigns the business to a general service, and the model then uses that category as primary. For commercial roofing, this can mean a slide toward general contractor; for bookkeeping, a shift toward tax preparation or financial services; for urgent care, an expansion into hospital service or general medical clinic. Another manifestation is directory source hierarchy, where the answer is built as if the presence of a card in a general directory matters more than the local page. This is especially visible when the company site is mentioned later, weakly, or not at all.
The vocabulary helps with field-card reading. It keeps all directory effects from collapsing into one heap. Sometimes the problem is language, sometimes geography, sometimes category, sometimes the source that has been allowed to speak first. Without these distinctions, the researcher sees only that “models like directories.” For the Citability Field Lab, that brush is too thick.
Two composite cards on the table
Study object A: a composite scenario. A small bookkeeping firm in a Midwestern suburb works with home-service contractors — HVAC, plumbing, roofing, cleaning. On its site, in this scenario, the firm explains that it handles books for service companies, handles regular operating bookkeeping, and does not sell a software product. In one Field run, AI-search retells the firm through a general directory: local tax services, payroll help, financial services. The name is recognizable, the city is almost the same, but the category has become wider and colder.
The lab does not conclude from this that the directory is “wrong.” The directory may honestly preserve a broad category that once fit or still partly fits. The problem is that the AI-search answer begins using this broad category as the company’s face. The site speaks in one voice, the card in another; the model chooses the more convenient voice. As a result, the owner sees a strange reflection rather than a total visibility failure: the company is named, but its niche work for home-service contractors has dissolved.
Study object B: a composite scenario. A regional commercial roofing contractor in North Texas appears differently. A general directory may attach the company to a broad construction category or to the city frame of a neighboring metro market. AI-search then describes the business as a general contractor or ties it to a market that sounds plausible but does not match the company’s narrower image. The small crooked detail matters: in one card, the model correctly held part of the service area, while the source beside it spoke only about the office address. Dependency then intertwines with citation-description split.
Both cards show why directory dependency rarely looks like a simple mistake. The model almost always takes something real: name, address, category, old line, area. The failure comes from weighting. The general directory becomes heavier than the company-owned source, and the answer begins to sound like a reconstructed card from someone else’s database, though the reader expected a description of the business.
What counts as an observation, and what counts as a finding
The lab treats a specific AI-search answer as the observation: query wording, regional frame, model, answer mode, named businesses, descriptions, and sources. If a general directory is the main citation trace in that answer, this is not yet a finding. The team needs to compare related observations and see a repeated form: the directory again sets the category, again broadens the region, again sounds stronger than the local source.
Separately, the team looks at the form of the source itself. A directory may be precise on the address and poor in meaning; it may hold the city but not the service area; it may give a category that is too broad for the current site. Without that cut, directory dependency is easy to confuse with the ordinary presence of a directory in an answer.
This discipline protects the work from an overly convenient story. It is easy to say: “models trust directories more than sites.” The lab phrases it more carefully: in this series of runs on local-service queries, a general directory often became the practical voice of the business, especially when the company site offered less compact text or language that matched the query less directly. This is an assessment from a series, not a universal rule.
For reading these answers, a simple order helps. First, the team looks at which page is named as the source. Then it asks which part of the description that page supports. Then it looks for a more local source that might have said the point more precisely. Only after that does the record receive the directory dependency label. Otherwise, dependency on a directory can be mistaken for the mere presence of a directory in the citation trace. The directory itself is not the problem; the problem begins when its blunt wording controls the whole scene.
There is another signal the team marks separately: the order in which sources appear. If a general directory stands first in the answer and sets the description’s vocabulary, while the company site appears as secondary support, dependency is already visible in the composition of the answer. Even when the site does not disappear, it may sound quieter. For a local business, that is especially unpleasant: the company’s own text is present, but it does not govern the meaning.
Sometimes the company site appears below the directory and seems to correct the broad category, but the order has already done its work. The first line gave the user a frame; the remaining sources are read as refinements. For the observation, this is an important detail: dependency can appear not through the absence of the site, but through the site arriving too late in the answer’s composition.
Limitations and a cautious finding
This material does not prove that a general directory is always stronger than the company site. It does not show the internal rules of ChatGPT, Perplexity, or Gemini, and it does not rate the quality of specific directories. In different source-access modes, the answer may change. Local companies update sites and external cards unevenly; sometimes the directory truly contains a detail that the site lacks. The lab therefore records an observed pattern, not a final truth about the source channel.
The cautious finding is this: directory dependency appears when an AI-search answer takes a general directory as the practical voice of a local business and lets its broad language outweigh the company’s narrower description. In these scenes, the business does not disappear. It speaks through a borrowed mouthpiece, where every word is slightly flattened and the local category starts to resemble a tag from a large storage box.
Last Local Pass
Last local pass: the tag on the sample reads US local markets, local services, directory dependency. The specimen carries someone else’s stationery: a general directory gives a short category line, and the answer starts speaking in that voice instead of the narrower company site. The team hears the listing line become the spine of the description, thin on live neighborhood context and current specialization. The next repeat run keeps the regional frame unchanged and checks whether the directory still speaks first.