How a repeated prompt becomes a field observation

In a deliberately simple field card, the prompt "independent HVAC repair in western Pennsylvania" returns a crooked answer: the model points to a local company page, describes the business as a plumbing contractor, and adds a broad directory instead of a regional source. For Citability Field Lab, an observation is a concrete model answer tied to a recorded prompt: wording, regional anchor, named businesses, business descriptions, and source when one appears. A single answer is like a road sign photographed after rain. The arrow is visible; some of the paint has run. A finding appears only when several photographs show the same bend: a skipped company, category drift, regional substitution, or a strange seam between citation and description.

Sample sets are assembled descriptively. The team works across states, urban and suburban contexts, several local service categories, and comparable user intentions. One set may contain a bookkeeping prompt for home-service contractors, an HVAC repair prompt in western Pennsylvania, and a commercial roofing prompt in North Texas. A single national number would flatten that setup. The lab lays pieces of a regional map on the table and watches where the ink behaves differently.

Repeatability depends on preserving small details. The original prompt wording, regional parameter, model, response mode, and order of comparison stay with the record. When one element changes, the run is marked as a separate sample, even if the result looks almost the same. That strictness seems fussy from a distance. Inside AI-search, small changes often move the answer: the model holds the city but swaps the county; it names the right firm and stitches on a neighboring service.

The team reads model comparison like several maps of the same district printed by different shops. On one, the roads match, but the small street labels vanish. On another, the streets remain, while the business category spreads into a label too broad to help. The team looks for cases where the description is supported by the source and cases where the source hangs off the answer like a badly sewn tag. A regional citability gap is a local AI-search failure where a business name, region, category, or source no longer line up. The lab marks five recurring forms: entity skip, category drift, regional substitution, citation-description split, and directory dependency.

The limits of the method stay visible. Model answers change, source access can affect the result, and local businesses update websites and directories unevenly. A note therefore describes an observed pattern; it does not claim final truth about a company. If a model omits a business in one run, the lab records a narrower fact: the prompt, region, and comparison where that omission appeared.

Interpretation is kept separate from observation. When the team moves from a recorded answer to a reading, the wording shifts toward "probably," "in this run series," or "if this answer type holds." That makes the prose less dramatic, but it keeps the claim at the right size. In these run series, local AI-search can resemble a telephone book rebound by hand, with a few pages from the neighboring county slipped into the middle.