Does Your AI Visibility Report Count the Right People?

Suppose you ask an AI assistant to recommend an estate agent. The answer might name an individual, a firm named after its founder, or a local office. A dashboard that searches for name-shaped text can confuse those three outcomes.

A second problem appears if the dashboard checks only a list of professionals supplied in advance. It can tell you whether those names occur. It cannot, by itself, tell you how many other people the assistant recommended.

Who Gets Named examines both problems. The study collected 2,400 web-grounded API responses across professional-service categories and European markets. Its main detector classified 619 responses, or about 25.8%, as naming an individual. A separate comparison matched output against a roster of 939 people assembled from public LinkedIn search.

The distinction between those instruments is the useful part of the paper. One asks whether the answer names a person. The other asks whether an identified name overlaps with a particular list.

A name-shaped string is not yet a person

The study found 27,293 name-shaped mentions and 128 matches to its roster. Dividing those numbers gives 0.47%, the striking figure behind the paper's title.

But the denominator contains candidate strings, not a verified count of individual professionals. A place name or an eponymous business can look like a personal name. Treating every such string as a person makes the roster appear to capture a smaller share of real naming than that calculation establishes.

The paper also reports 128 roster matches among 1,876 text spans that the detector classified as referring to individuals: 6.8%. Those are mentions, which can repeat the same person, rather than a count of unique people. That denominator better describes the roster's overlap with identified person naming. It still does not establish the roster's share of every professional in those markets.

Neither number makes the matching problem disappear. The detector can miss people or misclassify entities. The study's reported precision and recall—96.9% and 61.7%—describe the roster-free detector's validation against roster-based ground truth. They are not a guarantee that every match is correct or that the validation covers all possible naming. Lower recall is a reason to inspect missed naming, rather than assuming that a zero count proves absence.

The answer and the roster can both be incomplete

A roster is useful when you have a defined question such as “Does this answer include our partners?” It becomes misleading when the score is presented as a measurement of all individual recommendations.

If the assistant names someone outside the list, the roster-based score records no match. That may be a perfectly correct result for partner coverage while saying little about whether the answer names professionals generally.

For service businesses, keep three records separate: whether the answer names any individual, which individual it appears to name, and whether that person belongs to your monitored list. Where identification is uncertain, retain the uncertainty instead of forcing a match.

A firm's name needs particular care. “Smith Insurance” is not evidence that the answer recommended a person called Smith. The surrounding sentence, linked page and business identity may be needed to decide what was meant.

Source type is a clue, not a promotion tactic

The study also compared the sources cited in answers that named individuals with those that did not. Personal sites and category portals were somewhat more common in naming responses. Citation volume alone did not explain the pattern.

That association is worth inspecting because a citation count hides what the sources contain. A page about an individual supplies different information from a directory of firms. It does not follow that buying a portal listing or publishing a personal page will cause an assistant to recommend someone.

The source-type differences were small, and category and model behavior varied. The design did not isolate a causal naming trigger. It also does not establish a reliable way to suppress unwanted naming by changing one source type.

Inspect the matches behind the percentage

For a person-level audit, ask to see a sample of counted matches, rejected candidates and ambiguous cases alongside the complete answers. Check names against firm names and locations. Record which monitored people were never found, but keep that result distinct from a claim that the system never recommends individuals.

Then examine the roster itself. Does it include the people relevant to the market and buying question, or simply the people easiest to collect from one platform? A precise matching calculation cannot compensate for a poorly defined population.