Why Does AI Name a Particular Expert or Agent? Avoid the Roster Trap in Individual GEO Monitoring

In real estate, automotive, consulting, or insurance, a user may not be looking for a company at all. They may ask whom to contact. An AI may name a person, list an organisation, or do neither. When names are shared, roles have changed, or sources are unclear, treating one mention as a personal-reputation verdict is risky.

A common shortcut is to match answers against an employee or LinkedIn roster and call the matches “individual visibility.” That can miss aliases, namesakes, former staff, and unverified model-generated names.

A preprint first published July 26, 2026 and updated August 1 studied individual naming in 2,400 grounded model calls made on July 24, 2026. The paper reports that 25.8% of answers named an individual under its conservative detection rules and that the four models differed by about fourfold. It also reports that a 939-person roster built from public LinkedIn search matched only 0.47% of 27,293 name-shaped mentions. Those are lower-bound results for the study's markets, languages, and identification rules - not universal rates.

Start with the answer, then verify the person

The safer workflow runs in the opposite direction: retain the full answer and cited source first, then decide whether the name resolves to a verifiable individual. Record the question, date, model, name as written, role, organisation, city or market, supplied link, authoritative confirmation status, and namesake ambiguity.

Only then compare verified records with a company directory or public information. A roster is useful as a supporting check, not as the denominator for every person an AI has named. Mark unverified names as requiring review; do not contact, rate, or publish a ranking of people on that basis.

Segment results by task and category

The study suggests large category and model differences. Use real selection tasks such as “a local agent experienced with first-time buyers” or “an adviser with business-insurance experience,” while fixing location, language, and constraints. Track separately:

• whether the answer names a person, organisation, both, or neither;
• whether the name matches the need, location, and current professional status;
• whether the source is a personal site, category portal, organisation page, or nothing verifiable; and
• whether the result is stable across models, time, and language.

An occasional model mention is not a certification, ranking, or endorsement. Never try to “fill in” answers with false profiles, impersonating pages, or personal data used without permission.

Add privacy and human judgment to individual GEO

For personal information, minimise retention, control access, provide a correction or deletion route, and follow privacy and employment rules in the relevant market. In medical, financial, legal, and employment contexts, an AI answer should not itself be used as a recommendation basis.

GEO Radar at https://www.georadar.top can retain observations, sources, and changes across platforms for fixed questions so a compliance owner or business reviewer can assess them. It does not validate professional qualifications or guarantee that an individual will be named.

Sources for this article

• arXiv, first published July 26, 2026; version updated August 1, 2026, *Who Gets Named: Citation Type Predicts Individual Naming by Grounded Language Models, and a Roster Instrument Captures 0.5% of It*: https://arxiv.org/abs/2607.23893 (2,400 calls, naming rate, model differences, and the roster-matching limitation)