Skip to content

Research

Dated studies of what AI assistants say

Each study states when its answers were collected, how many it covers and what it cannot show. Findings below are the ones each study page states for itself. Datasets are linked where we publish one, so you can check or cite the numbers.

Updated · Data changelog

  1. Searched

    From a Google AI Overview to a Sacramento plumber’s estimate form

    Sample
    One Google AI Overview for one Sacramento water heater question; three linked pages read.
    Headline finding
    Google’s AI Overview for one Sacramento water heater question linked to guides on what to ask before replacing a unit.

    Read the study →No dataset file for this study

  2. Collected

    Which credit unions AI recommends, compared with who members choose

    Sample
    480 AI answers in eight US metros, compared with NCUA data.
    Headline finding
    The largest credit union is named less than its size suggests, and nearly 1 in 4 answers uses a retired name.

    Read the study →Dataset (JSON)

  3. Captured

    Eight plumbing markets, two AI assistants, different shortlists

    Sample
    Eight public plumbing leaderboards, 80 sampled ChatGPT and Gemini answers (10 per board).
    Headline finding
    The most-mentioned plumber appeared in anywhere from 5 to 10 of 10 sampled AI answers, depending on the city.

    Read the study →Snapshot (JSON)

  4. Reviewed

    One destination report is a snapshot

    Sample
    20 stored destination reports across ten towns, giving 30 list pairs.
    Headline finding
    Every one of the 30 lodging, dining and activity list pairs changed between the two most recent reports.

    Read the study →Report aggregates (JSON)Comparison results (JSON)

Method and related reports

These are first-party observations of AI answers on the dates shown. They are not controlled experiments, and they do not show why an assistant named a business or whether a mention led to a customer.