GeoCheckTool Research

Original measurements for the parts of AI search dashboards cannot explain.

We publish the frozen questions, raw measurement boundaries, confounds, coding decisions, and open data behind each technical note. A result stays an observation until the evidence supports something stronger.

Open datasetsFrozen methodsExplicit limitations

Published

Technical notes

Results are published with the smallest defensible claim, including failed collection, missing verification, and unresolved confounds.

Technical Note 01Published July 28, 2026

Same Question, Same Engine, Two Devices: Retrieval Asymmetry in China's Biggest AI Assistant

We ran 30 frozen questions on both the web version and the mobile app of Doubao, China's most-used AI assistant. Web displayed more sources in 14 of 16 valid pairs; mobile never exceeded it. Full dataset, question bank and confounds published under CC BY 4.0.

Read the note

Measurement program

Next in the series

TN-02 · Planned

Query expansion

What the engine silently turns a question into before retrieval.

TN-03 · Planned

Source adoption

When a retrieved page informs the answer—and when it does not.

TN-04 · Planned

Entity boundaries

How adjacent companies, credentials, products, and capabilities merge.

TN-05 · Planned

Claim escalation

Commercial claims that become stronger than their cited sources.

TN-06 · Planned

Recommendation sets

How candidate lists widen, contract, and change across surfaces.

TN-07 · Planned

Retest stability

What repeats when the question, entity, method, and window stay frozen.

R1 · Planned

China AI Search Index

A larger recurring measurement release built from the sealed instrument.