Data

The files behind every number on this site. Each one says how big it is, when I collected it and which tools I used.

All of it so far is CLI data: answers from command-line developer tools with no shopper location set, not the apps people use on their phones. It isn't comparable to later tests that set the shopper's city. How I test

Four shopping questions in 8 metros

Each question asked once per metro through each tool, counted by what the answer named first and whether it named a local independent agency anywhere.

Size: 64 answers
Collected: October 11, 2026
Tools: OpenAI Codex CLI (gpt-6.1-sol), live web search; Google Antigravity CLI (gemini-3.8-flash-high), search_web tool. CLI data.
Cities: Atlanta; Columbus OH; Dallas-Fort Worth; Grand Rapids MI; Indianapolis; Kansas City; Phoenix; Tampa
File: shopping-questions-2026-10-11.csv (CSV, 8 rows)

One question, 20 times per city

“recommend a local insurance agency in [city]” asked 20 times per city through each tool. Agency names replaced with letters.

Size: 80 runs
Collected: October 11, 2026
Tools: OpenAI Codex CLI (gpt-6.1-sol), live web search; Google Antigravity CLI (gemini-3.8-flash-high), search_web tool. CLI data.
Cities: Grand Rapids, MI; Phoenix, AZ
File: ai-naming-test-2026-10-11.csv (CSV, 4 rows)

Named agencies and Expertise.com lists

Independent agencies the tools named, and agencies picked at random that were never named: how many sit on an Expertise.com car or homeowners list for their city. A pattern, not proof.

Size: 157 agencies
Collected: October 7 and 11, 2026
Tools: OpenAI Codex CLI (gpt-6.1-sol), live web search; Google Antigravity CLI (gemini-3.8-flash-high), search_web tool. CLI data.
Cities: Phoenix; Grand Rapids MI; Tampa; Kansas City
File: expertise-listing-2026-10-11.csv (CSV, 10 rows)

Update log

October 15, 2026: page published with three files from the October 2026 scans.