Open dataset

Every poll we've ever run, unedited

Don't take our rankings on trust — audit them. This is the append-only history behind every leaderboard on this site: every answer ChatGPT, Claude, Gemini and Grok have given us, timestamped, failures included, nothing overwritten or filtered.

12,770

poll answers

1,430

What's inside

polls-devtools.jsonl — 7,143 records · software / dev tools / crypto ranking polls
polls-gear-opinions.jsonl — 5,534 records · consumer gear + opinion polls
polls-claude-variants.jsonl — 93 records · Claude Haiku/Sonnet/Opus variant comparison polls

Schema

ts	ISO timestamp the answer was received
asked_at	date the poll batch was launched
slug	category id — matches modelsagree.com/best/<slug>
type	“ranked” (top-5 list) · “pick” (single best) · “verdict”
question	the exact question asked
model	ChatGPT · Claude · Gemini · Grok
pick	the model's #1 answer (null on failure)
ranked	the model's full ranked list, first = best
reason	the model's verbatim one-line reasoning
err	error text when a poll failed — kept, we don't cherry-pick

Read this before citing it

These are answers from the consumer surfaces of each model, one fresh session per question, at whatever settings each product ships — we don't control temperature, which is exactly why each category is re-polled weekly and the full time series is published. A single answer is noise; the repeated record is the signal. This dataset measures what AI assistants recommend — not independent product testing — and that matters because millions of people now ask assistants what to use. Full protocol on the methodology page, including the verbatim prompt template.