Every prompt, the versioned vendor table and each edition's report, archived permanently at stable URLs. The raw judge labels and full answers are sent on request.
edition-2026-10-scores.jsonJSONOne row per named product, with the name exactly as written and a verbatim evidence quote.64.4 MBBy requestedition-2026-10-responses.jsonJSONEvery call with its full answer text, timestamp, version string, search flag, cited URLs, latency and token counts.50.0 MBBy requestedition-2026-10.mdMarkdownThe derived report for the edition, including the appendix on loose paraphrase stability.1.5 MBFree nowprompts.jsonJSONEvery prompt verbatim: six variants across 136 categories for each buyer segment, 816 per segment. Each prompt carries its set; the September 2026 mid-market set, asked as a B2B software company, is kept under its own ids beside the current one.1.1 MBFree nowmodels.jsonJSONPinned models, endpoints, search tools and list prices as of the edition.17 KBFree nowvendor_aliases.jsonJSONCanonical names, aliases, category-scoped readings, exclusions and the change log.1.9 MBBy requestmai-20261004-1101-calibration_b-scores.jsonJSONThe same prompts asked again on October 4, 2026 with nothing changed, over 12 categories. One row per named product, as in the edition file.5.7 MBBy requestmai-20261004-1101-calibration_b-responses.jsonJSONEvery call of the repeat with its full answer text, timestamp, version string, search flag, cited URLs, latency and token counts.4.4 MBBy requestmai-20261004-1101-calibration_b.mdMarkdownThe derived report for the repeat, the per-model flip rates behind the noise floor.141 KBFree nowThe pages, the standings and the method files are published under CC BY 4.0, free to cite. Until 4 October 2026 the September and October 2026 records were offered for download under CC BY 4.0; copies obtained then keep those terms. From that date no edition's raw labels or full responses are offered for download; they are sent on request, under the record license: for the research or reporting it was sent for, with commercial use, redistribution and model training by agreement.
The app reads this record for one company: every category it is named in, across all four indexes, what changed, who the models pick instead, and every answer behind every figure.
See the appSeven days free, no card.
| Cut | What it is |
|---|---|
| The flagship models | Claude Opus 5, Claude Opus 4.8, GPT-6 Astra, GPT-5.6 Sol, Gemini 3.1 Pro, Perplexity Sonar Pro, Muse Spark 1.3, on your category every month. |
| Job-to-be-done framing | A seventh prompt per category that describes the work instead of naming the category. |
Prices are on the Subscribe page. A cut that enough subscribers ask for can be promoted into the public index; that is announced in the edition log above and on the method page's change log, never done quietly.
A category enters the addressable set when it is a product a buyer asks an AI to recommend for going to market. Additions and removals are announced in the edition log and on the method page, never done quietly.
Every prompt set is one combination of four choices: how big the buyer is, what industry it is in, how the question is framed, and who answers. Each added value multiplies the prompt set, so these are taxonomy decisions first and scope decisions second.
A monthly edition on the same date. A special edition within days of a major frontier release, published against the prior generation. Quarterly written analysis, only once enough editions have accumulated to say something with substance.
When the vendor table changes, every edition in the series is regenerated under the new version, so the series is never a mix of table versions. A trend that appears only after a table change is tooling, not model behavior.
The historical series is the part that gets harder to replicate the longer it runs. Every edition stays at its stable URL, and the final edition is marked final on the site.