From memory. For the record.
The same question in two languages — once in Chinese, once in English — to the same panel of mainstream LLMs, on the same day, with both rankings side by side. Change the language and the list changes. Every raw response stays on file for anyone to check, including the ones that never answered.
Pet Dryer Box / Dog Washer
Ranked 1st of 56 candidate brands
2026-09-1312 models
Sample layout, built from the most recent evaluation — real brand, real score, real per-model ranks. A certificate number is issued at purchase; this brand has not bought one.
These aren't marketing figures — every one of those responses is stored verbatim and reprinted on the certificate page, so you can open any ranking and check the arithmetic yourself. How we compute it →
One question goes to every model on the panel in the same run, then the same question goes out again in the other language, on the same day. No follow-up prompts, no re-rolls, and no discarding an answer we don't like.
The answers come back in different languages and spellings. Brands are merged into one canonical name, then each model's ranking is weighted by its panel weight and aggregated into a single index.
Each response is stored verbatim alongside its model id, weight and contribution — and reprinted in full on the verification page of every certificate.
The same method, pointed at a question people are asking right now.
Re-measured every Monday: three questions, each asked once in Chinese and once in English, six cards in matched pairs — each card is one dated measurement.
For readers who evaluate AI engineering. This group stays online and is re-measured with exactly the same method as every other ranking.