How clearly do AI assistants understand a category? Every Friday we measure one — whether the answers it gives are settled, accurate, and anchored to the real market. One locked method, all year.
The Index asks whether AI assistants give a coherent account of a whole category — settled, accurate, and anchored to the real market. One category a week, the same instrument every time.
Whether it answers the same way twice, whether what it says is true, and whether its ranking matches the real category. Those three roll into one number from 0 to 100, and a word for which kind of clarity it is.
Where each brand lands is how we read the category. How clear the category is then decides whether that position can be moved, which is why the verdict matters more than the row.
Fifty-three symbols, one a week. The symbol belongs to the category, not to us — assigned when the category is scheduled, unique across the season, unchanged if we run it again.
A hundred and twenty years after Upton Sinclair published The Jungle, the cliche about asking what's in a hotdog is still the same: “you don't wanna know.” A fifth of the frankfurter labels filed with the federal government don't name a specific meat. That isn't an omission, it's an admission: a pack that just says “Franks” is legally disclosing it's a melange of meats. But what we didn't expect is that the AI answering our shopping questions has absorbed the manners of the category it describes: where it knows, it is precise, but where it doesn't, it just keeps talking.
Read the edition →Nearly every answer said the same thing: the bag dies at the wheels, the handle or the zipper, so the only question worth asking is who pays. That makes warranty fine print the deciding factor in this category — and the fine print is exactly where the answers are least current.
Read the edition →Arrid: 0 of all 50 runs, both nouns — and not for want of knowledge. Asked directly, the model puts Arrid in this category 3 of 3 times and recites its history down to the ad slogans. Known, accepted, never recommended.
Read the edition →TD Bank: 0 of all 50 runs, both nouns — and not for want of knowledge. Asked directly, the model puts TD in this category 3 of 3 times and returns 15 specific facts about it. Known, accepted, never recommended.
Read the edition →