Look up the same token on two rarity sites and the ranks often differ, sometimes by hundreds of places. The disagreement is not an error; it comes from how a multi-trait item gets reduced to one number.
Traits are counted independently
A generative collection assembles each item from separate trait categories, each with its own set of options and its own distribution.
Counting how often a single option appears is unambiguous. Every tool agrees on how many items have a particular background or headwear.
The ambiguity begins when several such counts must be combined into a single ranking, because there is no objective rule for comparing an item with one very rare trait against one with several moderately rare traits.
Summing and multiplying give different winners
One common method scores each trait as the inverse of its frequency and adds the scores. This rewards items that carry a single exceptional trait.
Another treats the item as a joint outcome and multiplies the trait probabilities, which favours items whose traits are collectively unusual even if none is individually extreme.
A third approach borrows from information theory and weights each trait by how much it narrows down the possibilities, which sits between the other two and penalises heavy weighting of one attribute.
Trait count itself is a hidden variable
Many collections give items different numbers of traits, and some leave categories empty on purpose.
Whether an absent trait is scored as its own rare option or ignored entirely changes the ranking substantially, and tools split on this decision.
Items with fewer filled slots often rank very high under one treatment and unremarkably under another, which explains a large share of the visible disagreement in any collection that uses empty categories.
Market prices follow attention, not arithmetic
Buyers rarely price a token by any of these formulas. They price recognisable traits, established sub-collections and whatever the community has decided is desirable.
A statistically rare combination that looks unremarkable can trade near the floor, while a common trait with a name and a following commands a premium the arithmetic never predicts.
Rarity scores are best read as one input describing supply, with demand determined by factors the trait table does not record.
Ranking is a whole-collection calculation
A rarity rank is not a property of an item. It is a position within a set, computed from the trait distribution of every token in that collection.
That makes ranks unstable while minting is still under way, since each new item shifts the frequencies the scores are built from.
It also makes ranks incomparable across collections. A high rank in a set of ten thousand items with twelve trait categories describes something quite different from the same rank in a smaller set with four.