What 'chance level' means, and why it beats a percentile

By Eric, BlinkBench Founder · Last reviewed: July 2026

A percentile tells you how you did against a crowd. It needs that crowd to exist, to be described honestly, and to have been measured on comparable equipment — which, as the companion guide to this one goes into, is rarer than the internet's supply of percentiles would suggest.

A chance level asks a smaller, cleaner question: if you had answered at random, with no memory at all, how often would you have gotten this exact score by luck? It needs no crowd. It needs only the rules of the test, which is arithmetic we already have. For three of BlinkBench's span tests, that arithmetic comes out differently — and the difference is the shape of the task, not a coincidence.

Three tasks, three counting problems

“How many possible answers are there?” sounds like one question. It isn't. It depends on whether an item can repeat and whether order counts, and each of the three span tests below answers those two questions differently.

Number memory: digits, with repeats, order fixed

You are shown a string of digits, 0 through 9, and asked to type it back in the order shown. Each digit is drawn independently, so 77 is a legitimate string. A blind guess has to get every position right, and each position has 10 equally likely digits regardless of what the others were — so the odds multiply: 10 × 10 × 10 …, once per digit. For an n-digit string, the odds of guessing it are exactly 1 in 10ⁿ.

Sequence memory: positions, no repeats, order matters

Sequence memory lights nine fixed blocks one at a time and asks for them back in the same order, and the same block is never used twice in one sequence. That is a permutation: the first tap has 9 possible blocks, the second has 8 remaining, the third has 7, and so on. The odds of guessing the first n taps of a sequence are 1 in P(9,n) — the product 9 × 8 × 7 × … stopping after n terms.

Order is what makes this a permutation rather than a combination — swap two correct blocks and the answer is wrong, even though you tapped the right set. That is also why it grows faster than number memory's: guessing gets one fewer option to work with on every position instead of the same ten each time.

Visual memory: cells, no repeats, order doesn't matter

Visual memory is the odd one out on purpose — it was built to strip the sequential part out of the block-tapping task above and leave only shape. The whole pattern appears at once, and you mark the cells in any order you like, so guessing right means choosing the correct set of k filled cells out of 2k total, with no credit for the order you marked them in. That is a combination, and the count of possible sets is the binomial coefficient C(2k,k).

Computed by guessOdds() in lib/pattern.js — the same function the test page itself calls to print the figure beside your score.

Why order-blind counts faster than it looks

It's worth sitting with why the visual-memory numbers climb so much faster than the sequence-memory ones, because it isn't obvious. Removing the order requirement looks like it should make guessing easier — fewer constraints, surely more ways to satisfy them. But the two tasks aren't counting the same thing: sequence memory always has 9 fixed slots to draw from and stops multiplying once it runs out of blocks, while visual memory's matrix grows with the level — 15 filled cells at the ceiling live in a 30-cell board, not a 9-cell one. A bigger board with no order constraint still produces astronomically more possible answers than a small board with one. The lesson isn't that order-blind tasks are harder to guess in general; it's that the size of the answer space is doing the real work, and order is just one factor multiplying into it.

What a chance level is good for, and what it isn't

A chance level answers one narrow question exactly: how much of a given score could have come from luck alone. It says nothing about how a typical person performs, because it isn't built from people — it's built from the rules of the game. That is also its limit. A test where the chance level is 1 in a billion at the ceiling isn't telling you the average visitor is a billion times better than random; it's telling you that reaching the ceiling by luck is effectively impossible, which is a different and much narrower claim.

It does one thing a percentile can't: it needs no crowd, so it carries none of a percentile's baggage about who took the test, on what hardware, filtered how. Where a defensible reference range exists, this site cites one. Where it doesn't — as on visual memory, whose norms sit in a printed manual we haven't read — the chance level is what's left, and it turns out to do most of the work a percentile pretends to do, honestly.