Chunking is the whole trick
Try to hold 4 7 1 9 2 8 3 5 as eight separate items and you will lose some of them. Read it as 47 19 28 35 and it is four items, each of which happens to carry two digits. The capacity that runs out is roughly a count of items, not a count of digits, so anything that packs more digits into an item buys you length directly.
People do this without noticing. Phone numbers, card numbers and dates are all written in groups for exactly this reason, and if a number happens to contain a year, an age or a familiar pair, most people grab it as a single unit without deciding to. Doing it deliberately — settling on two-digit groups before the number appears and reading it in that rhythm — is usually worth two or three digits immediately.
What the classic figure actually says
The number people quote is seven plus or minus two, from George Miller's 1956 paper on the span of immediate memory. It is widely cited and widely over-read. Miller's own point was partly that the limit sits in chunks rather than raw items, which is why the figure moves so much depending on what is being remembered. Later work on working memory proper tends to land closer to four items once chunking is prevented, and digit-span results sit higher than that precisely because digits chunk so easily.
So a result of eight or nine here is not evidence of an unusual memory. It is more likely evidence that you grouped well, which is a skill and improves with practice in a way raw capacity does not.
Why one mistake ends it
Span tests end at the first failure by convention, and the alternative is worse: a second attempt at the same length means you have already seen a number of that length once, and the practice effect contaminates the result. Restarting deals an entirely fresh sequence at every level, so repeated runs stay comparable with each other.
What the score does not tell you is anything about intelligence, learning ability or memory in the everyday sense. It measures how many digits survived a few seconds in a quiet room with a screen in front of you. Recognition, recall of meaningful material and the ability to learn are separate systems, and someone can be unremarkable at this and excellent at all of them.
Questions people ask
How long is each number on screen?
The setting controls the base time — two, two and a half or three seconds — and longer numbers get a small increment on top, about 150 milliseconds per extra digit. Without that, later levels would be unfair in a way that has nothing to do with memory: you would run out of time to read the number before you ran out of capacity to hold it.
Can I use a different strategy each run?
Yes, and it is worth doing. Try one run reading the digits one at a time, then a run grouping them in pairs, then a run saying them aloud in a rhythm. The difference between those three on the same person is usually larger than the difference between two different people using the same strategy.
Does it get harder in any way other than length?
No. The only things that change between levels are one more digit and a fraction of a second more display time. The digits are uniformly random apart from a non-zero first one, so a level is never deliberately made awkward.
Is anything recorded?
Nothing. The number and your answer exist in the page and nowhere else, and both are gone when the tab closes. No score is stored and nothing is sent anywhere.