Here is a true sentence about Genshin Impact: the share of new characters that are five-star rose from 38% in 2020 to 73% in 2026. Both figures are correct. Anyone can check them. And the sentence is close to worthless, because it describes a rise that the years in between do not contain.
The full series runs 38 · 65 · 53 · 53 · 71 · 60 · 73. 2020 is the outlier at the bottom, 2024 already reached 71% two years before the end, and the middle of the range wanders between 53% and 73% without settling. What the data shows is a step up after launch year and then noise around a level, which is a much less quotable finding and the one that is actually there.
This is what picking endpoints does. Any noisy series has a pair of years that, quoted alone, describe whatever shape you were hoping for — and the more years you have, the better your odds of finding one. The check is not sophisticated: show every point. If the claim survives seeing the whole series, it was a claim about the data.
Then there is who you count
That series counts the 118 characters you can pull for. The roster actually holds 132 rows; the other 14 are Travellers, who cost nothing to obtain. Count them too and the same question gives a different answer every single year: 47 · 68 · 58 · 59 · 71 · 65 · 77.
The opening figure moves from 38% to 47% and the closing one from 73% to 77%. Nothing has been fudged — every Traveller really is a five-star, so including them really does raise the share, in every year they appear. Neither series is wrong. A post that quotes one of them without saying which one is.
This is the most common way fan statistics go bad, and it almost never looks like an error. Two people can read the same roster, ask the same question, and get 38 and 47 — then disagree at length about Genshin, when what they disagree about is whether a Traveller is a character you pull for.
What the charts on this site do
Every chart here states its denominator in its caption, and the last bar of any time series is drawn hollow when the period is not over. Those two habits are not decoration. The roster’s most recent entry is from August, so this year’s bar is a partial year standing next to complete ones, and reading it as a like-for-like comparison invents a change that has not happened.
Neither habit is clever, and neither would survive being left to memory, which is why they are enforced: the chart data comes from a tested module rather than numbers typed into the prose, and a chart that does not declare its basis fails the build. It is a low bar. Most of getting this right is just refusing to round the awkward parts off.