G25 is not just PC1 and PC2
Most people's first contact with Global25 is a simple PCA plot with two axes: a horizontal one that roughly separates northern from southern Europe, and a vertical one that roughly separates eastern from western Eurasia. That is PC1 and PC2, and they are genuinely useful, but they are also only 2 of the 25 numbers in every G25 set. If you have ever wondered why two people can sit almost on top of each other on a PC1/PC2 map and still get noticeably different admixture models, the answer is hiding in the other 23 axes.
What PC1 and PC2 actually capture
PC1 and PC2 are the two directions along which the reference populations used to build G25 vary the most. In practice, across a worldwide dataset, that ends up being dominated by very large, very old splits, the depth of Neolithic farmer versus hunter-gatherer ancestry in Europe, and the broad east-west Eurasian gradient. They explain a large share of the total variance, which is exactly why they are the ones plotted by default, but "largest share of variance" is not the same as "the axis that matters for your specific question."
PC3 onward: where the finer regional signal lives
PC3, PC4 and the axes beyond them capture progressively finer, more regional splits, the kind of thing that separates a Croatian profile from a Serbian one, or an Anatolian Roman-era sample from an Italian one, differences that would be completely invisible on a PC1/PC2 plot because both groups sit in almost the same place on those first two axes. This is exactly why a period- or region-specific model (like the ones we run for Balkan or Mediterranean ancestry) needs the full 25-dimension distance calculation, not just a 2D plot, to tell closely related populations apart.
Why two similar-looking profiles can model very differently
If you paste two people's coordinates into a simple 2D PCA viewer and they land in almost the same spot, it is tempting to assume their ancestry is nearly identical. But a full 25-dimension NNLS model, the kind that actually estimates admixture percentages against a fixed set of reference populations, will pick up differences on PC3 through PC25 that a 2D plot simply cannot show. Two people who look like near-neighbours on PC1/PC2 can come back with meaningfully different percentage breakdowns once the deeper axes are taken into account, and that is not a contradiction, it just means the interesting part of the comparison was never on the plot to begin with.
Seeing it for yourself
If you want to compare two coordinate sets (yours against a relative's, or two different calculator outputs for the same person) and see exactly which of the 25 axes are driving the difference rather than guessing from a 2D plot, our free Compare Coordinates tool lays out every axis side by side and sorts them by how much they actually disagree. It is usually a short list, most axes agree closely, and that is exactly the point: once you can see which two or three axes are doing the work, the "why don't these match" question usually answers itself.