If your DNA results came back as "Broadly Southeast Asian" while a European friend's came back naming specific counties, that is not because your ancestry is vaguer. It is because the word "Broadly" has a technical meaning, and because the reference panel your sample was compared against is smaller. One company's own documentation puts 73 people behind maritime Southeast Asia and 2,000 behind Wales.

What the percentage actually counts

An ethnicity estimate is not a measurement of your ancestors. It is a count of chunks, which is a different thing from the relationship ranges our shared DNA calculator works with.

Both major companies describe the same shape of process in their published white papers. Your genome is cut into windows. 23andMe uses roughly 1,800 windows of 300 markers each; AncestryDNA uses 1,001. Each window is then compared against a panel of reference populations and assigned to the one it most resembles. The percentage you are shown is simply the share of windows that landed on a given label. Ancestry puts it plainly in its own consumer explanation: if fifteen per cent of the pieces look most like the France reference group, you get fifteen per cent France.

Because the number is defined against a panel, changing the panel changes the number. And a percentage here is not the probability that you had an ancestor from somewhere; it is a count of how the classifier sorted your windows.

What "Broadly" is admitting

The reference populations sit in a hierarchy, from narrow regions up to continental groupings, and each window is assigned to the lowest level the classifier can reach with enough confidence. If it cannot confidently place a window at a narrow level, it moves up until it can.

23andMe states the principle directly: where it cannot confidently determine whether a segment derives from one narrow region or a neighbouring one, it is often still possible to determine the broader region it came from. So "Broadly" is not a hedge added for politeness. It is the classifier reporting the narrowest level it could reach.

A result full of broad labels, then, is telling you something real, but it is about the test's resolution, not your ancestry. It reflects the limited data available for people like you in that company's panel.

The panel sizes, from the companies themselves

This is where the asymmetry stops being an impression and becomes a number. AncestryDNA's 2024 reference-panel documentation gives the numbers:

  • Maritime Southeast Asia: 73 reference samples.
  • Wales: 2,000. Ireland: 2,000. Scotland: 2,000. France, Sweden, Norway, Portugal, Spain and a dozen more: 2,000 each.

23andMe's documentation shows the same shape differently. It carries a single reference population covering four countries — Myanmar, Thailand, Cambodia and Indonesia — built from 77 people. Britain and Ireland alone is built from 1,014. Across the whole panel, Southeast Asian populations account for about 520 of some 14,445 individuals; European populations account for 6,350.

And the consequence is measured in the same documents. In AncestryDNA's own evaluation against nearly 21,000 people of known single origin, maritime Southeast Asia scored 53.9% mean overlap — the third lowest of the 85 regions it assessed. For 23andMe's four-country Southeast Asian population, recall drops to 57.9% at the stricter confidence setting, meaning roughly two in five windows that genuinely come from that population are assigned to something broader instead.

That last figure is the mechanical link between a thin panel and a vague result. When the classifier is less confident about a window, the window moves up the hierarchy, and what you read is a broader label.

Three things that would be unfair to say

The comparison is stark, and there are three ways it can be pushed further than the evidence supports.

The 2,000 figure is evidently a ceiling rather than a measurement — many European regions sit at exactly 2,000, which is what a cap looks like. Those panels are at the limit the company set, not necessarily at their natural size.

European regions are not uniformly sharp either. In the same evaluation, Scotland and the Netherlands both score 69.3%, Cornwall 70.0%, Germanic Europe 73.6%, France 74.8%. Neighbouring European regions bleed into each other for the same reason neighbouring Asian ones do. The point is about panel depth and granularity, not that European results are perfect.

And Southeast Asia is not uniformly thin. Vietnam gets 2,000 reference samples in that panel. So do the Philippines, twice over across two regions. It is maritime Southeast Asia specifically that is thin, not the region as a whole, and a guide that flattened that would be making the same error it is complaining about.

The companies say so themselves

In August 2026, 23andMe announced 137 new genetic groups across Thailand, Myanmar, Malaysia, Laos, Cambodia, Indonesia and Madagascar. The announcement stated that Southeast Asian communities have historically been underrepresented in genetic research, which has often meant broader labels and fewer specific insights for members with roots in the region.

That is this guide's argument, made three weeks before it was written by the company whose results prompted it. It is also a reminder that these figures move: panel documentation carries a date, and the numbers above are the ones published as of that date rather than a permanent property of the products.

What five companies did with one genome

In January 2019, CBC's Marketplace sent DNA from identical twins to five companies — AncestryDNA, MyHeritage, 23andMe, FamilyTreeDNA and Living DNA. No two companies agreed with each other, which is expected given different panels. The striking part is that individual companies did not agree with themselves: the twins, who have the same genome, were given different results.

CBC reported that one twin was assigned nearly ten per cent less "broadly European" ancestry than her sister. The two underlying percentages were not published, so the size of that gap cannot be pinned down further, and it is worth resisting the temptation to state it as a precise figure. Where CBC did publish both numbers, the gap is arithmetic: 28% and 24.7% Eastern European, for two people with identical DNA.

A calibration case worth knowing

Whole-genome sequencing of 177 Singapore Peranakans found Malay ancestry of 5.62%, against 1.08% in the Singapore Chinese comparison group, tracing to an admixture event roughly 190 years ago.

This documented signal in a culturally distinct community is smaller than the disagreement between two consumer results for the same person. No consumer panel carries a Peranakan label at all. If you are looking for something at that scale in a consumer estimate, the instrument is not sensitive enough to find it, and a number that appears to show it is as likely to be noise.

Where this comes from, and what will date

The panel sizes, the evaluation scores and the window mechanics are from AncestryDNA's Ancestral Regions white paper as updated October 2024 and 23andMe's Ancestry Composition white paper as updated December 2020, both retrieved 31 August 2026. The twins investigation is CBC Marketplace, 18 January 2019. The Peranakan figures are from a 2021 paper in Molecular Biology and Evolution.

The panel figures date quickly. That August 2026 expansion means the composition described above has almost certainly changed already, which is why every number here is tied to the date of the document it came from rather than to today. What does not change is the mechanism: your genome is cut into windows, each window is matched against a panel, and a confidence threshold decides how narrow a label you get.

One figure is deliberately missing above. You will often read that some large share of genomics studies has been conducted in people of European descent, and the number that follows is usually unreliable — not because it is invented, but because the denominator gets swapped between studies, participants and samples as the figure travels. One widely cited paper uses the word "studies" in a sentence whose next sentence makes clear it is describing samples. Counting participants, which is the measure that is actually well sourced, the share was 78% in 2018 and 88.2% on a live count in August 2026. It also fell to about 72% in 2012 before climbing back, so this is not a story of steady improvement.