Why your ancestry percentages differ from one company to the next

Same genome, different numbers

The study measured agreement simply: for each category, take the lower of the two percentages and add them up, so 20 and 40 percent Italian share 20 points^[1]. Because the three companies use different categories, the researchers first mapped each company's labels onto a common set, and where 23andMe reported a broad category such as Broadly European, they split it in whatever way maximised agreement.

23andMe and AncestryDNA agreed best, at 84.1 percent on average, and MyHeritage and AncestryDNA worst, at 52.7 percent, with one participant's pair of results overlapping by only 7 percent^[1]. The sample was small, all 42 participants were non-Hispanic white, and every company has rebuilt its panel since. What survives those caveats is the pattern: each company agrees with itself, and the companies disagree with one another.

A percentage is a comparison with a reference panel

Every ancestry estimate starts from a reference panel: people whose DNA is taken to represent a population, chosen for their family origins or for the DNA they share with one another. Your genome is described as a combination of these groups, and the result is reported in the groups' names^[2]. 23andMe draws its panel mostly from consented customers^[3]; AncestryDNA builds its panels from networks of customers who share DNA, annotated with their family trees^[4].

The number of groups changes what you can be told. MyHeritage went from 42 ethnicities to 79 and warned customers that they would now see more ethnicities, each with a smaller percentage^[7]. AncestryDNA went from 107 regions to 146 in 2025^[4]. A finer panel splits one slice into several; your DNA is the same either way.

Two ways to turn a genome into percentages

In the model behind the programs STRUCTURE and ADMIXTURE, a genome is a mixture: there are K source populations, each with its own frequency for every variant, and each person draws some share of their genome from each one^{[9, 10]}. Run unsupervised, the program learns the sources from the people in the study. Run supervised, it is told which reference people belong to which source and uses them to estimate everyone else's shares^[11].

The consumer companies work closer to the chromosome. 23andMe phases the genome into its two parental copies, cuts each into windows of about 300 markers and labels each window with a classifier trained on its panel^[3]. AncestryDNA cuts the genome into 1,001 windows and assigns each parental copy of each window to one of its 146 regions^[4]. The percentage is then a count of windows.

The number of components is a choice

An unsupervised model reports exactly as many components as it is given. ADMIXTURE can rank values of K by cross-validation^[11], which is not the same as counting peoples. A study of Uralic-speaking populations ran ADMIXTURE on 1,800 people from 111 populations at every K from 3 to 20, 100 times each^[12]. At low K, Uralic speakers looked like their geographic neighbours. From K = 9 a new component appeared, found mainly in Uralic speakers and peaking among the Ob-Ugric and Samoyed speakers of western Siberia, and K = 10 had the best cross-validation score^[12]. Ask for eight components and that ancestry is folded into its neighbours; ask for ten and it has a name.

Repeated runs at the same K do not always agree either. Independent runs can arrive at substantially different solutions, which is why tools such as Clumpak exist to sort replicate runs into modes and summarise each one^[13]. The Uralic study judged its results by the top tenth of its runs, ranked by likelihood^[12].

A component is not proof of a mixing event either. Lawson and colleagues simulated three histories: recent admixture, admixture with a population nobody sampled, and a recent bottleneck with no admixture at all. The bar plots came out almost the same^[8]. Strong drift, of the kind a small founding population goes through, can earn a population its own component, and its neighbours then appear to carry some of it.

Finnish DNA on someone else's panel

Finns are a textbook case of a drifted population. In a 2008 comparison of almost 250,000 markers, Eastern and Western Finns differed from each other more than Germans did from the British: an FST of 0.0032 against 0.0005^[14]. A STRUCTURE analysis of the same samples found a cluster dominated by Eastern Finns, while Germans, British and Utah residents of northern European descent shared one^[14]. In 2,376 Finns placed by their parents' birthplaces, Kerminen and colleagues found regional structure stronger than in the UK, with a main border following the 1323 treaty line of Nöteborg^[15]. Our guide to Finnish DNA covers that history.

Drift gives Finns a recognisable signature, and a panel with a Finnish group catches it. 23andMe states that because Finns are so genetically distinct, they have their own reference population, and in its held-out tests 95 percent of Finnish DNA was labelled Finnish^[3]. AncestryDNA's single-origin test people from Finland received 98.5 percent Finland on average^[4]. The neighbours fare less well.

Russian is the hardest of these for both. Russian DNA that 23andMe fails to call Russian is almost always assigned to its single Belarusian, Polish & Ukrainian population^[3]. Neighbours share much of their history, and the line a panel draws between them is the panel's own.

Finns also share something with peoples farther east. Modern Finns, Saami, Mordovians and Russians all share an excess of alleles with the Nganasan of northern Siberia, a signal ancient genomes from the Kola Peninsula date to at least 3,500 years ago^[16]. Tambets and colleagues estimated this Siberian-related component at up to a third of the genome in the Saami and in Volga–Ural peoples, below 10 percent across most of north-eastern Europe, and 5 percent in Estonians^[12]. On a panel with a Saami group, a Finn's share of that signal has a nearby label to land on. Saami languages were spoken across Finland before Finnish arrived, and the ancestors of Finnish speakers may have mixed with the people already living there, the likely ancestors of today's Saami^[16].

Salmela and colleagues saw the limit from the other side. They noted that a small degree of Saami admixture had been observed among Finns, and that their own data could not detect it in the absence of Saami reference data^[14]. A label appears only when the panel holds a group for it.

Those figures count users who see a label at all, not the size of the slice^[17]. A Baltic or East European slice on a Finnish result does not have to mean a Baltic or Russian ancestor in recent generations; it can be the part of a Finnish genome the panel's Finnish group did not explain, handed to the nearest groups.

Our test: take the Baltic Finnic group out of the panel

We can show the closed-list effect on our own panel, with the report's own estimator. We drew twenty genomes at random from the allele frequencies of its Baltic Finnic group, 142 people from Estonian, Finnish, Karelian and Veps samples, at all 71,550 markers, and fitted each three times: against all 31 groups, without Baltic Finnic, and without both Uralic-speaking groups, Baltic Finnic and Volga–Ural.

With the Finnic group in place, the genomes read 98 percent Baltic Finnic on average. Without it, the whole genome was still assigned: they read 43 percent East European, 33 percent Northwest European and 24 percent Volga–Ural. Remove the Volga–Ural group too and they read 54 percent East European, 39 percent Northwest European and 6 percent Siberian. That last figure sits within the under-10-percent range Tambets and colleagues estimated for the Siberian-related component across most of north-eastern Europe^[12]; here it surfaced only because the panel had no closer group to assign it to.

Three cautions. A genome drawn from a group's average frequencies is an idealised member of it, and real people are less tidy. The brackets show how far the result moved between twenty such genomes: 12 points for the East European share. And the test ran only the fitting step, without the card's resampled ranges, regions or panel-fit check.

Statistical labels, not a family tree

Genealogical ancestry counts the people in your family tree. Genetic ancestry follows only the paths by which your DNA was actually inherited, through a small share of those people^[2]. An ancestry percentage is neither. It measures how closely your genome resembles today's reference groups, labelled with places because the reference people come from places.

So a 25 percent slice does not mean one grandparent, and no slice carries a date. In AncestryDNA's own test, customers with all four grandparents born in Norway averaged about 95 percent Norway, and those with all four grandparents from northern Germany about 75 percent Northwestern Germany with around 10 percent the Netherlands^[4]. A family from one place still shows the neighbours, because neighbouring populations share ancestry. Haplogroups answer a different question, about a single line of descent, and our haplogroup guide explains where they fit.

What the Aimosti report shows

The report fits one mixture, with the same likelihood ADMIXTURE uses, against fixed allele frequencies for 31 reference groups. 30 are pooled from present-day people in the Allen Ancient DNA Resource, from its release on the Human Origins array of about 600,000 markers^[18]; the Ashkenazi Jewish group is built from published allele frequencies^[6]. With the reference frequencies fixed, the fit has a single best answer, so the same file gives the same estimate on every run.

The card groups the 31 components into 17 regions, sets of groups too close for one file to tell apart reliably^[6]. The Finnish region holds three:

Each region gets an estimate and a range, the spread of the estimate over 100 resamplings of the markers, widened at the top by the share the fit could not place^[6]. Inside a region the card prints ranges only, rounded outward to 5 points, such as about 35–50 percent Baltic Finnic, and names no point estimate. A region is named only when its printed lower bound is at least 1 percent. Whatever could not be placed appears as an Unresolved row. When the panel cannot reproduce a file's genotypes as well as it usually does, the card says so above the figures.

We test those ranges two ways^[6]. On genomes simulated from the panel's own frequencies, the generous case of our experiment, a region's range held the true value 97 percent of the time, and the ranges inside a region 91 percent. On genomes stitched together from real reference people left out of the panel, a region's range held 79 percent of the time at 23andMe chip density and 67 percent on whole genomes.

A chip's ranges are wider than a whole genome's, because it reads fewer markers. A file with fewer than 1,000 usable markers gets no composition, and the card says so.