Spit into a tube, post it off, and a few weeks later a website tells you that you are 38 per cent Scottish, 12 per cent Scandinavian and a sliver Iberian. The numbers feel like a verdict. They are closer to a weather forecast: a probabilistic estimate built from a model, revised as the model changes, and only as good as the data behind it.
Start with what the kit actually reads. Companies such as AncestryDNA and 23andMe do not sequence your genome. They genotype, using a chip that inspects several hundred thousand to roughly 700,000 pre-selected positions where human DNA is known to vary between individuals. These positions are single-nucleotide polymorphisms, or SNPs, single letters in the code that differ from person to person. Your three-billion-letter genome is sampled at those chosen points, and the pattern of letters you carry becomes the raw material for everything that follows.
The ancestry breakdown then comes from comparison, not recognition. The company holds a reference panel: DNA from people with deep, documented roots in a particular region, often four grandparents born in the same place. Your SNP pattern is compared, segment by segment, against those panels, and an algorithm assigns each stretch of your chromosomes to the reference group it most resembles. Add the segments up and you get percentages. Because it is a statistical fit rather than a lookup, the same raw data can yield different numbers when a firm expands its panels or rewrites its model, which is why users periodically log in to find their heritage has been quietly recalculated overnight.
Why siblings do not match
The clearest proof that these figures describe DNA rather than identity is the sibling test. You inherit half your DNA from each parent, but which half is decided by recombination, the shuffling of chromosomes that happens when eggs and sperm form. Two children of the same parents draw different hands from the same deck. If a grandparent's contribution lands in one sibling's inherited segments but not the other's, the kits will report different ancestry for people who share both parents. The estimate is honest about the segments you personally carry; it was never measuring an essence that siblings must hold in common.
Deep ancestry is read by separate machinery. Mitochondrial DNA passes intact down the maternal line, and the Y chromosome down the paternal line, so both can trace a single lineage back through many generations and sort people into haplogroups, the branches of the human family tree. But each traces only one thread. Your mitochondrial result says nothing about your father's mother, and a woman carries no Y chromosome to test at all. A haplogroup can be tens of thousands of years old and shared by millions, so it locates a branch, not a birthplace.
The database is the quiet author of the story. Reference panels are dense for northern and western Europe, where early adopters and academic cohorts clustered, and comparatively sparse for much of sub-Saharan Africa, Central Asia and Indigenous groups across the Americas and Oceania. That imbalance is why a British user may see "Aberdeenshire" while someone with West African heritage gets a single broad region spanning several modern nations. The granularity is a function of who has already been sampled, not of how mixed anyone's ancestry is.
Two other limits deserve flagging. Percentages assume today's populations are stand-ins for the past, yet people have always moved, so a label like "Scandinavian" reflects present-day Scandinavian DNA rather than a Viking certificate. And the relative-matching feature, which links you to strangers sharing DNA segments, is often the most concrete thing a kit offers precisely because a shared segment is a fact, while a percentage is an inference.
Then there is the data itself. Under UK GDPR, genetic information is special-category data with heightened protection, yet once a sample is analysed the resulting profile sits on a company's servers and is only as secure and as private as that firm's policy and solvency allow. Police forces have used consumer genealogy databases to identify suspects through distant relatives who never tested, and a spit kit implicates cousins who never consented. When a testing company is sold or folds, its most valuable asset is the database, and the terms you agreed to may outlive the brand.
None of this makes the kits worthless. They are a genuine window on the segments you carry and the relatives you can find. The mistake is reading a modelled estimate as a fixed fact about who you are, when the honest translation of "38 per cent Scottish" is closer to "the model's current best guess, given today's reference panels, for the share of your sampled DNA that resembles people it labels Scottish."