---
title: "How accurate is G25? What the coordinate can and cannot resolve"
description: "Global25 is a projection, not a test — so 'accuracy' has layers: the coordinate's own fidelity, the references it is compared against, and the models built on top. An honest audit of all three."
canonical: https://www.ancestrify.io/blog/how-accurate-is-g25
date: 2026-08-31T09:28:00+00:00
author: "Andi Thomaj"
---

# How accurate is G25? What the coordinate can and cannot resolve

"Is G25 accurate?" gets asked with three different worries behind it: whether the 25 numbers
faithfully describe a genome, whether the comparisons built on them are sound, and whether the
percentages people derive from them are true. Those are different layers with different answers —
and the honest audit is more interesting than either the fan version ("it matches papers!") or
the dismissal ("hobbyist toy").

Background if needed: [what the coordinates are](/blog/what-are-global25-coordinates) and
[where they come from](/blog/eurogenes-global25-history).

## Layer one: the coordinate itself

A [Global25 row](/blog/what-are-global25-coordinates) is a projection of chip-scale genotypes
onto 25 fixed PCA axes. Within its design, it is a *stable, repeatable* measurement: the same
genome projects to essentially the same point regardless of which vendor's chip produced the
file, siblings land near each other, and your row never changes as references grow. The known
sensitivities are inputs, not arithmetic: thin or corrupted files project noisily (pre-flight
with a [file check](/lab/file-check)), [simulated rows](/lab/g25-authenticity) are not
projections at all, and an [unscaled/scaled mix-up](/blog/g25-scaled-vs-unscaled) invalidates
everything downstream while looking like numbers.

The structural limit is compression. Twenty-five dimensions retain the broad and much of the fine
structure of West Eurasia — the space was built to — but **two populations distinct in
allele-frequency statistics can sit close in G25 space**. Drifted isolates and thinly sampled
regions fold worst. Nothing downstream can recover what the projection folded; that is the
ceiling every G25 result lives under, and the reason formal methods
[work from the genotypes directly](/blog/qpadm-vs-global25).

## Layer two: the comparisons

Distances and [PCA positions](/blog/g25-pca-explained) inherit the coordinate's fidelity plus one
more dependency: **the reference panel**. A
[closest-populations ranking](/blog/g25-closest-populations-explained) is exactly as good as who
is on the list — a missing population cannot rank, and its absence is invisible. Averages built
from two or three individuals wobble (an average of N carries roughly 1/√N of individual noise,
which is why our panel curation enforces member floors); era mixing blurs everything. Calibration
numbers from our own reference work give the scale of normal: a typical individual sits around
0.011 from their own country's modern average, and single individuals of a country scatter to
roughly 0.016–0.032 from a well-built calculator for it. Within those tolerances, distance
rankings against curated panels are *reliably reproducible and geographically sensible* — the
genes-mirror-geography result, live in a browser tool.

## Layer three: the models

Here is where "accuracy" usually breaks, and not because the arithmetic fails.
[nMonte-style fits](/blog/nmonte-explained) always return percentages; the
[fit distance has no significance interpretation](/blog/g25-fit-distance-explained); adding
sources improves fit mechanically while
[degrading the story](/blog/g25-source-selection-overfitting); collinear sources split their
signal arbitrarily. A G25 percentage is therefore accurate *as a description of a stated fit
against a stated panel* — and undefined as anything else. The same genome, honestly fitted
against two defensible panels, yields two different true descriptions. That is not G25 failing;
it is what model-dependence means.

The practical calibration, layer by layer:

| Claim | Verdict |
|---|---|
| "My row places me among these populations" | Trustworthy, within panel coverage |
| "I am closer to X than to Y" (same era, clear gap) | Trustworthy; check the gap is not noise-thin |
| "This fit says 48% source A" | A description of one panel, portable nowhere |
| "My 2% component is real" | Unconfirmed — quantisation and noise live at this scale |
| "This proves descent from X" | Never available from coordinates, at any fit |

## Where G25 sits among the instruments

Against a [testing company's estimate](/blog/g25-vs-23andme-ethnicity-estimate): more transparent
and deeper in time, smaller references, no segment view. Against
[component calculators](/blog/how-accurate-are-admixture-calculators): current references and
open arithmetic versus 2012 constructs. Against [qpAdm](/blog/qpadm-vs-global25): G25 is faster,
cheaper, always-answering — and unfalsifiable; qpAdm is slower, allele-frequency-based, and able
to reject a model, which is why our own G25 conclusions are drawn from G25 evidence alone and
claims that must survive argument go to the [formal analysis](/qpadm).

Used inside its envelope — curated panels, era discipline, fits read as descriptions — G25 is the
best exploration instrument the hobby has, and the free stack
([distances](/lab/g25-distance), [admixture](/lab/admixture), [PCA](/lab/g25-pca)) plus
[Ancestry Map](/g25), our worked Global25 analysis, all live inside that envelope deliberately. The inaccuracy people
fear mostly enters through the door the tools cannot lock: the reading.

Terms used here are defined in the [glossary](/glossary).