Solo Ranked Recon
Division 1B
1B's defining feature is an inversion: the top four seeds hold the ELO, the bottom three seeds hold the win-rates. Ralius (seed 6) and Topotic (seed 7) carry 76.0% and 75.4% lifetime — the two best rates in Division 1 — on samples a fraction the size of the players seeded above them.
India divides 1B (three list it a strength, two a weakness). Russia and Indonesia are shared strengths — four players each — so expect no free points there. And note Ralius's 258 km average distance: the sharpest accuracy figure in the entire league.
The cleanest data in the league: all eight carry live current ratings, geographic and per-mode records are complete. Sample-size watch-points are Ralius (818 duels) and Topotic (1,797) — real, but lighter than their cluster-mates'.
State of the field
A snapshot before the detail: where the eight sit on rating, experience and volume. The headline is the gap between seeded and current form — this is a field well off its collective best.
Seeded ratings average 2256; current form averages 1799 — the division sits 457 points below its enrollment ratings. Every player carries a live current rating, so no imputation is involved. The axis is scaled to this division's own range.
Tick marks are individual players. Mean ~6.5 years, but read it loosely: account age is only a proxy for GeoGuessr experience, and it breaks for fresh accounts of veteran players. Median lifetime volume is 4,928 ranked duels.
Geographic signal map
Every player against every country that appears in a top-3 or bottom-3 list. Warm = a ranked strength, cool = a ranked weakness, intensity = list position. Columns are ordered by how much the field cares about that country. Hover any cell for the value.
Ordinal signal only: a blank cell means the country isn't in that player's top or bottom three — not that they're necessarily neutral there.
Contested ground
Countries that are a strength for some players and a weakness for others — the regions where a matchup is won or lost rather than shared.
Field patterns
Where the field collectively excels and bleeds. Shared weaknesses are the soft ground worth drilling — common blind spots across skill tiers.
Common strengths
- RU — Firma, greguiz, Elements, Topotic
- IN — ALFA, rileyrhino, Ralius
- MX — ALFA, Firma, greguiz
- ID — Firma, rileyrhino, Ralius, Biquette
- BR — rileyrhino, greguiz, Elements
Common weaknesses
- CA — ALFA, Ralius
- AU — greguiz, Biquette
- CL — Elements, Topotic
- IN — Firma, Biquette
- AR — rileyrhino, Elements
Shared soft spots (2+)
- CA — ALFA, Ralius
- AU — greguiz, Biquette
- CL — Elements, Topotic
- IN — Firma, Biquette
- AR — rileyrhino, Elements
- NZ — Firma, greguiz
- CO — greguiz, Ralius
Player clusters — win-rate model
Our primary segmentation, built on lifetime per-mode win-rates over hundreds-to-thousands of ranked duels per player — measured outcomes, not imputed ratings, so no synthetic data props it up. Features: current ELO, lifetime overall and per-mode win-rate, a win-rate-derived mode lean, accuracy and volume. Ward and k-means agree. Two players resolve as singletons for real reasons (below). A second, rating-based model in section 06 triangulates this one. At n=8 these are reading aids, not hard tiers — silhouette ≈ 0.31, a real but soft structure.
Seeds 6–8, and the outcomes say that's wrong: Ralius 76.0% (818 duels, 258 km), Topotic 75.4% with an 89.4% moving rate, Biquette 58.5% with a hard moving tilt (70.9% moving vs ~47% restricted). Low volume is the only caveat — and see the note below on why it matters less here than usual.
Seeds 1–4. Big samples (1,900–10,000 duels), 52–60% outcomes, ratings that largely held through to now. greguiz is the form pick of the four: 59.7% with the smallest gap between enrollment and current.
A singleton, and not flatteringly: 49.5% lifetime — the only sub-50 in Division 1 — a 1,240 km average distance, and a current rating 370 below enrollment. Both models isolate him. The division's clearest pressure point.
How to read the two cluster views
The clustering runs in a 6-dimensional feature space; these two plots are different windows onto it. The PCA map flattens that space to its two most informative axes; the dendrogram shows the order in which players merged into groups. Same data, complementary views.
PCA map. Each player projected onto the two directions of greatest spread in this division's data. A principal component has no fixed meaning — it is whatever mix of the six features varies most here — so the axis readings below are derived from this division's own loadings, not carried over from any other division. PC1, the horizontal axis (48% of the variance): moving right means a higher lifetime win-rate, a moving-mode lean and more uneven win-rates across modes. PC2, the vertical axis (33%): moving up means a wider average guess distance and a lower current rating. Two of the six features deserve definition, since both are computed from win-rates rather than from how often a mode is played: mode lean = mean(NM%, NMPZ%) − Moving%, i.e. which way a player's results tilt; mode spread = the standard deviation across their three mode win-rates, i.e. how lopsided they are regardless of direction. A player who wins 74/75/71 across the modes and one who wins 49/52/47 both score flat on these; a 65/54/33 profile scores strongly on both. Distance on the plot ≈ overall dissimilarity. Together the two axes hold 81% of the variance in the six clustering features.
Dendrogram. Built bottom-up: each player starts alone, then the two most similar join, repeatedly, until all are one. Bar height = how different two groups were when they merged — low joins (ALFA+rileyrhino) are near-twins; high joins happen late and grudgingly. Reading down from the top, the first split separates the strong tier from everyone else; cut the tree at any height to get that many clusters. Dot colour marks final cluster membership.
Two models, triangulated
We trust two models, built on independent signals. The win-rate model (section 04) asks who beats their opponents, from measured lifetime outcomes. The ELO model asks where the ladder places them now, from current per-mode ratings with mode-lean computed only from real (non-imputed) data — the honest rating view, silhouette 0.29. Where they agree, a finding is robust; where they diverge, that disagreement is itself signal worth a second look.
| Player | Win-rate model | ELO model |
|---|---|---|
| Biquette | The under-seeded hitters | Lower rated |
| Ralius | The under-seeded hitters | Higher rated |
| Topotic | The under-seeded hitters | Higher rated |
| ALFA | The ladder holders | Higher rated |
| Firma | The ladder holders | Higher rated |
| greguiz | The ladder holders | Higher rated |
| rileyrhino | The ladder holders | Isolated |
| Elements | Adrift | Lower rated |
What the divergences mean. Normally a low-volume win-rate is an unconfirmed rumour. What makes 1B unusual is that the ELO model already agrees: it groups Ralius and Topotic up with the top seeds, because their current ratings (2216, 2241) back the win-rates. When both the outcome record and the live ladder say the same thing about a bottom seed, the seeding is stale, not the models. rileyrhino is the reverse case — isolated by ELO on a sagging current, mid-pack on outcomes.
Geographic similarity
Where each player's competence is concentrated — thread weight is the overlap in best-country lists. GeoGuessr ranks these countries within each player, so the lists say where someone is strong relative to their own baseline, not how strong they are.
Reading it
Thick threads mean two players' strengths sit in the same places — on a map drawn from those countries, neither gets an edge from familiarity. Thin or absent threads mean their competence is concentrated elsewhere, so map choice cuts between them. Because each player's top-3 is ranked against their own record, this maps the shape of a player's geographic knowledge with their overall level divided out: a Division 8 player and a Division 1 player can share a thread while being nowhere near each other in strength. It is a map of where someone is strong, not how strong they are — which is why it is kept separate from the clustering, and why a thread is a matchup cue rather than a prediction.