Methodology

There is no Steam API that lists TF2 players, so the corpus is discovered two ways: Steam store review authors (a casual-skewed seed pool) and friend-graph expansion from active players with 1,666+ hours. Friend-graph sampling over-represents socially connected and veteran players — the 2022 style.tf data famously showed a 1% Golden Frying Pan rate on sniper for this reason. Every player's discovery source is recorded, and sample sizes are shown on every page.

Loadouts come from GetPlayerItems, which only exposes the currently active loadout preset per class (A/B/C/D presets are invisible). Stock weapons never appear in backpack data, so an empty weapon slot is counted as the stock item. Troll loadouts are counted as-is; we assume swaps roughly cancel out.

Engineer's destruction PDA and toolbox occupy different physical equip slots than the construction PDA — we remap them all into one semantic "PDA" slot so slot filters mean one thing (4 = sapper, 5 = PDAs, 6 = watch).

The Steam API treats renamed/strange stock weapons ("Upgradeable"), festives, botkillers, australiums, and warpainted weapons as separate items. By default we merge anything functionally identical: items sharing a weapon class and gameplay attributes after resolving items_game prefab inheritance, ignoring a curated list of cosmetic-only attributes (kill counters, paintkits, killstreak effects, viewmodel offsets…). Two community-consensus reskins with technically-real quirks are merged manually: The Original (centered projectile) and The Rainblower (pyrovision). The unmerged view is one toggle away, and every merged row expands to its variants.

The all-class melees (Frying Pan, Conscientious Objector, …) merge into each other globally, and additionally fold into each class's stock melee on class-specific views — a Pan is a Bat for a scout — but not on the Any-class view, where transitively merging Pan→Bat→Bottle→… would collapse all stock-stat melee into one row and erase the "how many people pan" signal.

For a specific class, usage = players equipping the item on that class ÷ all players whose loadout we can see. For Any class, an equip is counted per class-slot opportunity: total equips ÷ (players × classes that can use the item). For slot-filtered Any-class views, the denominator uses classes that equip the item in that slot — the Panic Attack in the primary slot is an engineer-only opportunity even though four classes own it. Merged groups use the union of their members' classes.

Every player contributes all 9 class loadouts, whether or not they play the class — this matches how the game presents loadouts and how style.tf counted in 2022.

Roughly half of encountered accounts hide their backpack or game stats. The inventory endpoint signals a private backpack with a bare HTTP 503 — but the TF2 Game Coordinator behind it also fails transiently under load (community measurements show ~40% failure in busy periods), so we only classify 503 as "private" after it persists across three spaced retries, and transient errors are retried without being recorded. Live rates are on the health page.

Deaths do not exist in Steam's TF2 stats — the schema defines a deaths field, but the game client never uploads it — so no true K/D is possible; we show kills/hour and points/minute instead. Per-map stats only include playtime, not points, so per-map performance isn't derivable. Casual win/loss and MMR exist only in each player's private Game Coordinator data. Per-player stats accumulate across all servers, not just official casual — there is no server-type filter in Steam's player stats.

"Points per minute with this weapon" averages the lifetime per-class rates of players who currently equip it (10+ hours on the class, 3+ players minimum). This is correlation, not causation — lifetime stats span every loadout a player ever used, so read it as "the kind of player who equips X", not "X makes you better".