Player and puck positions at 10 Hz before every goal of the 2025-26 season (8,521 goals), reconstructed from the NHL's sprite tracking data.
Clips are not a fixed length. The footage before the shot runs a median of 9.4 seconds, with 90% of clips between 8.1 and 10.4, and each one continues a little past the shot to the puck crossing the line.
A clip carries the entities the tracking supplies for that stretch: the puck, both goalies, and typically ten skaters. That is not a guarantee of who was legally on the ice: about 37% of clips carry more than ten skaters, mostly players at or near the bench mid-change, and a handful carry no goalie at all.
This site is not affiliated with or endorsed by the NHL. Video clips, where they are available, are embedded from the NHL's own player as served.
The corpus is one row per goal and nothing else. Saves, blocks and misses are not counted anywhere. There is no shot-level table behind this, and therefore no denominator.
That is a statement about the counting, not about the clips. Saves, blocks and missed shots happen constantly inside the clip you are watching, and a goal scored off a rebound means the save that created it is right there on screen. What is missing is any record of the ones that did not end in a goal.
So nothing here is a rate, a share of attempts, or a save probability. A count of 40 means forty goals looked like this, not that anything happened 40% of the time, and not that this is a common or an uncommon way to shoot.
The goalie's sight line to the puck spans an angular width. Screening is the part of that width covered by other bodies at the release frame, reported in degrees, using a 1.5-foot body radius. That radius is calibrated rather than assumed: sustained player-to-player contact in this tracking settles at 2.0–2.5 feet of sensor separation, and 1.5 allows for a sight line passing a screener at hip height rather than shoulder height.
It is horizontal occlusion only. The tracking has no vertical dimension, so a player crouched, sprawled or leaning has the same footprint as one standing, and a puck lifted over a screen is not distinguished from one through it.
It is a Voronoi tessellation of the ice over the skaters' positions: every patch of the sheet is coloured for whichever skater is closest to it, and the boundaries are the lines where two skaters are equally close. It is recomputed every frame, so the cells move as the players do.
In hockey terms it answers one question continuously: who is closest to this piece of ice right now? That is worth having because a goal that looks like a scramble often turns out to have a large piece of undefended ice in it that somebody found. The cell that opens up before the pass is usually the reason the pass worked.
Ice further than 25 feet from every skater is left uncoloured rather than handed to whoever happens to be nearest. Twenty-five feet is about two seconds of skating at a median pace, and it is also the distance from centre ice to the blue line, so it is a length that can be checked by eye. Without that cap a single player on an empty half of the rink would be shown owning it, which claims a contest that is not happening.
Two things it is not. Goalies are excluded, since a crease is not contested ice. And it is plain straight-line distance: how fast a skater is moving and which way he is already facing are not in it. Football analysts use the same geometric idea as a starting point, but their pitch-control models go further and weight the area each player controls by velocity and how quickly they could actually get there; this does not, and the two should not be read as the same measurement.
Corrections and questions: hello@visualpuck.com. This is an early version and reports of anything that looks wrong are genuinely useful.