Expected goals has become the most quoted number in football analysis and one of the most misread. Understanding what goes into it explains both its usefulness and its limits.
The number describes the shot, not the finish
The model assigns each shot a probability based on characteristics such as distance, angle, body part used and the type of pass that created it.
That probability reflects how often shots with those characteristics have historically been scored, by all players, not by the one taking it.
The metric therefore measures the quality of the opportunity created rather than the quality of the attempt at it.
Single matches contain too few shots
A match typically produces a small number of shots, and a small sample of low-probability events is dominated by chance.
A team can accumulate several good chances and score none, which the model correctly describes as unlucky and which the scoreline records as a defeat.
Quoting the figure from one match as evidence that the wrong team won misapplies a tool built to describe tendencies across many matches.
What the model cannot see
Most versions do not fully capture defensive pressure, the position of the goalkeeper, or the precise arrangement of defenders at the moment of the shot.
Two shots from identical positions can face completely different situations, and the model may score them the same.
It also says nothing about the chances that were not taken, so a team that declines good opportunities registers as having created little.
Its value is in prediction rather than description
Over a season, the chances a team creates and concedes describe its underlying performance more stably than results do, because finishing varies more than chance creation.
Clubs use it to identify sides whose results are running ahead of or behind their process, which informs recruitment and managerial decisions.
That predictive use is where the metric earns its place, and it is a different question from who deserved to win on a given afternoon.
Different models give different numbers
Because each provider builds its own model on its own data with its own variables, published figures for the same match frequently disagree.
The differences are usually small but large enough to reverse a close comparison, which is a reason to treat any single published figure cautiously.
Comparisons are only meaningful within one model, and mixing sources produces the kind of contradiction that has made the metric an easy target for criticism.