A shot's expected goals (xG) value is a model's estimate of how often a chance like it becomes a goal, so two models can disagree about the same shot. When RubiScore shows a match xG figure that differs from one quoted elsewhere, the gap usually comes from model design and data collection rather than from an error.
This listicle explains seven reasons xG numbers differ between providers, how large each effect tends to be and how to compare figures fairly. Match pages with shot data and expected goals are available at https://rubiscore.com, and the same checklist applies to any public xG source. RubiScore treats xG as an estimate with a method behind it, which is the right starting point for any comparison.
xG is not measured the way goals are counted. A goal is an event; xG is a probability produced by a statistical model trained on historical shots. Every provider decides which shots to train on, which features describe a shot, how to handle unusual situations and how its event data is collected in the first place. Each choice is defensible, and each moves the output slightly.
Disagreement is therefore normal. The useful question is not "which number is right?" but "which differences matter for the conclusion I am drawing?" For most matches, reputable models agree on the broad story — who created more chances and better ones — while differing by a few tenths of a goal. A reader who knows where the gaps come from can separate a real disagreement from a methodological one. On RubiScore, a match xG total is best read in exactly that way: as a summary of chance quality under one consistent method.
The simplest models use only shot location, meaning distance and angle to goal. More detailed models add body part, the type of pass that created the chance (cross, through ball, cut-back), whether the shot followed a dribble, whether it came from a set piece or open play, and whether it was a first-time attempt.
Each feature changes the estimate for a subset of shots. A header from eight metres is a much lower-probability chance than a shot with the foot from the same spot, so a location-only model will overrate many headed chances. When two providers disagree about a match full of crosses, feature choice is often the reason.
A growing number of models use freeze-frame or tracking information: where the goalkeeper stood, how many defenders were between the ball and the goal, and how much pressure the shooter was under. A shot from the penalty spot with an empty net ahead and a shot from the same spot through a crowd of bodies look identical to a location-based model. They look very different to a model that can see the defenders.
Models with this extra context tend to spread values more widely. Open goals go higher and crowded attempts go lower, so match totals can move in either direction depending on which situations dominated.
A model learns conversion rates from the shots it is trained on. A provider that trains mainly on top European leagues learns the finishing and goalkeeping standards of those leagues. Another that trains on a wider set of competitions, or on recent seasons only, learns slightly different base rates.
This matters most when a model is applied outside the data it learned from: a lower division, a women's league with different shot patterns, or a competition where goalkeeping standards differ. Some providers build separate models per competition or adjust for league strength; others apply one global model everywhere. Neither approach is wrong, but they produce different numbers for the same match.
Most models give every penalty a fixed value based on historical conversion, typically somewhere around three in four. The exact figure varies by provider and by the period used to estimate it. Some sources quote xG including penalties, others publish non-penalty xG (npxG) as the headline figure, and some show both.
A single penalty can therefore explain most of the gap between two match totals. Before comparing, check whether both figures include penalties. If the match had one, subtract it to compare open-play and set-piece chance quality directly.
When a goalkeeper saves and the rebound is shot again, there are two shots but only one possible goal from that sequence. Adding the two xG values can produce a total above what the sequence could realistically yield. Some providers sum every shot anyway. Others treat the sequence as one chance with a combined probability: one minus the chance that every attempt in the sequence fails.
The difference is small in a typical match but noticeable in a game full of goalmouth scrambles. It also changes how teams that generate repeated attempts inside the box are summarised.
Most xG still begins with event data, a record of where each shot was taken and what happened. Collection methods differ between providers — live tagging, post-match video review or automated systems — and so does the precision of the coordinates. A shot logged half a metre closer to goal, or an attempt classified as a header rather than a shot off the shoulder, changes its value.
Definitions also differ at the edges. Whether a blocked attempt counts as a shot, whether a deflected cross that loops towards goal is logged as an attempt, and how own goals are treated (normally excluded from a player's xG) are all provider decisions. These hidden choices are why two sites can show different shot counts as well as different xG.
Providers retrain models as more data arrives, and they sometimes revise past values when they do. A figure quoted in a season review may not match the number shown on the same provider's site a year later. Models are also calibrated, which means checking that chances rated at 0.10 really do become goals about one time in ten across thousands of shots. Small calibration differences shift totals systematically, making one source slightly higher or lower than another across every match.
That systematic offset is why mixing sources is risky. A team's season xG from one provider and its rival's from another can create a gap that exists only in the methods.
For a single shot, the disagreement can be large in relative terms, especially for headers, chances under heavy pressure and rebounds. For a whole match, differences usually shrink because errors partly cancel out. Over a season, totals from reputable models normally track one another closely, although teams with unusual styles — heavy crossing, frequent set pieces or many penalties — show bigger gaps.
A useful rule of thumb is to treat match-level differences of a few tenths of a goal as methodological noise and to look for larger, consistent gaps before concluding that two sources tell different stories.
A short checklist prevents most misreadings:
A hypothetical example shows the point. Suppose one source gives a match as 1.9 xG to 0.8 and another as 1.6 to 0.9. Both say the same thing — one team created clearly better chances — even though the numbers do not match. If one source included a converted penalty and the other did not, the real methodological gap between them may be close to zero.
The existence of several xG models is a strength rather than a weakness. It reminds readers that xG estimates chance quality rather than recording a fact about the match, and that the method matters. Platforms such as RubiScore publish expected goals beside shots, shots on target and the match timeline, so a figure can always be checked against the events that produced it.
The most reliable habit is consistency: pick one source, understand its method and compare like with like. When two providers agree on the direction of a match, the conclusion is strong. When they disagree sharply, the seven reasons above usually explain why, and they give RubiScore readers a quick way to decide whether a difference is about the game or about the model.