How the analysis works

The demo is read tick by tick: where every player stood, every kill, grenade, flash and bit of damage, the bomb, and each team's money. That is all the analysis has to work with, and all of it can be checked in the 2D replay.

Code, not the AI, decides what's worth saying. A finding is a pattern that happened more than once in the match, with the rounds it happened in and, where the demo shows one, its cause: a death nobody could trade, a flash that blinded no one, a 1-vs-2 played where pros save.

Each one is compared with pro play in the same spot or situation, from 342 parsed pro matches. A number is only called good or bad after that comparison, so you aren't told something every team does.

A number is only called "well above", "only", "better than usual" or "worse than usual" when chance alone would rarely put it there. We tested this: for a team that played exactly like pros, about 1 match page in 20 still gets such a line. Before we added the test it was 3 in 4. The same rule applies to your stats against your own recent matches (each player card as a whole), to trends on Progress, and to the Compare page. A team's favourite bomb site is named only when a team choosing 50/50 would rarely be that one-sided: 6 of 6 plants, 8 of 9, 10 of 12. A team's position and grenade habits are only listed on maps with enough pro matches to compare them with; on a newer map the page says so instead. A death spot only counts as a pattern after 3 deaths there, and only if pros don't die there just as often. No other opening is suggested on an eco: the ecos that won more were the ones with more money.

Roles and defaults come from where each player stands 20 seconds into each round (pistol rounds left out). A player is the AWPer when they hold the AWP in at least 3 of every 10 of their rounds (and at least 3 rounds), more than anyone else on the team. On CT they're an anchor when they're at the same site in at least 3 of 4 rounds. On T they're the lurk when they're alone in at least a third of their rounds, and the entry when they take at least a third of the team's opening duels. Below those lines they're shown as flex or support, not guessed. "Rounds by situation" puts each team's record next to what pros win in the same situation, and marks a row only when a pro team would land that far out by chance less than 1% of the time.

When your team plays the same map again, last match's team tips are counted again the same way, and shown only when there are enough rounds to tell whether they changed. If your full-buy T plants went mostly to one site in your earlier matches on the map (up to three) and again in this one, the page says so: anyone scouting you can see it too. We don't compare round wins or opening duels between matches: over one match those move by chance more than by anything you changed.

"You got 7 opening kills, all on CT, and 6 of them came with the AWP. Four were at Bombsite A, in rounds 18, 26, 27 and 36." (woxic, Aurora · Mirage)

How each death is read

Every death gets one of these, checked in this order (the first that fits); the others that also fit are kept as notes. Distances are game units.

• Killed by a grenade, no duel: An HE grenade or fire killed them and they fired no shot in the last 3 s. (confidence 90%)
• Died while flashed: The game's own flash event still had them blind when they died. (confidence 80%)
• Killed through a wall or smoke: Shot through a wall, or through smoke while standing still behind it (under 150 units walked in 2 s), without shooting back. (confidence 80%)
• Couldn't shoot back: A grenade, knife or the bomb in hand, an empty magazine, a different weapon 0.3 s before (still switching), or scoped in with the killer within 600 units. Not an AWP or Scout fired in the last 1.5 s and switched away from: that quick-switch is a duel they took. (confidence 90%)
• Died alone: No teammate within 800 units (height counts double) and not traded within 2 s. (confidence 60%)
• Went for a trade with nobody following: A teammate had just died to the same killer, they moved 150+ units in, and no other teammate was within 800 units. Never counted as a habit on its own. (confidence 60%)
• Took a fight alone while ahead: Up two or more players for over 3 s, moved 150+ units toward the killer with no teammate within 500 units, and not traded within 2 s. A T side hunting after the plant doesn't count. (confidence 60%)
• Pushed alone while behind: Down on players, moved 150+ units in with no teammate within 500 units, unless the clock forced it (CT under 12 s with the bomb down, T under 20 s without a plant) or they were the last one alive. (confidence 60%)
• Caught in a crossfire: Hurt by a different enemy in the 2 s before the kill. (confidence 80%)
• Swung wide into a held angle: Moved 120+ units toward a killer who moved 60 or less, still 150+ units apart, and fired or hurt them. (confidence 60%)
• Killed before firing a shot: Fired no shot and didn't hurt the killer in the last 3 s. (confidence 70%)
• Lost a straight duel: None of the above: a fair duel, lost on aim. Counted as good news, never a habit. (confidence 50%)

A kind of death is only called a habit at 8 or more of them over at least 2 of your matches, one of them in this match, and only when pros on the same maps die that way far less often (less than a 1% chance it's luck). Without line of sight, the rules that guess at movement stay at 60% confidence.

Shooting on the move: the share of your rifle shots fired faster than 100 units a second, in a match with 150 or more. It's only named past what 98 pro matches in 100 reach, so a player who moves like pros gets it about 1-2% of the time.

How you move: dropping off high ground (more than 100 units) with an enemy within 1000 units, leaving a spot you'd held 3 s as CT before any contact or the plant, and a teammate within 800 units shot while you didn't move up, shoot or get shot in 3 s. Each is counted per round and named only past what 199 pro matches in 200 reach, at 4 times or more; the card names the one furthest past.

Not built yet: seeing a reload with bullets still in the magazine. The demo only shows an empty one, so only that counts as "couldn't shoot back".

• How pros play it. Under each player's main fix, the same fight (the spot, the side, before or after the plant) in about 330 pro matches: how often pros win it, land a flash on the enemy first, get traded, or have a teammate within 10 m, next to your own count. A row is marked as different only when chance would rarely give your count (under 1%); with fewer than 30 pro fights there, it says so instead of comparing.
• Practice it at the spot. Console lines that put you where you died, facing where you were shot from, built from your typical death there (never an average of positions), with three steps from the fix.
• Try this fix. The fix is counted in your next matches on that map, the same way it was found. From the third, you get "it's working", "it's getting worse" or "no clear change yet", from a test that treats both your before and after as samples (Fisher's exact test, 2.5% each way): a player who hasn't changed gets told either way about 2% of the time.
• The economy's other timeline. A lost force buy is named only when saving would have given four or more players rifle-and-armour money next round and the real round couldn't, and only where pros with that money per player save at least 60% of the time.
• Utility that did nothing. A flash that blinded no enemy. A molly that burned no one, cleared no one out of its fire and held back no one walking at it. Smokes are never judged: a smoke blocks sight, and nobody walking near it doesn't mean it failed.
• What the coach thinks. The role read from the demo and the habits it found are listed per player; your team can mark one wrong or set the role, and the page and the chat go by that from then on.

No AI writes the review: everything above is worked out in code. The one AI feature is the match chat, Ask the coach (paid plans).

• It answers from the match's own data, and before you read the answer, code checks its counts, round numbers, times, places and clutches against the match, with the AI made to fix what doesn't match.
• Every claim links its rounds. One click opens the replay at that moment, so you can see it for yourself.
• Think a line is wrong? Flag it with "This is wrong?" or 👎. Each one goes on the list of what to fix next.

• A demo has no voice. It can't know the call, who was meant to be where, or why someone peeked: it shows what happened, not what was planned. On the Team plan your team can add its own comms recording to a match; it plays next to the replay for your team only, and the analysis doesn't listen to it.
• Compared with pros, not with your level. The baseline is pro matches. A gap to pros is a direction, not a grade.
• Win % is a model. It's trained mostly on pro rounds, from what's visible at that moment (players alive, health, weapons, utility, bomb, time). It doesn't know aim or nerves. Checked on 69 matches it never saw: when it gives CT 0-20%, CT wins 5% of the time; at 40-60%, 51%; at 80-100%, 90%. But in an even fight (same players alive, similar money) it can barely tell who will win. The "Round by round" graph shows its read as each round starts, from both teams' money, weapons and armour.
• One match is a small sample. A pattern in one match can be a bad day. Habits show up once the same thing happens across your matches.
• Their FACEIT maps come from FACEIT's official API. From their last 40 ESEA league matches in a year: each match's map pool and the map played. A map is called out only when chance can't explain it: in a BO1 one map of seven is played, so "they almost never end up on Nuke: 1 of 31" is said, with what chance alone would give beside it. It says where they end up, not who banned what: the other team's bans decide half of it. Nothing from FACEIT is stored.
• The AI can still be wrong. The round links are the ground truth: if a sentence and the replay disagree, trust the replay.