How the BANG Score works

The BANG Score is one number, 0 to 100, for how much game you get per dollar. A 41 is a 41. You can argue with it, and people do. That argument is the point.

Reviews set the ceiling

The base of the score is the Steam review rating, shrunk toward the middle when a game has few reviews so a 100 percent from nine people cannot outrank a 96 percent from ninety thousand. A game that reviews badly cannot climb, no matter how cheap it is. Price moves you down inside that ceiling. It never lifts you past it. Reviews lead, length follows.

This ordering matters more than any single weight. An early version led with dollars per hour, and every short, brilliant game scored like a ripoff while padded grinds floated to the top. Flipping reviews to the front fixed it in one change.

Price is measured against real length

Value here is price weighed against content. The content number is a real length, taken from HowLongToBeat wherever a game can be matched. A free game has no price to divide, and a multiplayer or roguelike game has no ending to measure, so for those the value read is how many people own and play it instead, from SteamSpy's estimates, which are old (see the last section). Short-by-design games get a per-genre floor, so a five-hour narrative game is measured against other narrative games instead of against a sixty-hour RPG. A campaign game you keep replaying, in co-op or with mods, gets credit for that replay depth instead of one short run, when its completionist time runs far past the main story.

When a game has no verified length at all, the score says so and carries a dock, because the biggest input is then a guess. A verified length removes that dock.

What raises and lowers a score

Every game page shows the drivers, sorted into what raises the score, what lowers it, and what sits neutral. The common ones:

• Reviews. The gate. Strong reviews unlock the top band, weak reviews cap it.
• Value. Price per hour of content, the largest single mover for a paid game with an ending.
• Reach. How many people own it, by SteamSpy's estimate. For a free game and a game with no ending it is the value read; everywhere else it is context the score does not use.
• Finish rate. The share of players Steam counts at the game's ending achievement, read from Steam global achievements. This is a bonus that reserves the very top of the scale, and a game with no ending never gets it.
• Monetization. Loot boxes and pay-to-win apply a haircut, and a game documented as selling nothing extra earns a small credit.
• Refunds. When an unusual share of a game's Steam-purchase reviews come from buyers who refunded it, a small dock.
• AI content. When the developer has disclosed AI-generated content on Steam, a heavy dock.

Reading the grades on a game page

Every input in a game page's breakdown gets a grade, S to D, read from the buyer's side, so no monetization at all earns an S. The score is multiplicative, so an input only counts as raising it when the input sits near the top of its own scale. Anything with real room left is holding the score back, because that room is the gap to 100. ▲ marks a lift, ▼ a drag or a dock, ? an input the site does not have, and • a fact this genre's score does not use.

The player quotes on a game page are Steam reviews: the most helpful of all time, plus the newest that clear the same quality bar, each with the hours the player had put in when they wrote it. The mix of Recommended and Not recommended follows the game's own Steam ratio, and the set rotates daily. The opinions are theirs.

A worked example

Take a $19.99 game that reviews at 92 percent from forty thousand people, runs a verified 11 hours, and sells no extras. The reviews are strong, so the ceiling is high. Eleven real hours at twenty dollars is about $1.80 an hour, and the value read takes a real bite out of that ceiling. No monetization haircut. It lands around 70, in worth it, and the page names the price where it would become a BANGER.

Now hold the reviews and shrink the game: same 92 percent, but 3 hours long. The ceiling is identical. The value read now pulls harder, and the score settles lower, into fair. The game did not get worse. At $19.99 for 3 hours you are paying a premium for quality, and the number says exactly that. Put it on sale at $7.99 and it climbs back into worth it. The score follows the price in real time, which is the entire reason it exists.

Flip it once more: a $19.99 game with 60 advertised hours and 68 percent reviews. Dollars per hour says it is the best deal of the three. The score disagrees and puts it in skip, because weak reviews cap it, and Steam's own achievement data says almost nobody meets those 60 hours anyway. Measured on 2026-09-23 across the 2,868 games where this site can verify a finish rate, the median is 11.8 percent. Hours you will never see are not value.

What moves a score day to day

The pipeline crawls fresh data every hour and rebuilds every score every six hours, so a number can move under you. Three things move it:

• Price. The biggest mover. A discount can carry a game across a tier line, and the page shows the exact threshold price it crossed.
• Reviews. A review-bombing, a redemption patch, or a store-policy change shifts the recent score first, then slowly the lifetime one.
• Data fills. When a missing length arrives from a fresh crawl, the length dock lifts and the score corrects, and a new finish rate can add its bonus. Until then the page names every estimate under "What we assumed".

What it does not do

The score rates value. Whether a game is for you is still on you. It is deterministic: the same data gives the same number every time, with no model guessing in the loop. Every input fact behind a score is printed on the page. The weights that combine them are not, because a published formula gets gamed in a week.

Where it is weak, plainly

Global achievement rate is noisy, and a game with no achievements has none of it. HowLongToBeat does not list every game, and a missing length weakens the value read until it is filled in. Owner estimates are rough at the long tail, and they are old. The owner counts, the count of players at once that multiplayer and roguelike games are valued on, and the user tags that sort games into genres all come from SteamSpy, which has not updated its per-game numbers since about mid-2025. No page puts a date on them. The players online number on a game page is Steam's own, printed with the time we checked. We are checking whether Steam's count should replace SteamSpy's in the score, and until it does the score reads SteamSpy's.

The refund signal counts reviewers. It is the share of a game's Steam-purchase reviews whose writer refunded it, read from Steam's own Product refunded label, and someone who refunds inside two hours rarely leaves a review at all, so it understates refunds for every game. It ranks games against each other; read as an absolute refund rate it is too low. Steam keeps a refunded buyer's review and counts it in the review score; this site checked on 2026-08-27 (the 41% of 829 on Gollum reproduces only with its 21 refunded reviews counted). The site also shows Steam's own summary of the reviews written inside the two-hour refund window.

The AI content dock reads Valve's own record of the disclosure every developer fills in before release on Steam, and nothing else. It is not a guess from screenshots, reviews or interviews. It only reaches developers who disclosed: a studio that ships generated art and says nothing is not docked, so the dock lands on the honest ones. Valve records two kinds, content made with AI during development and content the game generates while you play, and both take the same dock.

The score is at its best on games with a real review base and real length data, and it says "limited data" when a game has only a handful of reviews to go on. Treat a provisional number as provisional.

Some unreleased games get a PREDICTED score, which is a forecast. It is worked out from the studio's earlier Steam games and its publisher's games from other studios, two ways and mixed: their reviews run through the same formula at the announced price with no length, and the scores those games actually got here, both pulled toward the genre's recent median. A studio with a short record leans on a publisher with a long one, and a studio with a long record mostly speaks for itself. A developer's AI disclosure docks the forecast the way it will dock the real score. It ships with an 80 percent band and the odds of landing above the typical released game in its own genre. The page prints that range and that probability instead of a tier, because a guessed tier scored worse than always saying skip. A game only gets a forecast when the studio itself has a real record here, and when that kind of record, sorted by how many games and how many reviews they drew, beat quoting the genre's typical score in each of the past few years of real releases, with a range narrower than the genre's own. A publisher's record never earns a forecast on its own, so a studio with no scored game here gets none. Most unreleased games have no such record and get no number at all, and nothing stands in for one. The ranges are still wide, and the about page prints the latest check against real releases, dated. Read it as a bet with stated odds and wait for the real number.