How we measure

afterpatch shows how Steam players met a game update. Not “good or bad,” but how far it is from what's usual for that game.

Where the data comes from

Public Steam reviews only: whether the player recommends the game, the date, the language, playtime and the text. We don't store who wrote a review.
Reviews are collected automatically: every hour for active games, at least once a day for the rest. When a game joins the log, its reviews from up to a year back are pulled in; the chart says which day the history starts.
Reviews that Steam hides by default as off-topic (waves of reviews over things outside the game) are counted too: a reaction to those is still a reaction. And we check separately whether a single language audience is moving the numbers.
Game events come from Steam news and patch notes: patches, seasons, expansions, launches. The kind and scale of each event are labeled by a language model (see below).
We show players online only next to the reaction: how it has changed compared with the game's usual. No sales or prices: this is a log of how updates were met, not a market tracker.

Usual for the game

Every game has its own normal: 95% positive is an ordinary week for one game, while 70% is already a good one for another. So everything is compared with the game's own usual — the median of weekly positive shares over the 90 days before the week in question, never before launch.

A median, not an average: a wave of reviews that lasts a week or two doesn't move the baseline. And each week's share is computed from totals, not averaged by day, so a day with three reviews doesn't weigh as much as a day with three thousand.

Shifts and their strength: M0-M5

The last 7 days are compared with the game's usual. A shift counts only if it is both:

Real: bigger than the game's own swings. Every game has its own noise, and we measure it.
Noticeable: at least 2 pp.

The strength is the lower of two steps: by significance (from 2, 3, 5, 8 and 13 standard errors, adjusted for the game's noise) and by size (from 2, 3, 5, 8 and 12 pp).

M0 · calm: usual for this game
M1 · tremor: a little off the usual; can happen for no reason
M2-M3 · shaking: clearly off the usual
M4-M5 · quake: a strong shift, rare for this game

M means “not as usual,” not “something happened”: a game's mood sometimes drifts on its own.

Reaction to an event

The week after an event against the week before, with a 95% interval. The outcome uses the same words as the game page:

Players loved it · Mood dipped: the shift is at least 2 pp and more than two standard errors, adjusted for the game's noise
Within the game's usual swings: the shift is visible (3 pp or more) but within this game's usual noise: swings like this happen without events too
No notable reaction: the shift is under 3 pp and doesn't clear the bar in the first row
Too few reviews to tell: too few reviews before or after the event to draw a conclusion

Until a week has passed since the event, the outcome is preliminary. A shift soon after a shift the other way that brings the share back to where it was is a rebound, not a reaction to the event.

Launches are a special case: before release, different people write the reviews, so we don't read before and after a launch as a drop or a rise.

What we don't claim

Causes. If reviews moved in the week after a patch, that's exactly what we say: in the week after the patch. Why is up to the reader; player quotes are right there.
Ratings of games or studios. afterpatch doesn't score games or judge developers.
Predictions. The log shows what has already happened, not what will.

How we check ourselves

Every week the same code that draws the site runs on synthetic series where nothing happens (anything found is a false alarm), on series with a step shift (did we notice, and how fast) and on real series with a random fifth of the reviews dropped (does the result hold).

Shifts without an event are shown only above a strict threshold — three standard errors, adjusted for the game's noise: with a softer one, the detector would find shifts in pure noise.

Latest check results. Median across games; the range across games is in parentheses.

False shifts: 0.7 a year on series where nothing happens (0-2.6). For comparison, the games' real series show 1.2 shifts without an event a year.
False alarms: 4% of weeks on series where nothing happens get M1 or higher (0%-12%) — at a 2σ threshold no more than 5% is expected.
Drift: On real weeks with no notable events nearby, M1 or higher shows up in 6% (0%-38%): a game's mood drifts slowly, and a week can differ from the previous three months for no reason at all. M says “not as usual”, not “something happened”.
Reaction interval: On series without changes, the 95% interval of an event reaction misses in 5% of cases (2%-11%); the target is 5%.
Sensitivity: A shift without an event — a 3, 5 or 8 pp shift — is found by the log in 5%, 5% and 20% of cases (±5 days). Less often for quiet games: the threshold is set not to invent shifts.
Stability: With a random fifth of reviews removed, a shift without an event stays in place in 85% of cases (0%-100%), and the week's M level is unchanged in 100%.
Noise floor: The weekly stamp reliably (80% of the time within a week) catches shifts from 8 pp — the median across games; quieter games have a higher floor.
Speed: A shift of 8 pp shows on the weekly stamp in 80% of cases, usually by day 5; a 20 pp shift by day 3.
False alarms: False tremors across the whole log: about 7.1 a week — that many games would get M1 on series where nothing happens. The more games, the more such alarms, which is why M1 reads tremor rather than shaking.
Usual level: The 95% baseline forecast interval covers 97% of days (93%-99%): slightly above target — the baseline errs on the cautious side.

What the check misses: real series have promo waves, review bombing and slow drift, so there may be more false shifts in practice. The shift is tested as a step, while real reactions can be gradual.

Checked on Oct 12, 2026 across 179 games, 20 runs per game.

Where a language model helps

A language model labels news — the kind and scale of an event, which decides what counts as a major one — and tags reviews with topics. It also writes texts under our rules: event headlines and breakdowns, digests, Russian titles for news posts, translations of review excerpts. The numbers, the strength of a shift and the event outcome are computed by code. Code checks the numbers in breakdowns and digests against the data, and review excerpts word for word against the reviews themselves; text that fails the check isn't shown. The model can still get a label or an interpretation wrong — that's an opinion, not a fact.

Quotes and images

Player quotes are excerpts from public Steam reviews, without authors' names or profiles. Game cover art and titles belong to their rights holders; on social cards, the cover is credited to its rights holder.

Data comes from Steam (Valve). afterpatch is not affiliated with or endorsed by Valve. Steam is a trademark of Valve Corporation.

Mistakes

Found a mistake? Write to hello@afterpat.ch. Mistakes in the numbers are fixed first; if a fix changes a conclusion that was already published, we say so.