The storefront
×2.2
times as negative: the reviews Steam shows first, against all of a game’s reviews
Part 01 of 05 · The findings
Eight things updates really do
Ranked by how much they should change the way a studio works. Each one opens into a chapter with the evidence.
01StrongRe-tested
Ch. 0602StrongRe-tested
Ch. 0303StrongRe-tested
Ch. 0104No effect
Ch. 0205StrongRe-tested
Ch. 0506StrongRe-tested
Ch. 0407StrongRe-tested
Ch. 0808Suggestive
Ch. 07
×2.2
how negative Steam’s top reviews are, against the score
Steam’s top reviews are twice as harsh as the game’s score.
The reviews with 10+ helpful votes, the ones buyers see first, are about 2.2 times as negative as the overall rating (80 of 82 games; ×1.96 on new games). They are written largely by experienced players, and their complaints are mostly about how the studio treats its players.
Ch. 0602−40%
reviews after a year without a release
Silence doesn’t anger players. It just loses them.
A year without a release costs 40% of a game’s reviews, yet the share of negative reviews doesn’t budge. The slide starts early: −7% after only one or two quiet months.
Ch. 0303+33%
reviews in a major-release month; a hotfix adds 0
Big releases bring players back. Hotfixes don’t.
A major release or DLC brings a third more reviews in its month and a fifth more the month after. Hotfixes and balance patches bring no lift at all: they simply ship where players already are.
Ch. 01040.0 pts
effect of how an update is packaged
How you package an update makes no difference.
Announcing it weeks ahead, a trailer, “based on your feedback”, a weekend launch: across about 40 features and 135 games, none changed how players reacted. Negative reviews spike for a week after any release, then settle.
Ch. 020539%
chance of losing a fifth of your reviews after a negative spike
Negative reviews warn you three months before players leave.
After the share of negative reviews jumps by 5+ points in a month, there is a 39% chance review volume falls by more than a fifth within three months, against 24% otherwise. It held up on new games (41% against 25%).
Ch. 0506+5.5%
reviews the month after a studio post
One post a month is worth 5.5% more reviews, and most studios skip it.
A month with at least one studio post is followed by 5.5% more reviews, and roadmaps do best at about 6%. Yet two in three months without a release have no post at all. Posts don’t touch the rating.
Ch. 0407121 of 135
games where November reviews are kinder than usual
November is the kindest month on Steam.
In 121 of 135 games November reviews are more positive than the game’s own average (112 of 124 on new games), and it is the busiest month too, with 36% more reviews. September and October are the trough.
Ch. 080868% vs 44%
reviews that still recommend the game
Players forgive grind. They don’t forgive bugs.
A review that mentions grind still recommends the game 68% of the time; one that mentions bugs, 44%; performance, 42%. Technical complaints come mostly from newcomers. This one rests on the first 136 games only.
Ch. 07Myths and reality
Studio folklore, checked against the data.
- Myth
“Going quiet builds anticipation for the next release.”
A release after a long pause brings reviews back to normal, never above it (126 games).
Evidence - Myth
“Updates tank your rating.”
Negative reviews spike for about a week. Over one to three months, content updates and hotfixes stay within normal drift.
Evidence - Myth
“A hotfix in launch week is a sign of failure.”
+0.7 points of negative reviews: indistinguishable from zero (125 games).
Evidence - Myth
“Release in a Steam sale, or in Q4, for a bigger lift.”
No extra lift beyond the season itself (136 and 110 games).
Evidence - Myth
“Discounts bring in outsiders who trash the score.”
More newcomers arrive; the rating stays the same.
Evidence - Myth
“React fast to a review storm and it passes sooner.”
A release or a roadmap right after a negative spike doesn’t speed recovery: about two months either way.
Evidence - Unclear
“Paid content in an update triggers a backlash.”
Zero in our first 136 games, about +3 points on new games, and unstable between sets. Not settled.
Evidence - Partly true
“Big DLC costs you rating.”
On average about +1 point of negative reviews in the first month, but the rise shows up in only 53% of games. A hint, not a rule.
Evidence
What to do with this
Seven working rules for studios, each tied to the number behind it.
- 1
Ship big, not just often.
A major release lifts reviews by a third; a hotfix by nothing. A year with two major releases is 27% busier, with one 13% busier. Chapter
- 2
Judge a launch after 30–90 days, not in week one.
The first-week spike of negative reviews (+1.65 points) fades on its own. Chapter
- 3
Don’t go dark between releases.
Silence costs 7% of reviews within two months. A post a month adds 5.5%; a roadmap about 6%. Chapter
- 4
Set a negative-review alarm.
A jump of 5+ points over the three-month norm raises the risk of losing a fifth of your reviews from 24% to 39%. Chapter
- 5
Talk to your veterans before you change things.
They write the reviews buyers see first, and their complaints are about the relationship with the studio. Chapter
- 6
Make the first hours stable.
Bugs and performance are the least forgiven topics, and newcomers raise them most (suggestive). Chapter
- 7
Plan around the calendar.
November is the kindest and busiest month; September and October are the trough. Chapter
For journalists: quote-ready figures
Each line is exact and carries the context it needs. The chapters hold the evidence, and we are glad to share more.
- On Steam, reviews with 10 or more helpful votes are about 2.2 times as negative as a game’s overall rating (median of 82 games; ×1.96 across 118 games tested separately).
- A year without a release goes with 40% fewer reviews, while the share of negative reviews stays the same (124 games; ×0.53 on 63 new games).
- A major release or DLC brings about a third more reviews in its month; hotfixes and balance patches bring no lift (266 games).
- After the share of negative reviews jumps by 5 points or more in a month, the chance that reviews fall by a fifth within three months rises from 24% to 39% (115 games).
- In 121 of 135 games, November reviews are kinder than the game’s own yearly average.
These are associations within games, measured on Steam’s English-language reviews. They say nothing about sales or revenue. Press page · [email protected]
Part 02 of 05
After an update
What a release does to a game’s reviews and rating, and how fast a quiet spell costs you your audience.
Chapter 01
What brings the audience back
The size of an update brings an audience back, not the fact that something shipped. A major release or DLC lifts the number of reviews a game gets (our stand-in for how active its audience is) by about a third in the month it ships. Hotfixes and balance patches do not.
Strong evidenceHeld up on new games
Figure 1
A major release lifts reviews by a third, and the lift is gone within a quarter
Reviews written in the month, as a multiple of an ordinary month of the same game. The capsule on each column shows the range we are 95% sure the true value sits in.
- release month
- the months after
- hotfix or balance patch
- ordinary month = ×1.00
×0.9×1.1×1.2×1.3×1.4×1.5×1.00×1.33+33%release month×1.21+21%+1 month×1.10+10%+2 months×1.06+6%+3–5 months×1.000%Hotfix or balance patchMAJOR RELEASE OR DLC · AND THE MONTHS AFTERHOTFIX OR BALANCE PATCHthe lift is gone within about a quarter
×0.9×1.1×1.2×1.3×1.4×1.5×1.00×1.33+33%release×1.21+21%+1 mo×1.10+10%+2 mo×1.06+6%+3–5 mo×1.000%hotfix
Counted from Steam’s own counters; 266 games, the weeks around major releases.CritField · Issue 02
RELEASE MONTH
+33%
A major release or DLC brings a third more reviews that month
Compared with an ordinary month of the same game. A game that usually gets 100 reviews a month gets about 133.
- 266 games
- ✓ ×1.34 on 102 new games
SIZE OF THE UPDATE
+27%
The bigger the update looks in its patch notes, the bigger the lift
Updates the notes call “large” add this on top of what their release type brings. A new game mechanic is a weaker signal: about +6%, and not consistent from game to game.
- 119 games
- ✓ ×1.29 on 82 new games
THE YEAR
+13% / +27%
Two major releases in a year beat one
A year with one major release is 13% busier than a year with none. With two, it is 27% busier. Four to twelve content updates a year did not lift the yearly total.
- 192 games
- ✓ held up on new games
HOTFIXES AND BALANCE
×1.00
A hotfix does not bring players back
Compared with ordinary dates, there is no lift. Months with a hotfix only look busier because hotfixes ship where lots of players already are.
- 62 hotfix games
- 33 balance games
A game that usually gets 100 reviews a month gets about 133 in the month of a major release, about 121 the next month and about 110 the month after. Across our games, major releases brought roughly 553 thousand extra reviews, about 370 per release month on average.
How much data. 2.22 million reviews were written in the 1,486 months (283 games) that had a major release, counted from Steam’s own counters. Another 2.42 million fall in the weeks around 584 stand-alone major releases. The whole data set covers 21.31 million reviews and 27,740 game-months. On the new games, the same conclusion rests on 620 thousand reviews from release months and 552 thousand from the weeks around them.
What to do
Treat major releases and DLC as your main tool for bringing the audience back. Don’t expect hotfixes and balance patches to do the same job.
- Does the lift last? The first month held up on new games . It brought 1.72 times as many reviews as ordinary dates in the same game (51 games; 38 of 45 beat their ordinary dates). The third month just missed our statistical bar on the first 130 new games (×1.14). With 80 more games added (210 in all), both months pass: ×1.50 in month one and ×1.16 in month three. But in month three only 57% of games (50 of 87) beat their ordinary dates, below our 60% bar. So “the lift lasts a quarter” is true on average, not in every game.
- “Large” updates. Updates that the patch notes describe as large add a further 29% on top of what their release type brings (82 games, 1,734 releases), and the same holds with the added games (×1.30, 143 games, the same direction in 74 of 90 games).
- A new game mechanic. Our first search suggested +17–22% (124 games). On new games it shrinks to about +6%. With all 210 new games the number clears the statistical bar (164 games), but only 56% of games (70 of 125) point the same way, below our 60% bar. Treat it as a small average bump, not a rule.
- Content updates. In premium games they lift the first month by 15% (65 games, 51 above their ordinary dates). Free-to-play games show no lift (see the free-to-play chapter).
- Hotfixes and balance patches. Months with a hotfix look busier (×1.10), but compared with ordinary dates there is no lift: hotfixes ship where lots of players already are.
- A whole year. A year with one major release is 13% busier than a year with none (×1.13, range 1.07–1.19, 192 games; ×1.16 in the first group of games, ×1.08 in the new ones). With two major releases it is 27% busier. Four to twelve content updates a year did not lift the yearly total (×1.05 for four updates, range 0.95–1.16, 235 games), though we cannot rule out a small effect.
- A long pause does not build anticipation. A release after a pause of 90+ days brings reviews back to the game’s norm, not above it (126 games).
- 130 gamesMajor releases and DLC lift reviews compared with ordinary dates in the same game (months one and three)Reviews behind it552 thousand countedNot confirmed
Month one: ×1.72 (above ordinary dates in 38 of 45 games). Month three: ×1.14, just short of our statistical bar (21 of 44 games). 51 games.
- 210 gamesMajor releases and DLC lift reviews compared with ordinary dates in the same game (months one and three)Reviews behind it10.57 million countedConfirmed
Month one: ×1.50 (66 of 88 games above ordinary dates). Month three: ×1.16, but only 50 of 87 games (57%) beat their ordinary dates, below our 60% bar. 97 games.
- 130 gamesA major-release month lifts reviews, and the lift fadesReviews behind it620 thousand countedConfirmed
×1.34 in the release month and ×1.22 the month after. 102 games.
- 130 gamesA “large” update beats its release type in the first monthReviews behind it2.06 million countedConfirmed
×1.29. 82 games, 1,734 releases.
- 210 gamesA “large” update beats its release type in the first monthReviews behind it10.57 million countedConfirmed
×1.30, and the same direction in 74 of 90 games. 143 games.
- 130 gamesA new game mechanic lifts first-month reviews beyond its release typeReviews behind it2.06 million countedNot confirmed
×1.06, just short of our statistical bar. 95 games. Our first search had suggested +17–22%.
- 210 gamesA new game mechanic lifts first-month reviews beyond its release typeReviews behind it10.57 million countedOnly on average
×1.06 clears the bar (164 games), but only 70 of 125 games (56%) point the same way.
- 130 gamesA year with a major release is busierReviews behind it5.51 million countedConfirmed
×1.08 (×1.16 in our first search). 89 games.
- 130 gamesThe number of content updates in a year makes no difference (within ±10% for four updates)Reviews behind it5.51 million countedNot confirmed
×1.03. The range (0.90–1.19) is wider than our ±10% margin. 112 games.
- 210 gamesThe number of content updates in a year makes no difference (within ±10% for four updates)Reviews behind it10.02 million countedNot confirmed
×1.02, range 0.90–1.16, wider than our margin. 190 games.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- Review volume is a stand-in for audience activity. It is not a head-count of players.
- These are associations within games: a studio may ship a big update when things are already going well.
- The third-month lift is an average. It holds in 57% of games, below our 60% bar.
Chapter 02
Do updates damage the rating?
In a typical game, no. Apart from a short first week, neither content updates nor hotfixes move the share of negative reviews by more than normal drift over the following one to three months. Major releases carry a small caveat: about 1.1 percentage points more negative reviews in the first month, though that is an average pulled up by some games, not a broad effect.
Strong evidenceSome checks not confirmed
Figure 2
Content updates and hotfixes stay within normal drift; balance patches and major releases may not
Change in the share of negative reviews in the first month after a release, compared with dates in the same game that had no release.
- ±1.5 points: normal drift
- range stays inside
- range reaches outside
normal drift: ±1.5 points
- Content update198 games+0.1−0.3+0.4within normal drift
- Hotfix169 games+0.3−0.8+1.4within normal drift
- Balance change76 games+0.7−0.1+1.5may exceed normal drift
- Major release / DLC109 games+1.1+0.1+2.1may exceed normal drift
-3-2-10+1+2+3
change in the share of negative reviews in the first month, in percentage points; bar = the range we are 90% sure the true value sits in
Counted from Steam’s own counters; the weeks around single releases, compared with ordinary dates in the same game.CritField · Issue 02
FIRST WEEK
+1.65 pts
Every release brings a short spike of negative reviews
It passes on its own: the rebound almost equals the dip, and hotfixes, apologies or a roadmap do not speed it up.
- 127 games, higher in 79
- 95 games for the rebound
NORMAL DRIFT
±1.5 pts
Content updates and hotfixes do not move the rating
Over one to three months the share of negative reviews stays within normal drift, in all three monthly checks.
−3−20+2+3
- 198 and 169 games
MAJOR RELEASES
+1.1 pts
A major release may cost about a point of rating in its first month, but not in every game
Negative reviews rise in only 53% of games, so this is a hint, not a rule.
rises in 53% of games
- 109 games
- 90% range +0.2 to +2.1
- suggestive
PACKAGING
0.0 pts
How you package an update does not change the reaction
None of these moved the share of negative reviews in both halves of the games. Paid content is the exception we can’t call: see “Gaps”.
−20+2
- 135 games
- about 40 features
How much data. 18.29 million reviews, counted from Steam’s own counters, in the weeks around 5,284 stand-alone releases. On the new games, the verdict on the rating rests on 4.06 million reviews around releases and 259 thousand around releases with patch-note details. We read and coded 651,642 review texts, 103,387 of them negative.
What to do
Judge a release over 30–90 days, not by its first week. Don’t spend effort “packaging” an update to manage the reaction: the announcement, the tone, a video or the day of the week do not change the rating.
- The first week. Negative reviews run 1.65 points above normal in the first week, and the rebound almost equals the dip (95 games). Hotfixes, apologies and roadmaps do not speed up the recovery.
- How an update is packaged. We looked at 7,840 releases and about 40 features of their patch notes against 10 measures of reaction. No effect repeated in both halves of the games. Examples, within a game: an announcement 1–45 days ahead −0.2 points (117 games), removed features −0.9 (73), an event −0.1 (112), “based on your feedback” −0.2 (110), a video +0.2 (110), a weekend release +0.3 (100). None can be told apart from zero. Paid content showed 0.0 in this first search (76 games), but the larger data muddied that: see “Gaps”.
- Checked on new games. Our strict test for “no effect” asked the whole range of doubt to sit within ±1.5 points. On the first 130 new games, not every case passed. With 80 more added (210 games), 6 of 12 cases passed for the release types with at least 50 games (major release or DLC: 86 games; content update: 160; balance patch: 59; hotfix: 136). For the packaging features, none could be confirmed as “no effect” within ±1 point: the ranges are too wide. Paid content now has 95 games (47 before), and its estimate is positive but weakens from one set of games to the next: +3.1 points (range 0.0 to +6.1).
- A caveat, not a finding. Patch notes that call an update “large” go with slightly more negativity (+1.4 points, 119 games), and so did releases that only weaken things, the “nerf-only” releases (+3.4 points in the first month, 60 games). The nerf-only result did not repeat on the larger set (−2.0 points on 84 games, with only 5 of 11 games moving up), so it stays a hypothesis.
- 130 gamesNo release type moves the share of negative reviews (within ±1.5 points, months one to three)Reviews behind it4.06 million countedNot confirmed
3 of 12 ranges fell outside the margin. Major releases or DLC (44 games) and balance patches (36) had under 50 games.
- 210 gamesNo release type moves the share of negative reviews (within ±1.5 points)Reviews behind it10.57 million countedNot confirmed
6 of 12 ranges sit inside the margin (release types with 50+ games): major release or DLC 86 games, content update 160, balance patch 59, hotfix 136.
- 130 gamesHow an update is packaged does not change the shift in negative reviews (announcement, paid content, removals, event, “your feedback”, video, weekend; within ±1 point)Reviews behind it259 thousand countedNot confirmed
Every range was wider than ±1 point (about ±2). Paid content had only 47 games. 119 games.
- 210 gamesHow an update is packaged does not change the shift in negative reviews (announcement, paid content, removals, event, “your feedback”, video, weekend; within ±1 point)Reviews behind it388 thousand countedNot confirmed
None of 7 features fits inside ±1 point. Paid content: +3.1 points (range 0.0 to +6.1), 95 games.
- 210 gamesA release that only weakens things (a “nerf-only” release) raises negative reviews the next monthReviews behind it388 thousand countedNot confirmed
−1.99 points on 84 games. Only 5 of 11 games moved up.
- 210 gamesA surprise major release brings more negativity than an announced oneReviews behind it388 thousand countedNot confirmed
−1.22 points, the other way. 56 games with both kinds.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- Releases and patch-note details were sorted from the developers’ news and checked blind on a sample.
- To say “no effect”, we set a margin in advance and required the whole range of doubt to fit inside it. Where the range was wider, we say “not confirmed”.
- The 1.1-point rise after major releases is an average pulled up by a few games: negative reviews rise in only 53% of games.
Chapter 03
What a quiet spell costs
A pause costs reviews, not rating. The longer a game goes without a release, the fewer reviews it gets, while the share of negative reviews does not rise. A regular rhythm adds nothing on its own beyond the size of the releases.
Strong evidenceSome checks not confirmed
Figure 3
Silence costs audience gradually: −7% after a month or two, −40% after a year
Reviews per month by time since the last release, as a multiple of the month right after a release. The shaded band is the range we are 95% sure the true value sits in.
- month right after a release = ×1.00
- reviews lost
- 95% range
×0.5×0.6×0.7×0.8×0.9×1.00the month right after a release×0.93−7%1–2 months×0.90−10%2–3 months×0.87−13%3–6 months×0.78−22%6–12 months×0.60−40%12+ monthsreviews lost compared with the month right after a release−7% already after one or two months−40% after a year
×0.5×0.6×0.7×0.8×0.9×1.00×0.93−7%1–2 mo×0.90−10%2–3 mo×0.87−13%3–6 mo×0.78−22%6–12 mo×0.60−40%12+ mo
Counted from Steam’s own counters; 124 to 252 games per point.CritField · Issue 02
A MONTH OR TWO
−7%
Silence starts costing within two months
After a pause of one to two months, a game gets ×0.93 of the reviews it got right after a release, and the curve keeps falling.
- 252 games
A YEAR
−40%
After a year without a release, reviews run 40% lower
×0.60 compared with the months right after a release. A pause does not store up demand: the next release brings reviews back to normal, not above it.
- 124 games
- ✓ ×0.53 on 63 new games
THE RATING
−0.07 pts
A quiet spell does not hurt the rating
The share of negative reviews is the same after a pause of 45 days or more.
−1.50+1.5
- 127 games
- range inside ±1.5 pts
RHYTHM
×1.05
An even rhythm is not the secret; release size is
An irregular rhythm looked costly (×1.12) until we accounted for release types. Then it shrank to ×1.05.
×0.9×1.0×1.1
- 227 games
- a ±10% effect can’t be ruled out
How much data. 1.39 million reviews, counted from Steam’s own counters, in months that came after a year or more without a release (3,930 months, 169 games; on the new games, 742 thousand reviews). By our estimate (×0.60 against months with a fresh release), those months lost about 915 thousand reviews. To judge negativity in quiet spells we read 40,762 reviews from 127 games (354 samples of 200 reviews).
What to do
Don’t hold back a long pause to “build anticipation”: a pause does not store up demand, and review volume slides gradually (−7% already after one or two months). Choose the rhythm your production can sustain.
- Pause and rating. On the new games the estimate is −0.1 points, but its range (−1.8 to +1.5) is slightly wider than the ±1.5 margin we set. With all games together it is −0.07 points (range −1.11 to +0.97), which fits inside it.
- Rhythm. In our first specification, an irregular release rhythm seemed to cost ×1.12 (range 1.03–1.21). The effect vanishes once release types are taken into account (×1.05): it was release size, not rhythm. We could not fully rule out an effect within ±10%.
- Relative pauses. How long a pause is compared with the game’s usual gap between releases predicts reviews no better than the plain number of days since the last release.
- The same pattern in five slices. A pause of 12+ months costs reviews in every slice we cut: ×0.58 for premium games, ×0.64 for free-to-play (15 games), ×0.65 for 2019–2022, ×0.79 for 2023–2026, ×0.80 for large games (49 games). For small games the model could not produce a stable estimate, so we don’t use it.
- 130 gamesReviews fall the longer a game goes without a releaseReviews behind it742 thousand countedConfirmed
After 12+ months without a release: ×0.53. 63 games.
- 130 gamesA regular release rhythm makes no difference to review volume (within ±10%)Reviews behind it5.51 million countedNot confirmed
×1.11. The range (0.98–1.27) reaches outside the margin. 100 games.
- 210 gamesA regular release rhythm makes no difference to review volume (within ±10%)Reviews behind it10.02 million countedNot confirmed
×1.11, range 1.00–1.24. 178 games.
- 130 gamesA pause does not change the share of negative reviews (within ±1.5 points)Reviews behind it5.51 million countedNot confirmed
−0.12 points. The range (−1.76 to +1.52) is a hair wider than the margin. 63 games.
- 210 gamesA pause does not change the share of negative reviews (within ±1.5 points)Reviews behind it10.02 million countedNot confirmed
−0.60 points, range −1.89 to +0.68. 108 games.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- A pause of a year is a different situation from a pause of a month, and few games have both.
- The rhythm effect vanishes once release size is accounted for; we could not rule out an effect within ±10%.
- For small games the 12-month estimate was unstable, so we don’t use it.
Part 03 of 05
Between updates
What a studio can do in the quiet months, and which warning sign shows that an audience is leaving.
Chapter 04
Between releases: what studio posts do
A month with at least one studio post (not a release) lifts the next month’s reviews by about 5.5%, and a roadmap post by a little more. That comes on top of the game’s current activity and holds in months without releases. Posts do not affect the rating.
Strong evidenceHeld up on new games
Figure 4
Roadmaps and announcements work best; sales only inside their own month
Extra reviews the next month for each studio post, by kind of post. Bars: all months. Diamond: months with no release.
- strong evidence
- suggestive
- weak or none
- months with no release
- Any post130 games+5.5%strong
- Roadmap, announcement126 games+6.4%strong
- Community, marketing123 games+2.9%suggestive
- Stream, trailer90 games+6.0%suggestive
- Developer diary53 games+5.6%suggestive
- Event93 games+2.5%weak or none
- Sale85 games+0.9%weak or none
0+2%+4%+6%+8%+10%
Counted from Steam’s own counters; per studio post, on top of the game’s current activity.CritField · Issue 02
ANY POST
+5.5%
A studio post lifts the next month’s reviews by 5.5%
On top of the game’s current activity, and it holds in months without releases (+5.9%).
- 256 games
- ✓ +5.5% on 126 new games
WHICH POST
+6.4%
Roadmaps and announcements work best; a sale post works only in its own month
A roadmap or announcement adds 6.4% the next month. A sale post adds almost nothing the next month (+0.9%), but +14% in the month of the sale.
- 126 games
- ✓ +6.0% on 124 new games
THE RATING
0 pts
Posts don’t touch the rating
In quiet spells, a month with a post has the same share of negative reviews as one without.
−10+1
- 164 games
- ✓ inside ±1 pt
THE UNUSED LEVER
34%
Most studios stay silent between releases
At least one post appears in only 34% of months without releases, although a post is worth about +6% the next month.
34% of quiet months have a post
- 316 games
- 50,106 posts
How much data. 11.54 million reviews, counted from Steam’s own counters, were written in the months after a month with a studio post (10,385 months, 336 games, 51,650 posts). On the new games the verdict rests on 2.40 million reviews.
What to do
In the gaps between releases, post regularly, especially roadmaps and announcements. Most studios don’t: at least one post appears in only 34% of months without a release.
| Kind of post | Reviews the next month, per post | In months with no release | Games | How strong |
|---|---|---|---|---|
| Any post (not a release) | +5.5% | +5.9% | 130 | strong |
| Roadmap or announcement | +6.4% | +7.5% | 126 | strong |
| Community or marketing | +2.9% | +2.5% | 123 | suggestive |
| Stream, trailer or showcase | +6.0% | +8.3% | 90 | suggestive |
| Developer diary | +5.6% | +7.8% | 53 | suggestive |
| Event | +2.5% | +2.9% | 93 | weak or none |
| Sale | +0.9% | −1.0% | 85 | weak or none (only in its own month: +14%) |
- Cause and effect. We cannot prove it. Studios do write a little more often when the game is busier, but the effect of a post on the next month is clearly stronger than the reverse, and it holds after adjusting for current activity.
- Posts and the rating in quiet spells. In quiet spells (no release this month or last), a month with at least one studio post has the same share of negative reviews as a month without one: −0.16 points (range −0.66 to +0.34) in the same month and +0.02 (range −0.50 to +0.53) in the next, across 164 games and 6,039 game-months. Both ranges sit inside ±1 point, and the result held up on the new games. One caveat: we first ran this check as a test on the original 130 new games, so that part of the data was not blind. The 80 added games were.
- 130 gamesA studio post (not a release) lifts reviews the following monthReviews behind it2.40 million countedConfirmed
+5.5% per post. 126 games.
- 130 gamesA roadmap or announcement lifts reviews the following monthReviews behind it2.40 million countedConfirmed
+6.0% per post. 124 games.
- 210 gamesIn quiet spells, a month with a studio post has the same share of negative reviews as a month without (within ±1 point)Reviews behind it1.31 million countedConfirmed
−0.16 points (range −0.66 to +0.34). 164 games.
- 210 gamesAnnouncing a delay changes neither negativity nor review volumeReviews behind it143 thousand countedNot confirmed
Same month +0.2 points (−1.2 to +1.5); next month −0.7 (−1.7 to +0.3); reviews ×1.10. 82 games.
- 210 gamesA post about changing monetization changes neither negativity nor review volumeReviews behind it145 thousand countedNot confirmed
Same month −0.4 points (−2.4 to +1.5); next month −0.6 (−2.6 to +1.5); reviews ×1.02. 64 games.
- 210 gamesAn apology or explanation post changes neither negativity nor review volumeReviews behind it558 thousand countedNot confirmed
Same month +1.0 points (+0.1 to +1.9); next month +0.9 (+0.1 to +1.6); reviews ×1.03. 107 games.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- Cause and effect are not proven: studios write a little more when the game is busier.
- The kind of post (roadmap, community and so on) comes from post titles and labels, so it is noisy.
- Sale posts matter only in their own month.
Chapter 05
The early warning sign
A rising share of negative reviews runs about a quarter ahead of falling review volume. Each +10 percentage points of negative reviews in a month goes with about 10% fewer reviews three months later, even at the same current volume. After a jump of 5+ points, a big drop in reviews is about 1.5 to 1.7 times as likely as usual.
Strong evidenceHeld up on new games
Figure 5
After a jump of 5+ points in negative reviews, a drop in review volume is about 1.5 times as likely
Chance that reviews fall by more than 20% within three months, by the change in the share of negative reviews. The ×-figures show how many times the normal chance that is.
- average over all months
- large games
10%20%30%40%50%24%21%×0.88within ±2 points30%×1.26rise of 2–5 points37%×1.55rise of 5+ points48% in large games
10%20%30%40%50%24%21%×0.88within ±2 points30%×1.26rise of 2–5 points37%×1.55rise of 5+ points
Counted from Steam’s own counters; 115 games with a spike. The 48% for large games comes from the details tab.CritField · Issue 02
THE EARLY SIGNAL
−10%
Each 10-point jump in negative reviews now means 10% fewer reviews three months later
Even at the same current volume: negative reviews run about a quarter ahead of the audience.
- 266 games
- ✓ −9% on 113 new games
AFTER A SPIKE
39% vs 24%
After a jump of 5+ points, a collapse in reviews is much likelier
Chance that reviews fall by more than 20% within three months, after a negative spike and without one.
- 115 games
- ✓ 41% vs 25% on 50 games
HOW COMMON
1 in 13–15
Negative spikes are common: 215 of 256 games had at least one
6.7% of months in the first games and 7.8% in the new games are negative spikes, so a warning rule has something to catch in most games.
215 of 256 games
- first games: 117 of 136
- ✓ new games: 98 of 120
RECOVERY
2 months
Most games recover from a spike in about two months; one in five has not after half a year
A release or a roadmap in the next month does not speed recovery.
19% do not recover in six months
- median 2 months
- ✓ 1 month and 16% on new games
How much data. 1,388 months with a negative spike (1.18 million reviews, counted from Steam’s own counters) in 284 games. We estimated the signal on the whole data set (21.31 million reviews). On the new games it rests on 5.51 million reviews.
What to do
Set up an early-warning rule: a rise of 5+ points in the monthly share of negative reviews over the three-month norm means a 1.5 to 1.9 times higher risk of a drop in reviews next quarter.
| Change in the share of negative reviews, against the norm | Chance of reviews falling by 20%+ within 3 months | Times the normal chance |
|---|---|---|
| within ±2 points | 21% | ×0.88 |
| a rise of 2–5 points | 30% | ×1.26 |
| a rise of 5+ points | 37% (48% for large games) | ×1.55 (×1.9) |
- A feedback loop. A busier month also predicts a lower share of negative reviews three months later.
- Which complaint is the dangerous one? The data don’t say. Quarterly topic shares add no predictive power beyond the overall share of negative reviews.
- Recovery. The median spike takes 2 months to pass (1 month on new games), and 19% (16% on new games) have not recovered after half a year. A release or a roadmap in the next month does not speed recovery.
- 130 gamesA higher share of negative reviews in one month predicts fewer reviews three months laterReviews behind it5.51 million countedConfirmed
−0.95% for each percentage point; the same direction in 70 of 113 games.
- 130 gamesA jump of 5+ points in negative reviews raises the risk of a 20%+ drop in reviewsReviews behind it200 thousand countedConfirmed
41% against 25% (1.69 times as likely). 50 games.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- The warning rule ranks risk. It does not tell you why players leave.
- Quarterly topic shares are noisy (about 40 reviews per game-quarter).
- Large games carry more of the risk after a jump of 5+ points (48% against 37%).
Part 04 of 05
Players and context
Who writes the reviews buyers see first, what players forgive, which time of year is kindest, and how free-to-play games differ.
Chapter 06
Who writes, and what the buyer sees
Buyers see complaints before they see the score. The “helpful” reviews that Steam shows first are more than twice as negative as the overall rating, and they are written by experienced players, mostly about how the studio treats its players.
Strong evidenceHeld up on new games
Figure 6
The reviews Steam shows first are far more negative than the rating
Share of negative reviews among all reviews, and among reviews with 10 or more helpful votes (the ones the storefront ranks first).
- all reviews
- 10+ helpful votes
- First 136 gameswhere we searched15%all reviews45%10+ helpful votes+30 pointson the storefront
- New games130 games, our check17%all reviews44%10+ helpful votes+27 pointson the storefront
0%10%20%30%40%50%60%
Reviews read: 10,328 helpful reviews in the first 136 games and 5,974 in the new games.CritField · Issue 02
THE STOREFRONT
×2.2
The reviews Steam shows first are twice as negative as the rating
In the median game. The most prolific 10% of reviewers write 22% of the helpful reviews.
- 80 of 82 games
- ✓ ×1.96 in 102 of 118
WHAT GETS UP THERE
×2.6 / ×2
A negative review reaches the storefront more often than a positive one of the same length
At equal length, author experience and review age: 2.6 times as often in the first games, about twice as often (+4.7 points on a 4.1% base) in the new games.
- 124 of 136 games
- ✓ 109 of 130 games
WHO WRITES
16% → 39%
Newcomers write about bugs; veterans write about the studio
Among negative reviews, the share about distrust and the relationship with the developer rises from newcomers (under 2 hours played) to veterans (200+ hours).
- 64 of 72 games
- ✓ 100 of 128 games
EXPERIENCED AUTHORS
+5 pts
Reviewers with a long record are harsher
An author with 20 reviews is 5 points more negative than an author with one, all else equal.
- 134 of 136 games
- ✓ 126 of 130 games
How much data. We read reviews with 10+ “helpful” votes (20,123, of which 9,157 are negative), reviews by veterans with 200+ hours played (82,428), reviews by newcomers with under 2 hours (39,011) and reviews by authors with 20+ reviews (150,996). On the new games the findings rest on 146 thousand reviews: 25.4 thousand of them negative and 5,974 with 10+ votes.
What to do
Explain changes in advance when they affect heavily invested players. Their complaints are about their relationship with the developer, which is exactly the negativity Steam shows buyers first.
- What lifts a review to the storefront. Length gets a review to the top in all 136 games, and the author’s experience does in 125 of 136. The most prolific 10% of reviewers write 22% of the “helpful” reviews.
- Veterans and newcomers. Newcomers write about bugs and the barrier to entry. Veterans write about their relationship with the developer.
- Updates and the mix of reviewers. After hotfixes the share of newcomers fell by 1.5 points in our first search (87 of 122 games), but that did not repeat on new games (23 of 49). With all 210 new games the average fall is still there (−1.5 points, in 38 of 64 games, or 59%), but it misses our 60% bar, so we draw no conclusion. Major releases do not change the mix. After sales, newcomers rise by 3.8 points (39 of 49 games) while the rating stays the same.
- Edits. In the week after a release there are about 1.5 times as many edits to old reviews, but the final vote after an edit is no kinder.
- 130 gamesAmong negative reviews, complaints about the studio grow with hours playedReviews behind it25.4 thousand readConfirmed
+3.6 points for each step up in playtime; higher for veterans in 100 of 128 games.
- 130 games“Helpful” reviews (10+ votes) are more negative than all reviewsReviews behind it5,974 readConfirmed
In 102 of 118 games; typically ×1.96.
- 130 gamesProlific reviewers are harsherReviews behind it146 thousand readConfirmed
+1.7 points for each step up in reviews written; 126 of 130 games.
- 130 gamesA negative review is more likely to get 10+ “helpful” votesReviews behind it146 thousand readConfirmed
+4.7 points on a 4.1% base; 109 of 130 games.
- 210 gamesThe share of newcomers falls after hotfixesReviews behind it388 thousand countedNot confirmed
−1.50 points. 64 games, with a fall in 38 of 64.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- “Helpful” votes are a single snapshot from the day we downloaded the data.
- The link between hotfixes and fewer newcomers did not repeat on new games.
Chapter 07
What players forgive
Players forgive grind but not bugs or performance problems. A review that mentions grind still recommends the game 68% of the time. For bugs the figure is 44%, for performance 42%.
Suggestive
Figure 7
Players forgive grind far more readily than bugs or performance
Share of reviews that still recommend the game, among the reviews that mention the topic. The thin line marks half.
- complaint topics
- praise topics
- half of mentions
- Complaints
- Grind68%
- Bugs44%
- Performance42%
- Praise
- Story and world94%
- Gameplay93%
- Visuals and audio90%
0%25%50%75%100%
Reviews read: 38,208 coded reviews in the first 136 games. Suggestive.CritField · Issue 02
FORGIVEN, NOT FORGIVEN
68% vs 44%
Players forgive grind; they don’t forgive bugs or performance
Share of reviews that still recommend the game, among those that mention the topic.
- 38,208 reviews
- 136 games
WHO COMPLAINS
−0.6 / −1.0 pts
Technical complaints come mostly from newcomers
Mentions of bugs and performance fall steadily as hours played grow (by 0.6 and 1.0 points per step up on a logarithmic scale of playtime). Stability in the first hours matters more than late content.
- 107 and 79 games
- strong evidence
THE COSTLIEST COMPLAINT
+28.5 pts
A complaint about updates is the most damaging of all
Reviews that complain about updates have a negative share 28.5 points higher than reviews that don’t (bugs +18.3, matchmaking +16.8, performance +15.7).
same direction in 97% of games
- 62 games
- strong evidence
HOW CONSISTENT
93%
In almost every game, grind is forgiven more readily than bugs
A comparison inside each game that mentions both.
93% of 74 games
- 74 games
- suggestive
How much data. We coded 38,208 reviews from the first 136 games: bugs come up in 3,548, performance in 2,440, grind in 2,539, and gameplay (in praise) in 22,499.
What to do
Technical complaints come mostly from newcomers, so stability in the first hours matters more than late content.
- Read with care. A review that says “bugs” is negative almost by definition. This is the make-up of negative reviews, not the cause of the rating.
- After releases. No topic shifted after releases in a way that passed our two-halves check and the 50-game rule.
- More coded reviews. We coded 5,390 more reviews from 35 more games (43,453 in all). “Complaints about updates” now reaches 62 games and passes every check: reviews that complain about updates have a negative share 28.5 points higher than reviews that don’t, with the same direction in 97% of games and in both halves of the data. Matchmaking stays at 39 games, under 50. Using review topics as an early warning still can’t be tested (23–29 games per topic). We did not recompute topic shares after adding reviews, because the new ones come from 35 games and are not a random sample, so the shares in the first table stay as they were.
| Topic | Share of reviews | Still recommend if mentioned | Raises the negative share by | Games | Result |
|---|---|---|---|---|---|
| Bugs | 9.3% | 44% | +17.9 points | 107 | passes |
| Performance | 6.4% | 42% | +16.4 points | 79 | passes |
| Grind | 6.6% | 68% | +3.0 points | 75 | did not repeat in both halves |
| Gameplay (positive) | 58.9% | 93% | −32.9 points | 136 | passes |
| Story and world | 9.7% | 94% | −11.1 points | 54 | passes |
| Visuals and audio | 10.7% | 90% | −11.3 points | 107 | passes |
| Complaints about updates | 4.2% | 32% | +28.7 points | 41 | too few games (under 50) |
| Matchmaking | 3.3% | 45% | +16.2 points | 33 | too few games (under 50) |
After adding the coded reviews:
| Topic | Games before | Games after | Shift in negative share (points) | Same direction in | Result |
|---|---|---|---|---|---|
| Bugs | 107 | 109 | +18.3 | 99% of games | passes |
| Performance | 79 | 86 | +15.7 | 100% of games | passes |
| Grind | 75 | 80 | +2.7 | 65% of games | did not repeat in both halves |
| Complaints about updates | 41 | 62 | +28.5 | 97% of games | passes |
| Matchmaking | 33 | 39 | +16.8 | 97% of games | too few games (under 50) |
| Gameplay (positive) | 136 | 136 | −33.6 | 100% of games | passes |
| Story and world | 54 | 57 | −11.3 | 96% of games | passes |
| Visuals and audio | 107 | 113 | −10.8 | 94% of games | passes |
We did not write down a test for this chapter in advance. These findings come from the first 136 games only, so treat them as suggestive.
- A review that says “bugs” is negative by construction: read it as the make-up of negative reviews.
- Only the first 136 games, so there is no check on new games yet.
- Only 46% of the review queue was coded, in an order unrelated to what the reviews say.
Chapter 08
The Steam calendar
November is the kindest month for reviews, and September and October are the trough in review volume. Releasing in a sale month adds nothing beyond the season.
Strong evidenceHeld up on new games
Figure 8
November is the kindest and busiest month; September and October are the trough
By calendar month, across all games: review volume (top) and the shift in the share of negative reviews (bottom).
- average month = 1.00
- more negative than the game's own average
- kinder
0.81.21.41.000.821.36REVIEW VOLUME · reviews per month, 1.00 = the game's average monthNovember: 36% busier than an average monthSeptember: the trough
−4−2+20JJanFFebMMarAAprMMayJJunJJul+1.4AAugSSepOOct−3.2NNovDDecNEGATIVE REVIEWS · against the game's own yearly average, in percentage pointsNovember: 3 points kinder than the game's average
0.81.21.41.000.821.36REVIEW VOLUME · 1.00 = average
−4−2+20JJanFFebMMarAAprMMayJJunJJulAAugSSepOOctNNovDDecNEGATIVE REVIEWS · points vs average
Counted from Steam’s own counters; 136 games for review volume, 266 games for the share of negative reviews.CritField · Issue 02
THE BUSIEST MONTH
×1.36
November is the busiest month of the year
November has 36% more reviews than an average month of the same game, December 21% more. September is the lowest at ×0.82.
- 136 games
THE KINDEST MONTH
121 of 135
Almost every game gets kinder reviews in November
November has a lower share of negative reviews than the game’s own average (−3.2 points across all games).
121 of 135 games
- 135 games
- ✓ 112 of 124 new games
THE QUIET MONTHS
−30 … −35%
September and October are the trough
Review volume in September and October runs 30–35% below January.
- 136 games
THE HARSHEST MONTHS
+1.1–1.4 pts
May, August and September are the harshest months for the rating
The share of negative reviews runs 1.1–1.4 points above the game’s own average. A holiday release only looks worse because its “before” baseline is inflated.
- J
- F
- M
- A
- M
- J
- J
- A
- S
- O
- N
- D
kinderharsher than the game’s average
- 266 games
How much data. 2.22 million reviews, counted from Steam’s own counters, were written in Novembers. On the new games it is 537 thousand reviews: the share of negative reviews is 9.5% (first games) and 7.0% (new games) in November, against 12.9% and 10.3% in the other months.
What to do
Take the seasons into account when you plan: November to January is the peak of activity and brings the kindest reviews of the year.
- The harshest months. The most negative months are May, August and September (1.1 to 1.4 points above a game’s own average). December is also kinder than average (104 of 135 games). Across all 266 games, November sits 3.2 points below a game’s own yearly average, and in 88% of games November is kinder than their average.
- Review volume, without a statistical model. November is 1.36 times a game’s average month, December ×1.21, September ×0.82. We did not test the cause of the November peak; in time it coincides with the autumn sale and the Steam Awards nominations. A holiday release only “looks” worse because its “before” baseline is inflated.
- 130 gamesNovember is the kindest monthReviews behind it537 thousand countedConfirmed
Kinder than the game’s own average in 112 of 124 games (90%).
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- We did not test why November is kind. It coincides in time with the autumn sale and the Steam Awards.
- Review volume here means English-language reviews per month.
Chapter 09
Free-to-play versus premium
Updates and posts work less well in free-to-play games than in premium ones: in the month of a major release, the free-to-play lift is about 0.8 of the premium lift. Our 50-game rule is met only when the first 136 games are added, which is why the label is “suggestive”.
Suggestive
Figure 9
A major release lifts free-to-play games much less than premium ones
The same two effects in premium and subscription games and in free-to-play games.
- premium and subscription
- free-to-play (F2P)
- no effect = ×1.00
Major-release month
×1.2×1.4×1.6×1.00×1.47+47%×1.16+16%premiumF2P
Post → next month, %
2%4%6%8%+7.2%+1.4%premiumF2P
Monthly estimates across all games. The fact cards below also include the extra free-to-play games.CritField · Issue 02
MAJOR RELEASE
×1.43 vs ×1.13
A major release lifts free-to-play games much less than premium ones
The lift in the month of a major release: premium ×1.43, free-to-play ×1.13.
- 53 free-to-play games
- suggestive
STUDIO POSTS
+7.2% vs +1.4%
A studio post barely moves free-to-play games
Reviews the next month: +7.2% in premium games, +1.4% in free-to-play, which can’t be told apart from zero.
- 56 free-to-play games
- suggestive
SILENCE
×0.58 / ×0.64
A year without releases hurts both kinds of game
Reviews after a year or more without a release: premium ×0.58, free-to-play ×0.64 (15 games).
- 15 free-to-play games
NOT PROVEN YET
32 of 50
On new games, the free-to-play gap is still short of our bar
The new games point the same way (×0.84), but our rule needs 50 free-to-play games, so the label stays “suggestive”.
32 of the 50 games the rule needs
- 50-game rule
How much data. 78 free-to-play games in the combined data, with 4.14 million reviews in their monthly counters. The 50-game rule is met only on the combined data (61 free-to-play games with a major-release month, 74 with posts).
| Effect (monthly) | Premium and subscription | Free-to-play |
|---|---|---|
| Month of a major release | ×1.47 (163 games) | ×1.16 (49) |
| Studio post → next month | +7.2% (200) | +1.4%, not distinguishable from zero (56) |
| +10 points of negative reviews → reviews three months later | −10.3% (206) | −8.4% (60) |
| A year without releases | ×0.58 | ×0.64 (15 games) |
- Why? The likely reason is retention mechanics inside the games, but that is a guess. We did not work out how many individual games point the same way for these models, and the new games alone (32 free-to-play games) are under 50.
- After adding games. The ten free-to-play games among the 80 added show no gap on their own (×1.08, 8 with a major release). On all games together (354 games, 61 of them free-to-play with a major-release month) the free-to-play lift in a major-release month is ×0.81 of the premium lift, and the adjustment for a post is −0.059 (74 free-to-play games). The new games alone give ×0.84 on 32 free-to-play games: same direction, but under 50 games, so the label stays .
- 210 gamesFree-to-play: a major-release month helps lessReviews behind it10.02 million countedToo few games
28 free-to-play games, below the 50 we need to call it.
- 210 gamesFree-to-play: studio posts help lessReviews behind it10.02 million countedToo few games
33 free-to-play games, below the 50 we need to call it.
Before we looked at the new games, we wrote down each prediction and what would count as a pass. “130 games” marks that first test; “210 games” adds 80 more games. “Counted” means tallied from Steam’s own counters; “read” means review texts we read and coded.
- The 50-game rule is met only on combined data that includes the first 136 games.
- We did not work out how many individual games point the same way for the free-to-play comparison.
Part 05 of 05
How we know
The method in plain words, then everything that did not hold up, where we still need data, the limits, a glossary and the full method for specialists.
What this is, and how we did it
What this is
Studios ship updates all the time, but few can say what an update actually does for their audience. We followed 354 Steam games to find out. After a patch lands, and during the quiet stretches between patches, do players come back? Does the rating suffer? What do people write about, and which reviews does a buyer see first? The raw material is public: the studios’ own news posts on one side, the players’ reviews on the other.
How to read the labels
- Strong. Seen in at least 50 games, pointing the same way in at least 60% of them, and repeated in both halves of our data.
- Suggestive. A broad pattern, but it shows up in one direction only or in one kind of measure.
- No effect. We looked across many games and found nothing.
- Held up again on new games we tested afterwards.
- Not confirmed. The evidence is too thin or too fuzzy. That is uncertainty, not proof of the opposite.
- 652 thousandreview texts read and coded; 103,387 of them negative
- 354Steam games: 136 to look for patterns, 218 more to test them
- 26.8 thousandreleases logged from developers’ news posts, hotfixes included; about 7% are major releases or DLC
- 18.3 millionreviews counted in the months around 5,284 single releases
- 13 of 20predictions that held up on 130 new games. We wrote them down before looking.
How we did it
- 1We logged the updates. We collected the news posts of 354 games that were still being updated and sorted them into a log of 26,774 releases, every build a studio shipped: what, when and how big. About 7% are major releases or DLC, half are content or balance updates, and a third are hotfixes and small patches. We checked the sorting blind against a sample.
- 2We matched them to the reviews. For every update, and every quiet stretch between updates, we looked at the reviews the same game received: how many, how many were negative, who wrote them and about what. We read and coded 651,642 review texts. Another 21.31 million reviews we did not read: we took their counts straight from Steam’s own monthly counters.
- 3We hunted for patterns, carefully. We searched the first 136 games with safeguards against fluke results. The games were split into two halves, and a pattern counted only if it showed up in both. It also had to appear in at least 50 games and point the same way in at least 60% of them.
- 4Then we tested the patterns on fresh games. We tried every pattern on 130 other games (different developers, publishers and franchises). Before touching that data we wrote down exactly what we would test and what would count as a pass. Where a question was still short of 50 games, we added 80 more.
What doesn’t matter, and what only looked like a finding
Beliefs we tested across many games and found no support for, and early results that fell apart when we checked them. One line each, with the number of games behind it.
Tested on many games, no effect found (18)
| Belief | What we found | Games |
|---|---|---|
| How you package an update (announcement, tone, video, “based on your feedback”, release day) changes how players react | No feature showed a repeatable effect (about 40 features against 10 measures of reaction) | 135 |
| Paid content in an update triggers a wave of negativity | 0.0 points in our first search. On new games it is murkier: +3.1 points (95 games) and not stable from one set of games to the next, so we treat this one as unsettled | 76 |
| Removing features triggers a wave of negativity | −0.9 points, not distinguishable from zero | 73 |
| A quick hotfix within a week of a major release is a sign of failure | +0.7 points, not distinguishable from zero in both halves of the data | 125 |
| A major release in Q4 works better | The same lift as at other times of year | 110 |
| Late in a game’s life, major releases work less well | Game age makes no difference | 110 |
| Releasing in a Steam sale month gives an extra lift | ×0.96: no extra lift | 136 |
| Regular updates keep the audience | No effect on review volume or rating | 136 |
| A long pause builds anticipation | A release after a pause brings reviews back to normal, not above it | 126 |
| A quiet spell spoils the rating | −0.07 points: no | 127 |
| Discounts spoil the rating by bringing in “outsiders” | More newcomers, same rating | 59 (49 checked game by game) |
| Free keys spoil or inflate the rating | Within a game, reviews from keys are slightly more positive | 136 (47 checked game by game) |
| A content update 1–2 months after a major release amplifies it | Nothing beyond the sum of the two separate effects (+4%, not distinguishable from zero) | 87 |
| Posts only work at the start of a pause | Posts help just as much however long the pause | 130 |
| A studio’s reaction to a negative spike (a release, a roadmap) speeds up recovery | No | 69–82 |
| One kind of complaint is more dangerous for audience loss | Beyond the overall share of negative reviews, no topic predicts loss | 136 |
| Monetization announcements (store, paid DLC, prices) provoke a reaction | No change compared with other posts; the labels are noisy | — |
| Rhythm (slowing down, speeding up, growing irregular) changes the share of negative reviews | No change over 180 days | 34 and 16 |
Looked like a finding, but wasn’t (9)
| It looked like | Why it isn’t |
|---|---|
| A long pause leads to a burst of reviews after the release | The starting level was simply low. In absolute terms there is no extra effect. |
| A release during a Steam sale lifts activity by 29% | Every game grows in sale months. Beyond the calendar, the release adds nothing. |
| Balance patches cause a lasting rise in negativity | It vanishes against ordinary dates: it is the game getting older. |
| A change in Steam’s rating label at a threshold is a release effect | It happens without releases too: it is the noise of a 30-day stretch. |
| The longer the pause between major releases, the stronger the next one | With release size taken into account there is no effect (61 games). |
| New characters mean less negativity | It held in only 36 of 72 games, and the second half of the data did not confirm it. |
| A holiday release brings more negativity | It is the season: November is the kindest month, so the “before” level is inflated. |
| A surprise major release is worse than an announced one | The “surprise” group is contaminated (platform launches, quieter games). On 56 games the estimate points the other way (−1.2 points). |
| Negativity is stronger in Early Access | After leaving Early Access the same games age in the same way. |
“No effect” here means “no effect larger than the margin we state”. Where our check could not narrow the range enough, the chapters mark it as uncertainty, not as an opposite effect.
Not enough data yet
We only call something a finding if at least 50 games show it and at least 60% of them point the same way. After adding 80 games, some questions that were short of that bar have cleared it, and others have not. This is the honest ledger: what the top-up settled, what still has fewer than 50 games, and what has enough games but an answer too fuzzy to call “no effect”. Nothing in the last two groups is a finding. They are the shortlist for the next study, and none of our chapter labels rests on them.
What the top-up settled
- 50
Paid content in an update
Reached 50Games before47Games after95
Now testable. The answer is not a clean “no effect”: it is a positive shift of about 3 points that is not stable from one set of games to the next (see below).
- 50
Rating after a major release or DLC
Reached 50Games before44Games after86
Enough games now. Negative reviews rise about 1.3 points in the first month (range +0.8 to +1.8), so we cannot call it “no effect”.
- 50
Rating after a balance patch
Reached 50Games before36Games after59
Enough games now. +0.6 points (range −0.3 to +1.4): too fuzzy to say “no effect”.
- 50
A release that only weakens things
Reached 50Games before48Games after84
Enough games, but the effect did not repeat: only 5 of 11 games moved the same way.
- 50
Fewer newcomers after hotfixes
Reached 50Games before43Games after64
Enough games, but only 59% moved the same way, just under our 60% bar, so no conclusion. With 190 games it was 61% and we had called it confirmed.
- 50
Studio posts and the rating in quiet spells
Reached 50Games before47Games after164
Confirmed: a month with a post has the same share of negative reviews as a month without one (within 1 point).
- 50
Announcing a delay
Reached 50Games before15Games after82
Enough games, but too fuzzy to say “no effect”. A hint of about 10% more reviews the next month (range ×1.02 to ×1.19), which we did not test as a prediction.
- 50
Posts about monetization changes
Reached 50Games before7–39Games after64
Enough games, but too fuzzy to say “no effect”. No hint of any effect.
- 50
Apology or explanation posts
Reached 50Games before22Games after107
Enough games, but too fuzzy to say “no effect”. A hint of about 1 point more negative reviews in the same and the next month (range +0.1 to +1.9).
- 50
Surprise releases against announced ones
Reached 50Games before47Games after56
Enough games, and the estimate points the other way, so the idea that surprises hurt more did not hold.
Still under 50 games
- 50
Early Access against finished games
9 to goGames we have41(34 found earlier, 7 added)
No eligible games are left among our 800 candidates. We need a new candidate list.
- 50
Studios rolling back a change
9 to goGames we have41
Under 50 games, and we never set up a test for it in advance.
- 50
Compensation for players; shutting a game down
counted onlyGames we have60; 60
Enough games only when both groups are added together, and we did not set an outcome in advance. Counted, not concluded.
- 50
Matchmaking complaints in reviews
11 to goGames we have39
Under 50 games: only the first 136 games had their review texts coded.
- 50
Review topics as an early warning
27 to goGames we have23–29games per topic
Under 50 games per topic. We need reviews coded in a few dozen more games.
- 50
Free-to-play: the check on new games
18 to goGames we have32free-to-play games with a major release among the new games
The 50-game rule is met only when the first 136 games are added (61 and 74 games).
Enough games, not enough precision
Announcing a delay
82 games+0.2 points (−1.2 to +1.5)
The range of doubt is wider than the ±1.5-point margin we set in advance, so we can’t call it “no effect”.
A post about monetization changes
64 games−0.4 points (−2.4 to +1.5)
The range of doubt is wider than the ±1.5-point margin we set in advance, so we can’t call it “no effect”.
An apology or explanation post
107 games+1.0 points (+0.1 to +1.9)
The range of doubt is wider than the ±1.5-point margin we set in advance, so we can’t call it “no effect”.
Paid content in an update
95 games+3.1 points (0.0 to +6.1)
The estimate sits above zero and is not stable between sets of games (+5.2 on the first 130 new games, +0.7 on the 80 added games). Only 25 of 34 games with both kinds of release point up, and that is under 50 games.
Release rhythm and review volume
178 games×1.11 (1.00 to 1.24)
The range of doubt is wider than the ±10% margin we set in advance.
A pause and the share of negative reviews
108 games−0.60 points (−1.89 to 0.68)
The range of doubt is wider than the ±1.5-point margin we set in advance.
Number of content updates in a year
190 games×1.02 (0.90 to 1.16)
The range of doubt is wider than the ±10% margin we set in advance.
“Not enough precision” means we measured against a margin set in advance and the range of doubt came out wider than that margin: we can neither confirm “no effect” nor claim an effect. Paid content is the one line where the larger data moved against an earlier “no effect”. Our first search gave 0.0 points; the new games give a positive shift that is not stable. We keep the first-search result in the list of null results, with this caveat, and treat the line as unsettled.
What this study cannot tell you
What the numbers measure
Reviews are not players
We count how many reviews a game gets in a month and use that as a stand-in for how active its audience is. It is not a head-count of players.
No sales, no retention, no player counts
We have no data on revenue, concurrent players, wishlists or how long players stay. Nothing here should be read as an effect on money.
Steam only, English reviews only
We chose the games for variety, not to represent all of Steam.
The storefront
“Helpful” votes are a single snapshot from the day we downloaded the data. We control for the age of a review but cannot see Steam’s ranking algorithm.
What the data cannot settle
Associations, not causes
Comparing a game with itself removes differences between games. It cannot rule out that a studio ships an update when things are already going well.
Many tests
In the exploratory parts we ran many comparisons without formally correcting for it. We picked findings by whether they repeated across games, not by p-values.
Steam’s counters
Our counts are English-language reviews of all purchase types, including off-topic review activity. The score Steam shows leaves out key activations and counts all languages, so our numbers match Steam’s labels only roughly.
How the study was run
AI sorting of updates
An AI model picked out the releases and the details of the patch notes. A blind second rating checked its release calls: of the posts it called releases, 88% were, and it caught 78% of the real ones. Reliable details were checked by hand, unreliable ones were replaced by keyword rules. AI models also tagged the review texts against a fixed codebook.
Labelling of marketing events
Sales, monetization posts and announcements are sorted by titles and labels, which is noisy: “store”, for example, mixes in-game shops with game editions.
Developer replies
We do not treat replies to reviews as a lever: Steam advises against stepping into the review space. That observation stays in the scientific report.
Terms in plain words
Pick a term to see what it means on this page.
Activity (review volume)
How many reviews a game gets in a month, counted by Steam. We use it as a stand-in for how active the audience is. It is not a head-count of players.
Negative share
The share of a game’s reviews that are negative (“Not Recommended”).
Release
A new version shipped to all players, according to the developer’s news. We sort releases into major releases and DLC, content updates, balance changes and hotfixes.
Isolated release
A release with no other release within two weeks either side, so its before and after are not muddied.
Ordinary dates
Dates in the same game with no release. We hold a release up against them to separate its effect from drift and season.
Multiplier (×1.33)
How many times as many reviews as usual. ×1.00 means no change. ×0.60 means 40% fewer.
Percentage points (pts)
The distance between two percentages: going from 10% to 11% is one percentage point.
Range of doubt
The span in which the true value very likely sits (we show 90% and 95% ranges). A narrow range means a precise estimate. If the range includes zero (or ×1.00), we can’t be sure there is an effect.
Within a game
We compare a game with itself at other times, so differences between games do not distort the result.
Two-halves check
We split the games at random into two groups. A pattern counts only if it shows up in the first group and then repeats in the second.
The 50-game rule
Our rule for calling something a finding: at least 50 games that show the feature, and the same direction in at least 60% of them.
Tested on new games
Repeating the test on games that were not part of the search, after writing down in advance what would count as a pass.
The margin for “no effect”
To say “no effect”, we set a margin first (for example ±1.5 points) and require the whole range of doubt to fit inside it.
Verdicts
Confirmed. Not confirmed (we are uncertain; we are not saying the opposite is true). Too few games (we could not test it). Only on average (the average effect is real, but fewer than 60% of games point the same way, so it is not a rule).
Early Access
The period when a game is sold before its official release.
Free-to-play (F2P)
Games that are free to start, with in-game purchases.
Storefront
The reviews Steam shows first, picked by “helpful” votes.
For specialists: the full method
This is the full method, written the way a methods paper would be: the question, the units of analysis, the measurements, the estimation, the controls against false findings and the status of every conclusion. The numbers are the same as in the main part.
Research question and status of the conclusions
Question. How are game updates, and the pauses between them, associated with audience behaviour on Steam: with the flow of reviews (a proxy for the active audience), with the share of negative reviews (the rating), with the mix of authors, and with what they write about?
Status. An observational study on public data. All estimates describe associations within games. We make no causal claims, draw no conclusions about sales, retention or revenue (we have no such data), and do not generalize to "all of Steam": the games were chosen for maximum variation, not as a representative sample.
Design: discovery, then independent confirmation
| Stage | What was done | Games | Safeguards against false findings |
|---|---|---|---|
| 1. Discovery | Review corpus and monthly series; a hypothesis machine over about 800 tests | 136 | split-half by game, Benjamini–Hochberg correction, breadth rule |
| 2. Independent confirmation | 20 hypotheses; the plan was frozen before the data were fetched (SHA-256 of the plan in plan_freeze.json) | 130 new games (different developers, publishers, franchises) | pass criteria set in advance; deviations logged in DEVIATIONS.md |
| 3. Top-up to the breadth rule | Extra games where a line fell short of 50 games (plan F, frozen before labelling) | +80 games | a mechanical selection rule independent of outcomes |
Discovery sample: maximum variation over game age and update cadence (an age × cadence matrix) across three cohorts (live, winding down, control),
to avoid survivorship bias. Confirmation sample: 130 games drawn at random from 314 eligible ones (app_type = 0; at least 3,000 reviews; at least 60 news items;
released no later than 2024-06-01; no developer, publisher or franchise shared with the discovery games) and stratified by monetization model.
Units of analysis and data
- The game is the unit of independence. Standard errors are clustered by game, and the effective sample size is bounded by the number of games, not of reviews (reviews within a game are dependent). The number of games is therefore the condition for a conclusion; the number of reviews measures precision and coverage.
- Game × month. Steam's monthly counters (an
appreviewsquery with a date range,query_summary): the number of English reviews and of negative reviews. Activity is the logarithm of the number of reviews in a month. - Release × window. Windows before and after an isolated release (30 days before and three 30-day windows after; isolated means no other release within ±14 days), compared with "placebo dates" without releases in the same game.
- The review. Texts and metadata (hours played at the time of the review, the author's number of reviews, "helpful" votes).
- The release as an event. Taken from developers' news feeds: a frozen classifier plus rules, with a blind check (post precision 0.88, recall 0.78 in discovery). Types: major release or DLC, content update, balance change, hotfix.
- Patch-note features (about 20: kind of changes, size, announcement, paid content, removals, event and so on) are extracted by the same classifier. Reliable features were checked by hand; unreliable ones were replaced by keyword rules.
Models
| Family | Specification | Estimate |
|---|---|---|
| Monthly panel | y(game, month) ~ predictors + game FE + calendar-month FE + log(game age), SE by game; by developer as well in the check | activity multiplier (exponent of the coefficient on a log outcome) or pp of negative share |
| Windows around releases | ratio "after/before" relative to the same ratio on placebo dates of the same game; negative share as a difference in pp | ×-effect, 90% and 95% confidence intervals |
| Within release (reviews) | within-release DiD: shift of a review feature after the release against the "before" window, game and release effects absorbed | pp |
| Yearly panel | complete game-years, game and year FE | multiplier |
| Early warning | log(activity t+3) ~ negative share t + log(activity t); separately the probability of an activity drop larger than 20% | % per pp, probability ratio |
| Aspects | within-game linear probability model on 38,208 coded reviews | pp |
Activity effects were checked in two ways (as a ratio to the "before" period, and as a level relative to the game's norm), and seasonal effects relative to other games in the same calendar month. This removed plausible but false findings (see "What doesn't matter").
Control of false findings
- Split-half. Games are split at random (seed 336) into two halves. A finding counts if it is found in one half with BH correction (q < 0.10) and repeats in the other (one-sided p < 0.05, same sign).
- Breadth rule (binding). A conclusion needs at least 50 games in which the feature occurs, and the sign of the effect must agree in at least 60% of the games with data. Everything else is "insufficient data".
- Independent confirmation against a frozen plan. Three outcomes: CONFIRMED, NOT CONFIRMED, NOT TESTABLE (breadth not met). The criteria were never revised.
- "No effect" claims are equivalence tests. A bound is set in advance (±10% for activity, ±1 pp or ±1.5 pp for the negative share) and the 90% confidence interval must lie entirely inside it. If the interval is wider, the verdict is "not confirmed", not "no effect": there are not enough data to rule out a small effect.
- Placebos and artifacts. Placebo dates; comparison with other games in the same calendar month; re-checking findings under two definitions of a release (the primary "release" definition and the sensitivity "substantive" one).
The review-coding instrument
Reviews were coded with a closed codebook of 10 codes, fixed in advance. A reference coding covers 400 reviews; the test part (239 reviews) was coded blind. Agreement of the coding with the reference on the test part (Cohen's κ): hope/patience 0.85, disappointment 0.69, update talk 0.61, anger 0.46, developer blame 0.43, developer praise 0.32 (not interpreted). "Resignation" cannot be evaluated (2 positive examples). Conclusions about review texts (the chapters "Who writes, and what the buyer sees" and "What players forgive") therefore rely on codes that passed validation, and they are presented as exploratory until an independent sample confirms them.
Results of the independent confirmation (130 games, frozen plan)
Of 20 hypotheses, 13 were confirmed and 7 were not. The unconfirmed ones fall into two kinds: a shortfall of significance (C1 in the third month, p = 0.057; C9, p = 0.067) and five "no effect" claims (C2, C5, C6, C12, C19) whose intervals were wider than the bound. No hypothesis turned out to point the opposite way.
- C1Major releases and DLC raise activity against placebo dates (months 1 and 3)Reviews behind it552 thousand countedNot confirmedm1 holds, m3 misses p < 0.05
month 1 ×1.72 (p < 1e-9; above placebo in 38/45); month 3 ×1.14 (p = 0.057; 21/44); 51 games
- C2No release type moves the negative share against placebo (±1.5 pp, months 1–3)Reviews behind it4.06 million countedNot confirmedequivalence not shown
3 of 12 intervals outside the bounds; DLC (44 games) and balance (36) below 50 games
- C3A major-release month raises activity and the effect fadesReviews behind it620 thousand countedConfirmed
×1.34 in the release month, ×1.22 the next; 102 games
- C4Activity falls with the length of a pauseReviews behind it742 thousand countedConfirmed
12+ months without a release ×0.53; 63 games
- C5Release regularity does not affect activity (±10%)Reviews behind it5.51 million countedNot confirmedequivalence not shown
×1.11, 90% CI [0.98, 1.27] outside the bounds; 100 games
- C6A pause does not change the negative share (±1.5 pp)Reviews behind it5.51 million countedNot confirmedequivalence not shown
−0.12 pp, 90% CI [−1.76, +1.52] slightly wider; 63 games
- C7November is the most positive monthReviews behind it537 thousand countedConfirmed
more positive than the game's mean in 112 of 124 games (90%)
- C8A "large" update raises first-month activity beyond its typeReviews behind it2.06 million countedConfirmed
×1.29 (p < 1e-4); 82 games, 1,734 releases
- C9A new mechanic raises first-month activity beyond its typeReviews behind it2.06 million countedNot confirmedshortfall of significance
×1.06 (p = 0.067); 95 games; discovery +17–22%
- C10Among negative reviews the share of relationship complaints rises with playtimeReviews behind it25.4 thousand readConfirmed
+3.6 pp per log-hour; higher for veterans in 100 of 128 games
- C11"Helpful" reviews (10+ votes) are more negative than all reviewsReviews behind it5974 readConfirmed
102 of 118 games; median ×1.96
- C12The post-release shift in negativity does not depend on announcement, paid content, removals, event, "your feedback", video, weekend (±1 pp)Reviews behind it259 thousand countedNot confirmedequivalence not shown
every 90% CI wider than ±1 pp (about ±2); paid content 47 games; 119 games
- C13A studio post (not a release) in month t raises activity in t+1Reviews behind it2.40 million countedConfirmed
+5.5% per post; 126 games
- C14A roadmap / announcement raises activity in t+1Reviews behind it2.40 million countedConfirmed
+6.0% per post; 124 games
- C15The negative share in month t predicts lower activity in t+3Reviews behind it5.51 million countedConfirmed
−0.95% per pp; 70 of 113 games negative
- C16Prolific reviewers are harsherReviews behind it146 thousand readConfirmed
+1.7 pp per log-review; 126 of 130 games
- C17A negative review more often gets 10+ "helpful" votesReviews behind it146 thousand readConfirmed
+4.7 pp on a 4.1% base; 109 of 130 games
- C18A year with a major release is more activeReviews behind it5.51 million countedConfirmedsmaller effect
×1.08 (discovery ×1.16); 89 games
- C19The number of content updates a year does not change activity (±10% for 4)Reviews behind it5.51 million countedNot confirmedequivalence not shown
×1.03, 90% CI [0.90, 1.19] wider; 112 games
- C20A negative spike of 5+ pp raises the risk of a 20%+ activity dropReviews behind it200 thousand countedConfirmed
41% vs 25%, lift 1.69; 50 games
"Counted" means reviews tallied by Steam's counters in the months and windows the estimate rests on; "read" means review texts that were read and coded.
Top-up to the breadth rule (plan F)
- C1 (210)Major releases and DLC raise activity against placebo (months 1 and 3)Reviews behind it10.57 million countedVerdict on the poolConfirmedVerdict on 130 gamesNot confirmed
month 1 ×1.50 (p < 1e-4; above placebo in 66 of 88); month 3 ×1.16 (p = 0.005; 50 of 87); 97 games
- C2 (210)No release type moves the negative share (±1.5 pp)Reviews behind it10.57 million countedVerdict on the poolNot confirmedVerdict on 130 gamesNot confirmed
inside the bounds in 6 of 12 intervals (types with 50+ games); major release/DLC: 86 games; content update: 160 games; balance change: 59 games; hotfix: 136 games
- C5 (210)Release regularity does not affect activity (±10%)Reviews behind it10.02 million countedVerdict on the poolNot confirmedVerdict on 130 gamesNot confirmed
×1.11, 90% CI [1.00, 1.24]; 178 games
- C6 (210)A pause does not change the negative share (±1.5 pp)Reviews behind it10.02 million countedVerdict on the poolNot confirmedVerdict on 130 gamesNot confirmed
−0.60 pp, 90% CI [−1.89, 0.68]; 108 games
- C19 (210)The number of content updates a year does not change activity (±10% for 4)Reviews behind it10.02 million countedVerdict on the poolNot confirmedVerdict on 130 gamesNot confirmed
×1.02, 90% CI [0.90, 1.16]; 190 games
- C8 (210)A "large" update raises first-month activity beyond its typeReviews behind it10.57 million countedVerdict on the poolConfirmedVerdict on 130 gamesConfirmed
×1.30 (p < 1e-4); 143 games; sign per game 74 of 90
- C9 (210)A new mechanic raises first-month activity beyond its typeReviews behind it10.57 million countedVerdict on the poolConfirmedon p; sign per game 56%, below 60%: no per-game claimVerdict on 130 gamesNot confirmed
×1.06 (p = 0.037); 164 games; sign per game 70 of 125
- C12 (210)The post-release shift in negativity does not depend on announcement, paid content, removals, event, "your feedback", video, weekend (±1 pp)Reviews behind it388 thousand countedVerdict on the poolNot confirmedVerdict on 130 gamesNot confirmed
inside the bounds for 0 of 7 features; paid content +3.1 pp [0.0, +6.1] (95 games)
- E4A nerf-only release raises the next-month thumbs-down shareReviews behind it388 thousand countedVerdict on the poolNot confirmedVerdict on 130 games—
−1.99 pp; 84 games; sign 5 of 11
- E5The share of newcomers falls after hotfixesReviews behind it388 thousand countedVerdict on the poolNot confirmedper-game sign < 60%Verdict on 130 games—
−1.50 pp; 64 games, a fall in 38 of 64
- E7Negativity after an unannounced major release is larger than after an announced oneReviews behind it388 thousand countedVerdict on the poolNot confirmedVerdict on 130 games—
−1.22 pp; 56 games with both kinds
- E2F2P: a major-release month helps lessReviews behind it10.02 million countedVerdict on the poolToo few gamesbreadth: 28 < 50 gamesVerdict on 130 games—
coefficient −0.195; 28 F2P games
- E3F2P: studio posts help lessReviews behind it10.02 million countedVerdict on the poolToo few gamesbreadth: 33 < 50 gamesVerdict on 130 games—
coefficient −0.038; 33 F2P games
- F3Studio posts in quiet spells do not change the negative share (±1 pp)Reviews behind it1.31 million countedVerdict on the poolConfirmedVerdict on 130 games—
−0.16 pp [−0.66, 0.34]; 164 games
- F5A delay announcement does not change negativity or activityReviews behind it143 thousand countedVerdict on the poolNot confirmeda 90% CI exceeds the boundsVerdict on 130 games—
same month +0.2 pp [−1.2, +1.5]; next month −0.7 pp [−1.7, +0.3]; activity ×1.10; 82 games
- F6A monetization-change post does not change negativity or activityReviews behind it145 thousand countedVerdict on the poolNot confirmeda 90% CI exceeds the boundsVerdict on 130 games—
same month −0.4 pp [−2.4, +1.5]; next month −0.6 pp [−2.6, +1.5]; activity ×1.02; 64 games
- F7An apology / explanation post does not change negativity or activityReviews behind it558 thousand countedVerdict on the poolNot confirmeda 90% CI exceeds the boundsVerdict on 130 games—
same month +1.0 pp [+0.1, +1.9]; next month +0.9 pp [+0.1, +1.6]; activity ×1.03; 107 games
The extended pool is the 130 confirmation games plus the 80 new games (60 + 20 in a second tranche), independent of discovery. The frozen verdicts on 130 games were not changed; they stand next to the new ones. "Counted" means reviews tallied by Steam's counters in the months and windows the estimate rests on (for review-level tests, the reviews in the ±30-day windows around releases).
The same headline estimates re-run on all sets together (discovery 136 + confirmation 130 + top-up 80 = 346 games; exploratory, like the earlier pooled estimates):
| Headline estimate | All sets together | Games |
|---|---|---|
| Major-release month: activity | ×1.34 (95% CI 1.26–1.42) | 346 |
| The month after | ×1.20 | 346 |
| A year or more without releases | ×0.60 (95% CI 0.54–0.68) | 169 |
| Any studio post → next month (per doubling of posts) | +5.0% | 336 |
| Roadmap post | +5.4% | 330 |
| +10 pp negative → activity three months later | −9.3% | 346 |
| A year with a major release (per doubling) | ×1.11 | 257 |
| A negative spike of 5+ pp: risk of an activity drop | 39% vs 24% | 148 |
| First month after a major release against control dates | ×1.55 (95% CI 1.39–1.72) | 164 |
Reproducibility
The corpus is frozen (scripts/corpus_v2/freeze.py --verify), the verification plans are frozen by hash, every stage can be reproduced with the scripts in the
repository (scripts/, reports/), and deviations are logged. Code and versions are listed in the appendix.
Questions about the data or the method?
Ask about a number, a test or a caveat. A researcher replies.