More than half our sample is pre-launch, so we finally tested the assumption everyone carries into a pre-launch review: that a Coming Soon page is a worse page. Matched against released pages of the same era and the same obscurity, 208 Coming Soon pages score the same on all five rubric columns, carry the same number of issues, and differ on exactly one thing: media volume. They run a mean of 1.20 trailers against 2.03 (p < 0.0001) and a median of 8 screenshots against 10 (p = 0.036). They are not worse pages. They are thinner pages.
Two caveats first. This is a convenience sample of pages somebody chose to paste into a free tool, not a random draw from Steam. And the scores are our own AI rubric, not a Valve metric. No sales, wishlist or traffic data is joined to any of this, so nothing below is a claim about revenue.
Why the obvious comparison gives the wrong answer
Split the 373 pages by release status and pre-launch looks worse across the board. Coming Soon pages (n = 208) against all released pages (n = 165) gives a median overall of 66 against 69 and a median tags subscore of 62 against 70, and that tags gap comes back at p = 0.0001. Case closed, apparently.
It is not closed, because the released group is not a fair comparison. It contains every famous game anyone ever pasted into our tool out of curiosity: 32 of those 165 pages carry a Metacritic score, and 85 of them released before 2025. You are comparing indie pre-launch pages against a group padded with Valve, FromSoftware and Rockstar. Any difference you find is at least partly a difference in studio size.
How we built a fair control group
The control is released pages from 2025 or 2026 with no Metacritic score: 77 pages, 50 from 2026 and 27 from 2025. Same era, same level of obscurity, same kind of team. The only thing that differs is that their launch date has passed.
The matching is deliberately coarse. It is a two-variable filter on era and press coverage, not a propensity score, and we do not have the covariates (budget, team size, wishlist count) that a proper matched design would want. Read it as a strong sanity check on the naive comparison rather than as a causal design.
Do Coming Soon pages score worse than released pages?
No. On the matched comparison, every score column ties.
| Score | Coming Soon (208) | Matched control (77) | Median p | Mean difference 95% CI |
|---|---|---|---|---|
| Overall | 66 | 68 | 0.429 | [-3.78, +0.97] |
| Capsule | 72 | 72 | 1.000 | [-3.79, +1.78] |
| Description | 64 | 64 | 1.000 | [-3.03, +3.50] |
| Screenshots | 72 | 78 | 0.232 | [-5.46, +0.40] |
| Tags | 62 | 62 | 1.000 | [-5.65, +1.27] |
Three of those median p-values are exactly 1.000 and none of them means anything on its own. Capsule lands on exactly 72 on 152 of the 373 pages in the corpus and tags on exactly 62 on 87 of them, so any two subgroups of reasonable size will tie on the median by construction. That is why the right-hand column is there: the bootstrap interval on the mean difference is what tells you how large an effect the data still allows, and on the overall score it rules out a pre-launch penalty bigger than about 3.8 points.
A note on power, because the honest version is awkward. The permutation test on medians is badly behaved on a lumpy discrete score: its power depends on which group you apply the simulated shift to, giving 0.27 in one direction and 0.84 in the other at a 4-point gap. That is a broken instrument, not a weak one. The permutation test on means does not have the problem and reaches power 0.96 at a 3-point gap on this sample, so the null on the overall score is meaningful.
Is the issue load any different before launch?
Also no, which surprised us more than the scores did. Both groups carry a median of 6 issues (p = 1.000; means 6.60 against 6.42, p = 0.384). 61.1% of Coming Soon pages carry at least one critical issue against 55.8% of the control (p = 0.498), and both groups sit at a median of exactly 1 critical issue.
An unfinished game does not, on this evidence, get an unfinished page. Whatever it is that makes a store page score badly, it is present in full before launch and is not something that gets cleaned up in the last month.
What actually separates a pre-launch page?
Media volume, and only media volume. It is the one dimension that survives matching, and it survives it convincingly.
Trailers. Coming Soon pages average 1.20 trailers against 2.03, a difference of -0.82 with a 95% CI of [-1.34, -0.40] at p < 0.0001. The medians tie at 1, which is the mass point problem again, so look at the distribution instead: 19.2% of pre-launch pages have two or more trailers against 41.6% of the control (p = 0.0002), and only 4.8% have three or more against 26.0% (p < 0.0001). At the other end, 9.6% have no trailer at all against 1.3% (p = 0.019).
Screenshots. Median 8 against 10 (p = 0.036), mean 9.00 against 10.64 (difference -1.63, CI [-2.74, -0.49], p = 0.0067). 52.9% of Coming Soon pages sit below 9 screenshots against 35.1% of the control (p = 0.0086), and 43.8% sit below 8 against 27.3% (p = 0.014).
Text volume does not move. Short description length is a median of 232 characters against 228 (p = 0.683), and detailed description length 4,616 characters against 4,881 (p = 0.451, mean difference interval [-787, +429]). Whatever pre-launch teams are short of, it is not copy.
Why the median of 8 screenshots is the number to worry about
Our store page audit found that the useful screenshot band starts around 8. We re-ran it here on the full corpus: pages with 7 or fewer screenshots have a median screenshot subscore of 68, pages with 8 to 12 have 78 (p = 0.010), and pages with 13 or more have 78 as well, a difference of exactly 0 at p = 1.000. The gain is a step at the bottom, not a slope.
Pre-launch pages sit at a median of exactly 8, the bottom edge of that band. 43.8% of them carry fewer than 8 screenshots and 52.9% carry fewer than 9, against 27.3% and 35.1% in the matched control.
39 of the 208 pre-launch pages sit at exactly Steam’s five-screenshot minimum. Inside the Coming Soon group, pages with 9 or more screenshots have a median screenshot subscore of 78 against 68 for pages below that (p = 0.015), which is the same step the full corpus shows.
One hard limit on this claim: our analyzer sends only the first 10 screenshots to the model, and 121 of 373 pages (32.4%) have more than 10. Nothing here says anything about screenshots 11 and up, in either direction. The screenshot checker covers ordering and coverage rather than count, which is the part that actually varies once you are past 8.
Does the trailer gap mean you should cut more trailers?
Careful. This is the one place where causality plausibly runs the wrong way.
A game that has not shipped may genuinely not have footage worth cutting. A trailer is not a page asset you can produce on demand the way you can write a better short description: it needs a build, capture, an editor and usually money. The correct reading of “1.20 against 2.03” is that trailer count partly measures how far along a project is, so treating it as a lever means treating the symptom.
The within-group evidence is consistent with that caution. Inside the Coming Soon group, pages with two or more trailers have a median overall score of 69 against 65 for pages with one or none, but that gap comes back at p = 0.116 and is not something we would act on. Studios that have cut a second trailer have usually also finished their capsule and their copy. Our trailer best practices post covers what the first one has to do; a second one only earns its place when it shows something the first did not.
Screenshots are the cheaper half of this finding. Getting from 5 or 6 up to 9 is a capture session, not a production. If you are building the page before the game is finished, that is the part of the media gap you can actually close, and the Coming Soon page guide covers the rest of the pre-launch asset list.
Why the tags gap is about fame, not release status
Tags was the only score column where the naive split cleared significance on the median test: 62 against 70 at p = 0.0001. It vanishes on the matched comparison (62 against 62). A skeptic should immediately ask where it went, so here is the decomposition.
Compare the matched control against the rest of the released group, and the tags gap reappears inside released pages: 62 against 75, p = 0.0066. The 88 pages in that remainder include all 32 Metacritic-carrying games and 85 pre-2025 releases. The tags gap is a fame and age effect that happens to sit inside the released group, not a release-status effect. A plausible mechanism, worth one sentence and no more: Steam tags are partly applied by users after launch, so an old popular game accumulates a well-curated tag cloud that a brand new page cannot have yet.
One text-level signal does survive matching, weakly. Tag-category issues about chasing competitive or high-traffic terms appear on 14.9% of Coming Soon pages against 5.2% of the control, p = 0.043. We swept several issue-text patterns to find it, and it does not survive correction for that sweep, so treat it as a hypothesis rather than a result: pre-launch teams may be reaching for the big traffic tags because they have no data yet about which niche tags fit. The tag optimizer and our best Steam tags for 2026 post are the practical version of that argument.
What this analysis cannot tell you
The instrument has known defects. Two bugs in the scoring code were found after this corpus was scored and are fixed in the code but not in the stored data. Description subscores are depressed by roughly 11 points and tag subscores inflated by roughly 11 to 12 points on every page. That is harmless for group comparisons within one column, which is all we do here, and fatal for comparing one column against another, which we do not.
The tags column is the one exception, and it weakens the tags null. The tag inflation came from Steam feature flags being counted as tags, and that contamination is not balanced across our two groups: 37.5% of Coming Soon pages carry a tag issue naming a Steam feature against 24.7% of the control (p = 0.048). So tags is the one row of the matched table where the two groups were not distorted identically, and the tie there deserves less confidence than the other four.
The matching is coarse. Era plus absence of press coverage, nothing else. Better matching could reveal a difference this design cannot see.
These are nulls, not proofs of equivalence. The intervals rule out a pre-launch penalty larger than about 4 points on the overall score. They do not rule out 2 points, and they say nothing about pages outside this sample.
And nothing here connects to money. No wishlist, sales or traffic data is joined to these pages. A page that scores the same as a released page is not thereby a page that converts.
The practical version of all this is short. If you are building a pre-launch page, stop assuming it gets a lower standard because the game is not out. On our rubric it is already held to the same one, and it already fails in the same places. The gap you can genuinely close before launch is the screenshot shelf: get from 5 or 6 to 9, and you have bought the only difference this data can actually see. The leaderboard shows how other public pages scored, and the store page checklist covers the assets Steam requires before you get to the copy. The nulls we could not shake out of the rest of the corpus are collected in six things that do not predict page quality.