Six Things That Do Not Predict Steam Store Page Quality

We keep a list of things we expect to predict store page quality, and we test them against the 373 Steam pages our analyzer has scored. Most of the list does not survive. The single most common finding our rubric produces, a note that the small capsule may be hard to read at thumbnail size, appears on 265 of 373 pages (71.0%), and the pages that get it score higher than the pages that do not. Metacritic score, having a publisher, announcing a firm release date, shipping on three platforms and genre all failed to move the needle too.

Two caveats before any number. This is a convenience sample: pages somebody chose to paste into a free tool, not a random draw from Steam. And the scores are our own AI rubric, a design opinion about what a good store page looks like, not a Valve metric. A null here means “this variable does not predict our score on this sample”, which is a narrower claim than it sounds.

Every page was analyzed once between 2026-02-10 and 2026-07-26, deduplicated to the latest run per app ID, across 356 distinct developers. 208 of the 373 pages (55.8%) were unreleased Coming Soon pages, which confounds nearly everything, so every result below that looked significant was re-run inside released-only and unreleased-only subsets.

Two defects in the scoring code matter before you read subscores. The description floor was computed from BBCode patterns that never match Steam’s HTML, so description subscores here are depressed by roughly 11 points on every page. The tag count was padded with Steam feature flags such as Single-player, so tag subscores are inflated by roughly 11 to 12 points on every page. Both are fixed in the code and neither is fixed in the stored data. The consequence: never compare one category against another in this corpus, but comparisons between groups of pages on the same category stay valid, because every page was distorted the same way. Every test below is of that second shape.

Group comparisons use two-sided permutation tests, 20,000 resamples, fixed seed. Because several score columns pile up on a single value (capsule is exactly 72 on 152 of 373 pages), a median test can return p = 1.0000 that means nothing at all. Where that happens we report the bootstrap confidence interval on the mean difference instead, which is what tells you how large an effect the data still allows.

The recurring flag is some phrasing of “small capsule readability at thumbnail scale”. Grouping title variants gives 266 issues across 265 pages, phrased 110 different ways, 259 of them opening with the words “Small capsule”. It is the most frequent defect cluster we log, and it is mostly low severity: 181 warning, 76 info, 9 critical.

Pages carrying it have a median overall score of 69 against 65 for pages that do not. The plain median test on that gap gives p = 0.0543. The mean test is cleaner: 67.30 against 64.05, a difference of +3.26 points, p = 0.0025, bootstrap 95% CI [+0.87, +5.69]. Holding the page’s total issue count constant, which controls for the mechanical “more issues means lower score” effect, the advantage grows to +5.89 points using issue-count quartiles and +5.00 points using exact issue counts, both p < 0.0001, and both survive Bonferroni correction across the 24 defect clusters we tested (threshold 0.00208).

The mechanism shows up when you split capsule verdicts three ways.

Pages flagged only for thumbnail legibility and pages the analyzer had no capsule complaint about at all carry the same median capsule subscore of 72 (difference 0.0, p = 1.0000). The gap to the genre-signal cluster is 14 points at p < 0.0001. Our reading is that thumbnail legibility is what the rubric reaches for when it cannot find anything worse to say about a capsule. The rubric agrees with itself here: the estimated score gain it attaches to this flag averages 6.85 points against a corpus average of 7.70 (p = 0.0007), one of the lower-value fixes it recommends.

Three things stop this from being a clean bill of health, and we would rather state them than have you find them.

First, the medians tie because both groups sit on the 72 mass point. On means, thumbnail-only pages score 71.36 against 75.45 for pages with no capsule complaint, a real gap of 4.10 points, p = 0.0002. So the flag is close to silence, not identical to it.

Second, the score advantage is not evenly distributed. Within released pages the mean gap is +1.61 (p = 0.333); within unreleased pages it is +5.47 (p = 0.0002). The effect lives in the pre-launch half of the sample.

Third, the decile comparison is confounded. The top decile is 70.3% flagged against 48.6% in the bottom decile, but the top decile is also 35.1% Coming Soon against 59.5% in the bottom, so release status is moving with the flag.

What this is not. This is a fact about how our rubric distributes its attention, not a fact about Steam. Small capsules genuinely are rendered small, and legible key art at thumbnail size remains sound advice that we still give in our capsule design guide. What the data shows is that receiving this particular note from this particular analyzer is not a signal that your page is in trouble. The capsule finding that does predict trouble is the genre-signal cluster, pulled apart in the capsule genre signal post.

Thirty-two pages in the set carry a Metacritic score, all of them released. Correlating that score against our overall page score gives Spearman rho = +0.324, permutation p = 0.0711, bootstrap 95% CI [-0.063, +0.652].

That interval is the whole story. It contains zero, and it contains a strong positive relationship. We cannot confirm that critical acclaim predicts page quality and we cannot rule it out. With 32 pages, a simulated power curve says we would detect a true rho of 0.3 only 34% of the time, a true rho of 0.4 58% of the time, and a true rho of 0.5 81% of the time. This is an underpowered test, not a proven null, and anyone reporting it as “review scores do not matter” is over-reading it.

What the individual pages do show is that a great game is not a guarantee of a great page.

Half-Life 2 holds the highest Metacritic score in the sample and a page score of 64, three points below the corpus median of 67. Terra Invicta at Metacritic 83 scores 60. Hades at 93 scores 87. The spread among well-reviewed games is wider than the gap between well-reviewed and unreviewed ones. None of those studios participated in or endorsed this; their pages are public and somebody pasted a URL.

There is a seductive number here that we want to dismantle rather than publish. Pages with a Metacritic score have a median overall of 75.5 against 67.0 for released pages without one, p < 0.0001, which looks like decisive proof that good games have good pages. It is not. Having a Metacritic score means the press covered you, which correlates with budget and studio size, not with “the game is good”. More to the point, it is a selection effect in our own sample: famous games are the ones people paste into a free analyzer out of curiosity. The comparison is between games a stranger looked up and games their own developer submitted, and it tells you about our traffic, not about Steam.

One subscore does track, and honesty requires reporting it: Metacritic against the screenshot subscore is rho = +0.535, p = 0.0020, which survives Bonferroni across the five score columns. Capsule sits at +0.338 (p = 0.0585), description at +0.189, tags at +0.126. Budget buys screenshots. It does not appear to buy copy.

Treating a page as self-published when the developer and publisher strings match gives 277 self-published and 96 externally published pages. The naive comparison is median 66 against 69.5, p = 0.0669, close enough to tempt a headline.

It dissolves on the first control. Self-published pages are 62.1% Coming Soon against 37.5% for externally published ones (p = 0.0001). Publishers sign games that ship. Restricting to the 165 released pages, the gap becomes median 69 against 70, p = 0.7589, with a mean difference of -2.17 points and a bootstrap 95% CI of [-5.02, +0.79]. This null is reasonably well powered: a simulated 3-point true gap would be detected 90% of the time by the mean test at this sample size, so a publisher advantage larger than about 5 points is ruled out.

What partly survives is volume, not quality. Among released pages, externally published ones carry a longer detailed description (median 5,133 characters against 3,551, p = 0.0019) and more screenshots (11 against 9, p = 0.0163). Add a price control and even that thins out. Within the 124 released paid pages listing at $5 or more, the score gap is exactly zero (70 against 70, p = 1.0000), the screenshot gap is gone (10 against 11, p = 0.7156), and only the description-length gap holds on the median (4,191 against 5,541, p = 0.0278) while the mean test on the same comparison returns p = 0.2656.

Read together: a publisher buys you volume, and most of that volume is really a price effect, which we covered in price versus page quality. Nothing here says a publisher writes you a better page. If you are weighing the decision on other grounds, our publisher versus self-publishing guide is the honest version.

Inside the 208 unreleased pages, 27 name an exact day, 31 name a quarter, 45 name a month or a bare year, and 105 say something like “Coming soon” or “To be announced”. Comparing the 58 pages with a genuinely specific date against the 150 loose ones:

Screenshot count is identical at a median of 8, p = 1.0000. Trailer count is identical at 1, p = 1.0000. Narrowing to the 27 pages naming an exact day against the other 181 gives 69 against 66, p = 0.3306, mean difference +0.14 with a CI of [-3.88, +3.86].

The capsule and tags rows are ties on a mass point and carry no information, which is why the mean-difference intervals are there. Power on the mean test is 0.74 for a 3-point gap and 0.91 for a 4-point gap, so a moderate effect would have shown. A studio confident enough to commit to a date is not, on this evidence, a studio that has finished its page. Committing to a date is still worth doing for wishlist timing and for Steam’s upcoming-release surfaces, which is a separate argument covered in the Coming Soon page guide.

258 pages list Windows only, 73 list two platforms, 42 list three. Median overall scores run 67, 67 and 70, and a permutation omnibus across the three groups returns p = 0.5477.

Pooling into Windows-only against multi-platform gives a median of 67 against 69 (p = 0.5021) but a mean difference of -2.36 with a CI of [-4.38, -0.32] and p = 0.0287, which is the kind of marginal result that should be checked before it is believed. It does not hold up. Within unreleased pages the mean gap falls to -0.64 (p = 0.6583). Within the released paid pages listing at $5 or more it falls to -1.81 (p = 0.2499). Windows-only pages also list at a median of $11 against $15 for multi-platform ones, so price is doing part of the work.

We are reporting this one as not reliable rather than as a clean zero. There is a small, possibly real association in the pooled data, and it disappears under every control we can apply, which usually means it was budget wearing a Mac build as a disguise.

Our store page audit already found that genre does not predict the overall score. Nobody had checked the four subscores underneath it. We tested nine genre labels against each of the five score columns, with a Bonferroni threshold of 0.05/9 = 0.00556 inside each family.

No genre survives correction on any subscore. The smallest p-value in each family: capsule 0.0535 (Free To Play), description 0.1474 (Strategy), screenshots 0.0189 (Free To Play), tags 0.0646 (RPG), overall 0.0809 (Free To Play). Free To Play appears three times and rests on 16 pages, which is how a genre with a wide interval keeps almost clearing a bar.

One detail is worth knowing about even though it is not a finding. The median capsule subscore is exactly 72 for eight of the nine genres, with Free To Play the only exception at 68. That is not a discovery about capsule art converging across genres. It is a rubric mass point: capsule lands on exactly 72 on 152 of the 373 pages (40.8%), so any subgroup of reasonable size will have 72 as its median. Tags does the same thing at 62 on 87 pages. If you see a table of medians from this corpus where a column is suspiciously flat, that is why.

No, and finding that out was more useful than the answer would have been.

We wanted to test charm pricing: do pages priced at $X.99 score differently from pages at a round number? Restricting to the 154 released pages carrying a US list price, 150 end in .99 and 2 end in .49. That is 152 of 154, or 98.7%. The only two exceptions in the entire corpus are one game at $14.89 and one at $35.00. There is no control group. The question is untestable on this data, and it is untestable for a reason worth knowing: charm pricing on Steam is close to universal.

There is a trap here that we walked into first. Run the same test on the discounted price at the moment of analysis and you get n = 123 against 31, medians 69 against 68, p = 0.8506, which looks like a publishable null. It is nothing of the sort. Those 31 “round-priced” pages are pages that happened to be on sale, so that test compares discounted games against undiscounted ones. In Steam’s price_overview field, final_formatted is the sale price and initial is the list price. We made that mistake so you can skip it.

Sample. 373 unique Steam app IDs, one latest analysis each, scored between 2026-02-10 and 2026-07-26, from 356 distinct developers, 208 of them (55.8%) unreleased. Pages entered the sample when a person pasted a URL into a free tool, so people run it on pages they suspect need work and on famous pages out of curiosity. It is not representative of Steam’s catalog.

Statistics. Two-sided permutation tests on the difference in medians and in means, 20,000 resamples, fixed seed. Bootstrap percentile intervals, 20,000 resamples. Bonferroni correction inside each family, stated at each use. Power figures resample the observed distributions with an artificial shift applied, 1,200 iterations each. One note we owe the reader: the permutation test on medians is unstable on these lumpy scores and its power depends on which group you shift, which is why the mean difference and its interval are the primary evidence in every null above.

Instrument limits. Description subscores are depressed and tag subscores inflated corpus-wide, as described at the top. The analyzer sends only the first 10 screenshots to the model, and 121 of 373 pages (32.4%) have more than 10, so nothing here says anything about screenshots 11 and up. The analyzer’s evidence text quotes the small capsule as 231x87, which is a render size; the current Steamworks upload size is 462x174.

What a null does not mean. Six of these results say a variable does not predict our score on this sample. None of them says the variable does not matter to a player, to Steam’s algorithm, or to sales, because no sales, wishlist or traffic data is joined to any of these pages. The genre-signal and price results we did find are in the capsule genre signal and price versus page quality posts, and the pre-launch cut is in Coming Soon pages are not worse pages.

If your own report comes back with a small-capsule legibility note and nothing more severe, the data here says you are in the company of 219 other pages whose capsules the rubric had nothing real to complain about. Check the leaderboard for how public pages scored, and if you want a second opinion on your art rather than a score, the capsule validator is the tool for it.