Steam Store Page Audit 2026: 373 Pages, 2,400 Issues

Between February 10 and July 26, 2026, our analyzer scored 373 unique public Steam store pages and logged 2,400 individual issues against them. The opening line is the most common failure on a Steam store page: 298 of the 373 pages (79.9%) were flagged for a weak or buried opening hook, and 279 of the 373 (74.8%) are still flagged after we discard every issue touched by the two rubric defects described below.

Three caveats before any number. This is a convenience sample: pages somebody chose to run through a free tool, not a random draw from Steam’s catalog. The scores are our own AI rubric, not a Valve metric. And we found two defects in that rubric on 2026-07-26, after every page here was scored, so the next section lists the comparisons this data cannot make.

Every page was submitted to the free Steam Page Analyzer by a real person, deduplicated to one latest analysis per app ID. The 373 pages come from 356 developers, so this is not one studio’s back catalog. 208 of them (55.8%) were unreleased “Coming Soon” pages at analysis time, which skews the sample toward pre-launch work and matters for several results below.

The set mixes famous and unknown. Hades scored 87, the highest here; Hollow Knight: Silksong 85, Palworld 83, Stardew Valley 80, Balatro 77. None of those studios participated in or endorsed this analysis; their pages are public and somebody pasted the URL. We have pulled several apart element by element, including what the lowest-scoring pages get wrong.

Two, both found on 2026-07-26 while checking analyzer output against live Steam pages, both fixed the same day, neither applied retroactively. We chose not to pay to re-score all 373 pages, so everything below came out of the broken version. The full writeup has the detail; here is the cost to this study.

The description floor read the wrong markup. Eleven of the 30 deterministic points our rubric awards a description came from regexes hunting for Steam BBCode: [h1], [list], [img]. Steam’s API returns compiled HTML, so those regexes matched 0 of 555 stored analyses. Of the 373 pages here, 92.0% had HTML headers, lists or bold and 84.7% had an inline image or video, yet 314 (84.2%) sat at exactly 19 of a possible 30, and the average page lost 9.8 of the 11 it could never reach. The prompt also reported formatting: no, embedded media: no on pages full of headers and images.

The tag floor counted things that are not tags. User tags, Steam genres and Steam features were merged into one list before the count that sets the tag floor, padding a typical page by 11 entries and lifting its tag subscore by 11 to 12 points.

So description subscores here are depressed, tag subscores are inflated, and nothing in this post compares one category’s score against another’s. An earlier draft led on description being the weakest of the four sections. That rested entirely on the suppressed subscore: on the corrected floor the description median moves from 64 to about 75 while tags moves the other way, which probably inverts it. We are saying so rather than quietly dropping it.

What still holds is any comparison of the same category between two groups of pages, because every page was distorted identically. That is the shape of nearly every finding below.

A third limit is not a defect but bounds one result: the model is sent only a page’s first 10 screenshots, and 121 pages (32.4%) have more.

The median page scored 67 out of 100, mean 66.4, range 29 to 87. The distribution is tight: half of all pages land between 61 and 74, and only 16 (4.3%) cleared 80.

Read that table across a row, never down a column. The description row is depressed by up to 11 points and the tags row inflated by about the same, so the ranking of the four categories is an artifact of our code. The spread inside a row is not.

The average page carried 6.4 issues (median 6, range 2 to 12), split 1,215 warning, 902 info, 283 critical. Issue count correlates -0.537 with the overall score, close to tautological but a check that the rubric is internally consistent.

Subscores cannot answer that here, so we answered it from the issue text. Every issue was sorted into a defect theme by hand-checked rules, and each page counts once per theme. The chart is the conservative count: every tag issue naming a string that can only be a Steam feature, and every description issue mentioning formatting or media, is thrown out first, because those are the issues the defects could have manufactured.

Before that filter the hook theme sits on 298 pages (79.9%) and the two tag themes on 233 (62.5%) and 204 (54.7%). The filter costs the hook theme 19 pages. It costs the wall-of-text description theme most of its mass: 89 pages before, 26 after. That collapse is the fingerprint of the markup defect, because the model was told the description carried no formatting and duly complained about formatting.

Three quarters of the hook issues name the short description or the opening line explicitly; the rest fault the opening of the long field. Either way the top two themes share a surface: the 300 characters Steam prints under the capsule in search, wishlists and the Discovery Queue, and the image above them. Both fail before a shopper clicks, which lines up with our tally of top-seller capsule art.

If you only ever fix one thing, start with the 300 characters at the top of your page. Our short description guide has the formula; the description writing guide covers the long field below it.

Deciles give 37 pages each side. The top decile scored 77 to 87, the bottom 29 to 52.

Every row is one category compared between two groups, the comparison the defects leave standing: both distortions hit every page in both deciles. Read the gaps as approximate rather than ranked, but screenshots separating the deciles least is a wide enough margin to mention.

The behavioral differences are more useful than the score differences. Bottom-decile pages carried a median of 8 issues against the top decile’s 5, and averaged 1.81 critical issues each against 0.03: 36 of the 37 best pages have zero criticals, and only 2 of the 37 worst do. Top-decile pages also wrote more, a median detailed description of 4,887 characters against 3,074, and 10 screenshots against 7.

No, past a point. Screenshot count correlates only 0.251 with the screenshot subscore, and the relationship is a step, not a slope.

Pooled: pages with 7 or fewer screenshots (n=135) have a median screenshot subscore of 68; pages with 8 to 12 (n=160) have 78. A permutation test on that 10-point difference gives p = 0.010. That step is the finding.

The flat line to the right of it is not. The rubric sees a page’s first 10 screenshots and nothing else, so on the 121 pages (32.4%) carrying more, the extra images were never looked at. Comparing 8 to 12 against 13 or more (n=78) does return a median difference of exactly 0 at p = 1.00, but both groups were capped at the same 10 visible images, and 105 pages share a subscore of exactly 78, a mass point that manufactures p values of 1.00. On means the difference is -0.93 points, 95% CI -3.63 to +1.82, p = 0.498. Going from 5 screenshots to 9 moved our score. Past the tenth we are blind.

The 52 pages sitting at exactly Steam’s 5-screenshot minimum had a median overall score of 62, five points under the set median.

One confound: unreleased pages have fewer screenshots (correlation -0.237 with the Coming Soon flag) and score slightly lower. Controlling for release state, the count-to-score correlation drops only from 0.251 to 0.221, so the step survives. Run your set through the screenshot checker and read the screenshot guide before adding a tenth image that says nothing new.

We expected yes. The data says we cannot show it.

21 pages (5.6%) had no trailer at all. Their median overall score was 64 against 68 for everyone else, and a permutation test on that 4-point gap returns p = 0.175, not distinguishable from noise at this sample size. Worse for the hypothesis: 20 of those 21 are unreleased Coming Soon pages, and comparing them only against other unreleased pages shrinks the gap to 2 points at p = 0.493.

Trailer count does track with score, but read it as a maturity signal rather than a lever:

Correlation between trailer count and overall score is 0.269. Studios that have cut four trailers have usually finished their capsule, screenshots and copy too. Cutting three more for a page with none of that fixed is not what this table says. Our trailer best practices covers the one that matters.

Median in the sample is 232 characters, mean 223, range 49 to 319. Two thirds of pages (249 of 373) sit between 200 and 300.

Every description subscore in that table is depressed by the markup defect, so read the differences and ignore the levels. The honest reading is a floor, not a target. Pages under 150 characters (n=49) have a median description subscore of 60 against 65 for the 250-300 band (difference 5, p = 0.016). Comparing 250-300 against 150-249 gives 1 point at p = 0.729. Getting off the floor helps. Creeping from 210 characters to 290 does not.

The cleanest evidence that length is the wrong variable: median short description length is 232 characters for pages flagged with a weak-hook issue and 232 for pages that were not. Identical. What the rubric penalizes is the first ten words, which matches Chris Zukowski’s framework for describing your game to strangers: shoppers want the verbs, not the lore.

Nine pages sit at 300 characters or more, topping out at Dota 2’s 319. The Steamworks documentation only says the field is “limited to a few hundred characters”, so 300 is a practical ceiling, not a published rule.

Length matters more in the long field. The 25 pages under 1,500 characters have a median description subscore of 45; the 33 pages at 8,000 or more have 71.

Barely. Genre medians span 59.5 to 69, but the small genres carry error bars wide enough to swallow the difference.

RPG (71 pages, median 69), Strategy (101, 67), Simulation (110, 67), Early Access (47, 66) and Racing (12, 65) sit inside that band. Intervals are bootstrapped over 2,000 resamples. The MMO median of 59.5 rests on 10 pages and its interval runs from 48 to 75, overlapping every other genre. Action versus MMO returns p = 0.030, the only genre pair clearing the usual threshold, on 10 observations. Action versus Casual: 3 points, p = 0.403. Action versus Indie: 3 points, p = 0.230.

The variation within Action (standard deviation 10.3) is larger than the variation across the whole sample (9.6). Genre is not a useful predictor of store page quality here. Anyone quoting a genre median from this dataset as a benchmark is over-reading it.

Critical is the rubric’s top severity, reserved for problems that plausibly cost a page conversions on their own. There are 283 across 2,400 issues (11.8%), spread thin rather than pooled in a few disasters. 210 of 373 pages (56.3%) carry at least one, and only 6 pages carry three or more.

Description holds 51.2%, the one count the defects dent without destroying. The prompt asserted that the description had no formatting on nearly every page, which pushes issue generation toward description. Run it conservatively: drop the 17 critical description issues mentioning formatting or embedded media and the 22 critical tag issues naming a Steam feature, and 244 remain, split description 128, capsule 50, tags 33, screenshots 33. Description still carries more than the other three combined, 128 against 116. Read 145 as the upper bound, 128 as the lower.

The most common critical issue by title is “Weak opening hook fails to grab attention” on 31 pages. Second is “Capsule lacks visual impact and genre clarity” on 15. The only page with 5 criticals, an unnamed indie puzzle game, was flagged for a mascot-logo capsule with no gameplay, a buried hook, an unformatted description and 9 of 20 tag slots used. We checked it by hand: the description really does carry no headers, lists or images, so that flag stands. The tag count does not. 5 of those 9 entries were genres and features, putting the real figure nearer 4.

Pages with at least one critical issue had a median overall score of 64. Pages with none had 74.

Sample. 373 unique Steam app IDs, one latest analysis each, run between 2026-02-10 and 2026-07-26. Pages entered by a user pasting a URL into a free tool, so the selection effects are obvious: people run the tool on pages they suspect need work, and on famous pages out of curiosity. It does not represent Steam’s catalog.

Rubric. Four subscores of 0-100 plus an overall that is effectively their average. 2,400 issues, and 3,129 recommendation rows of which 851 are empty placeholders left by our own report-gating, so every figure here comes from the issues. 1,578 issues carry a score-gain estimate, median +8, a predicted movement in our score and nothing about revenue.

Defects. Both bugs were fixed in code on 2026-07-26 and neither fix reached these 373 analyses. Scores produced after that date are not comparable to the ones here, and the leaderboard holds both regimes. The defect writeup has the code, the tests and the counts.

Statistics. Correlations are Pearson. Group comparisons are two-sided permutation tests on the difference of medians, 20,000 resamples, fixed seed. Genre confidence intervals are bootstrapped medians over 2,000 resamples. Deciles are 37 pages each.

What we cannot claim. No causal link between any score and any commercial outcome, because no sales, wishlist or traffic data is joined here. No comparison of one subscore category against another. And every finding is conditional on our rubric being a reasonable judge of store pages, which is a design opinion, not a fact.

To check your own page, the analyzer above scores the same four sections with both defects fixed. The capsule validator tests the thumbnail legibility problem that hit 265 pages here. If it tells you your opening line is weak, you are in company: 298 of these 373 pages were told the same thing, and it is the cheapest sentence on the page to rewrite.