Steam Description Anatomy: What 373 Real Pages Contain

Steam’s own API reports the length of your About This Game section as a character count, and that number is misleading. The median Steam description field is 4,476 raw characters, but only 38.4% of it is readable text: the median page carries 1,655 visible characters, which is 268 words, or roughly 67 seconds of reading. The rest is HTML markup.

Every number in this post is measured directly from the detailed_description field Steam returns for each app, not from our scoring of it. That makes these figures independent of our rubric and of the two rubric bugs we disclosed elsewhere. They describe what real store pages contain. They do not claim what a page should contain, and the section on formatting explains why we deliberately stop short of that.

The sample is 373 unique public Steam pages that people submitted to our free Steam Page Analyzer between February 10 and July 26, 2026. It is a convenience sample, not a random draw from Steam’s catalog, and 55.8% of the pages were unreleased at the time.

Three different numbers get called “description length”, and they are far apart.

The 4,476 figure is what you get if you measure the field Steam hands you. It counts every <img src="https://shared.akamai.steamstatic.com/...">, every , every closing tag. On the median page, 38.4% of that field is text a human can read (bootstrap 95% CI 35.4% to 40.3%; the mean page is 42.7%, CI 40.7% to 44.7%).

So the honest length benchmark is 268 words. At a typical reading speed of 240 words per minute that is about 67 seconds, and almost nobody reads all of it. Half of all pages sit between 188 and 358 words. The longest description in the set runs 1,272 words; the shortest readable one is 10.

If you want a single number to compare your own page against, use visible words. If your CMS or spreadsheet only gives you the raw field length, the median there is 4,476 characters, and expect roughly a third of it to be markup.

Steam compiles the BBCode you type in Steamworks into HTML. Counting those elements gives a structural profile of the median page.

Three details worth knowing before you benchmark against that table.

Inline media is mostly video, not images. The median page embeds 3 videos (72.7% have at least one) against a median of 0 inline images (45.3% have at least one). Those <video> elements are the autoplaying, muted, looping clips Steam serves as paired webm and mp4 sources, which is what an uploaded animated GIF becomes. If you have been treating “add GIFs to your description” as an advanced move, 72.7% of the pages in this set already do it.

Steam never emits an h3. Across all 373 pages the h3 count is zero, and so is the horizontal-rule count. Practically, your header hierarchy on Steam is two levels deep at most, and [h1] is rare: only 11.8% of pages use one, against 57.4% using [h2].

Plain prose is nearly extinct, but partial formatting is normal. Only 13 pages (3.5%) have no headers, no bullets and no inline media at all. But 37.5% have no headers, 48.0% have no bullets, and 15.3% have no inline media. The median page picks two of the three levers, not all three.

The description does not sit alone, and the fields around it are more consistent than the copy is.

Two of those deserve a second look. 112 pages (30.0%) support exactly one language, which is almost always English, while the widest page in the set lists 103 languages. And 121 pages (32.4%) carry more than 10 screenshots, which matters because our own analyzer only reads the first ten, so nothing we publish says anything about screenshot 11 onward.

The short description median of 232 characters means the typical page leaves about 68 characters of its 300-character allowance unused. Our short description guide covers what to do with them.

Medians hide how differently successful pages solve this. All three of these were re-checked against the live Steam API on the day this was written.

Dota 2 uses no structure at all. 1,511 raw characters, 1,402 of them visible, 241 words, zero headers, zero bullets, zero inline media, four bold runs. It is 92.8% readable text, the opposite of the corpus median, and it is one straight block of prose.

Hades uses headers and nothing else. 3,551 raw characters, 1,889 visible, 300 words across 6 headers, one inline image, two embedded videos. That is close to the corpus profile, just tidier.

The Witcher 3 has almost no copy in the field at all. 1,446 raw characters and 10 visible words, because the description is six inline banner images with the text baked into the artwork. Only 3.8% of the field is machine-readable text. This is a real and common Steam pattern, and it has a cost worth naming: search engines, screen readers and AI assistants read none of it. If you do this, keep a text layer underneath.

For contrast at the other end, Stardew Valley runs 623 words at 83.9% readable text with 13 bullets and no headers, while Palworld runs 377 words at 24.6% readable text with 10 embedded videos. Both are commercially enormous, and they built the same field in opposite ways.

This is the question a post like this is supposed to answer, and we are not going to, because our instrument was broken in exactly the place it would need to be sound.

Our scoring code awarded part of the description subscore by looking for BBCode markers such as [h1] and [list] in a field that Steam returns as compiled HTML. Those patterns matched on 0.0% of pages. Every one of these 373 analyses was scored with the model being told formatting: no, embedded media: no for a page that, in most cases, had both. The bug is fixed in the code as of July 26, 2026, but these stored analyses predate the fix and were not re-run.

We ran the tests anyway, and here is what came back: pages with headers score identically to pages without (median 64 both ways, p = 1.000), pages with bullets differ by 1 point (p = 0.53), and the raw gaps that did appear for inline media and bold text disappear once you hold word count constant (media +1.8, p = 0.30; bold +2.7, p = 0.07).

Do not read those as evidence that headers and bullets do not work. A null from a blind instrument is not a null about the world. Our description writing guide recommends headers, bullets and images on grounds of shopper scanning behavior, and nothing here contradicts it. The only defensible statement this dataset supports is that we cannot measure formatting effects with this data, and the responsible thing is to say so rather than dress a broken measurement up as a finding.

One thing the data does show cleanly, because it needs no rubric at all: length and structure travel together. Pages at or above the 268-word median carry a median of 3 headers, 4 bullets and 5 inline media items; pages below it carry 1, 0 and 3. Whatever you conclude about formatting, formatting is what longer descriptions do.

Treat them as a position check, not a target.

• —If your description is under 188 visible words, you are in the bottom quarter of this sample. That is a deliberate choice for some pages and an oversight for most.
• —If it is over 358 words, you are in the top quarter. Nothing here says that is wrong, but the scan-pattern argument in our store page mistakes post applies with more force the longer you go.
• —If you have no inline media, you are in the 15.3% minority, and video specifically is the more common form.
• —If you support one language, you are in the 30.0% that does, and the localization languages breakdown has the market-size math on what changing that is worth.
• —If your copy lives inside images, count how many visible words survive. Ten is what The Witcher 3 has. That is a defensible choice for a page with that much brand behind it and a risky one otherwise.

The store page checklist turns those checks into a trackable list, and the 373-page store page audit has the scoring distribution these pages sit inside.

Structure counts come from a tag parser run over the raw detailed_description string. Visible text is what survives stripping every tag and collapsing whitespace. Medians carry bootstrap 95% confidence intervals from 2,000 resamples with a fixed seed. The 373 rows were pulled fresh from stored raw API responses and verified field by field against our analysis export with zero mismatches.

Four limits:

• —Convenience sample. These are pages somebody chose to paste into a free tool, so the mix of famous and unknown games does not match Steam’s catalog, and 208 of 373 (55.8%) were unreleased.
• —Whitespace word counts under-count CJK text. Chinese, Japanese and Korean copy does not separate words with spaces. Exactly one page in 373 is CJK-heavy, and it shows 418 visible characters against 12 counted words, roughly 35 characters per “word” where the corpus norm is 6.2. Removing it leaves the median at 268 words, unchanged. If you write in a CJK language, use the visible-character figure of 1,655 instead.
• —detailed_description is not always identical to About This Game. On 44 of 373 pages (11.8%) Steam prepends a promotional block (a DLC banner, an edition comparison) to detailed_description that the separate about_the_game field does not carry, a median of roughly 400 extra characters and up to 3,370. The other 329 pages (88.2%) are character-for-character identical.
• —Nothing here connects to sales. No revenue, wishlist or unit data is joined to this dataset, and no claim in this post is causal.

The measurements are a snapshot. Steam pages change, and the three named examples above were verified live, but any of them could be rewritten tomorrow.