Are AI-Flagged Steam Pages Lazier? 1,592 Pages Compared

On the ten things I counted, no. Across 1,592 unreleased pages from the top 62% of Steam’s popular Coming Soon listing, the 29% that carry an AI disclosure are not behind the rest on any of them. Tags and demos are level within the noise. On the others the flagged pages have a little more: a median of 4 languages against 2, “About this game” text about 35% longer, and a gap of 10 points in how many have the same developer and publisher. Short descriptions and screenshots differ by 6% and one image.

Six of the ten measures cleared the usual 5% bar in both of two separately drawn samples ( 332 games and 1,260 ). Some of the six are one fact seen twice, since languages and single-language pages move together, and so do the two text lengths. A repeat across samples rules out a lucky draw. It does not rule out the way the pages were chosen, which is covered under the limits.

“Lazier” here means “fewer of the things you can count”. Counting cannot see whether a capsule is readable or a hook is any good.

I started from Steam’s “popular coming soon” search listing ( about 65,000 matches on 23 October 2026 ) and drew random pages from its first 40,000 positions, spread across that range so popular and obscure pages both appear.

I drew a smaller sample, then a larger one that excluded every page in the first, and kept the two apart as a check on each other. About one id in ten was dropped as DLC, software or a page that failed to load ( I did not separate those ), leaving 332 and 1,260 games, 111 and 354 of them with an AI disclosure.

For each page I recorded what Steam’s own product data publishes ( screenshots, trailers, short and detailed description length, languages, developer and publisher ), plus the tag count and the “AI Generated Content Disclosure” section from the English store page. Genres, developer names and release dates were recorded in the larger sample only. All requests were read-only and spaced 1.5 seconds apart. I am not publishing the page-level list, because it would put small developers’ names next to a label some players dislike. These are aggregates only.

Limits:

• •The listing is ordered by popularity and I sampled only its top 40,000 of about 65,000. The least popular 38% is not covered. I assume thin pages are likelier to sit there, but nothing in these data shows it. If a disclosure lowers a page’s rank, the flagged pages that remain in the top 40,000 may be the stronger ones, which could produce “as full as the rest” on its own.
• •“Unflagged” means no disclosure, not no AI, and developers who used AI may differ from those who did not in ways other than the page, such as team size or experience.
• •These pages are unreleased, so nothing here is about sales. Valve changed the disclosure form in January 2026, so who discloses may shift, and a rerun would show it.
• •A flag is yes or no. It does not say how much AI went into the game.
• •Ten measures, two samples. By luck alone, any one measure would clear 5% in both samples about 1 time in 400, and at least one of ten about 1 time in 40. Related measures count as one, so six hits are fewer than six confirmations.

The flagged and unflagged columns pool both samples. The last two columns show each sample on its own: flagged against unflagged, then the p-value. A p-value is the chance of seeing a gap this large or larger if the two groups were really the same, so a small one means chance alone would rarely produce it. It says nothing about whether the gap matters, and a large one does not show the groups are equal, only that these samples cannot tell them apart.

I used a rank test ( Mann-Whitney ) for counts and lengths and a two-proportion test for shares.

Each of the six rows that clear 5% in both samples shows the gap in the same direction both times: About text, short description, screenshots, languages, single-language pages and developer-is-publisher. Trailers lean the same way and clear 5% only in the larger sample. Tags and demos do not move.

The biggest gaps are in text and languages. Flagged pages write about 473 more characters of About text, and I cannot say why. Some of it may be text drafted with a model, and some may be developers who write a lot and also use AI. The data does not separate those.

The language gap has an obvious suspect, since 27% of the disclosures mention localization or translation ( see what the disclosures say ). A rough check supports it. Among the 341 flagged pages whose disclosure does not mention either, the median is 2 languages, the same as the unflagged 2, and 36% list a single language against 43% of unflagged pages. The gap sits in the 124 that do mention it, where the median is 9. That is a keyword match, so treat it as suggestive.

Developer-is-publisher is a structural fact about the studio and not something on the page, so it is a confounder as much as a finding.

In the larger sample, the only one that recorded release dates, 28% of flagged pages give an exact release day against 22% of unflagged ones ( p = 0.01 ). It is one sample, so it is not among the six.

To get one number per page, I counted how many of five items a page has: at least 5 screenshots, at least 1 trailer, a short description of 150 characters or more, at least 15 of the 20 tag slots used, and at least 500 characters of “About this game” text.

I chose round, low cutoffs after the first sample’s results were in, so this section is exploratory. They are not Valve’s rules, and the check is not a quality score. Five screenshots is in fact Steam’s own minimum, which is why almost every page passes that item.

The last column is the pooled p-value. 62% of pages have all five items. Short description, trailer and About text each lean toward flagged pages, and I did not separate them. Screenshots and tags contribute nothing.

The gap is about 8 points in both samples. In the larger one that is p = 0.008, in the smaller one p = 0.14, so only the larger clears 5%. Pooled, flagged pages are ahead by 8 points ( 95% interval 3 to 13 ). The fair reading is no sign of laziness and perhaps a little more completeness.

Do not use this as a checklist. The cutoffs were chosen after seeing data, and a short, text-light page can be a deliberate choice.

Among the seven biggest genres, Simulation is highest ( 37% ) and Action and Adventure are lowest ( 23% and 24% ), against 28% overall in the larger sample. Strategy, Casual, RPG and Indie sit at 30 to 32% and are not clearly different from each other.

With 258 to 928 pages per genre, each figure is good to roughly plus or minus 3 to 6 points. Smaller labels swing further ( Early Access is 39%, on 139 pages ), so I stayed with the seven biggest, and I would treat the gap between Simulation and Action as a lead and not a finding. A page can carry several genres, so the groups overlap.

Because the samples are random pages from the listing, they double as a benchmark. These are the quartiles across all 1,592 games, flagged or not. They describe the top 62% of the popularity-ranked listing. If thin pages sit in the unsampled tail, the true middle of all Coming Soon pages is lower.

Use the tables to catch free fixes, such as a missing trailer or unused tag slots. Do not chase the median on languages or text length, and do not read a low number there as a deficiency: a 900-character About section that sells the game beats 2,200 characters of padding. The short description is capped at 300 characters in the Steamworks editor ( see our short description guide for where that number comes from; a few pages measure slightly over, probably because of markup ), and the median page uses about four fifths of it.

A page can have the median number of everything and still lose the click, because the capsule is unreadable at thumbnail size or the first sentence of the description says nothing.

The flag has the same limit. A disclosure tells you the developer used AI for something, not what, and our guide to what developers write in a disclosure looks at the text. To compare your own page, the conversion benchmarks cover what happens after the click, the screenshot guide covers the images, and the optimization guide ranks the fixes we see most often.

Nothing here separates the two groups on page content, and if anything the flagged pages are slightly fuller. The free analyzer below scores your page. It gives a quality score, which nothing in this post tested.