In late December 2025, one game developer toyed with a thought experiment: How much could you tell about a Steam game’s success based solely on one screenshot?
It’s easy to say not to judge a book by its cover, but the reality is that we’re visual creatures. The first thing we’ll notice about a game is what it looks like. Strong art direction can be the difference between watching a trailer or closing out of a browser window with total disinterest. Could a screenshot thus be a good predictor of whether a game might find an audience?
A game creator with the handle Newbie Indie Game Dev recently sat down to try and figure out how, exactly, one would go about answering such a question with some rigor. He downloaded one screenshot from every single game on Steam and fed the resulting information to a neural network that would then cluster games together based on what they looked like. The screenshots were automatically pulled specifically from whichever picture was displayed first in the image gallery on that game’s Steam page. Presumably, these pictures would be the best shorthand for whatever vibe a publisher wanted to sell for any given game.
From this data set, and in an effort to keep things computationally feasible, Newbie pared the selection down to 10,000 games. He included a screenshot from every game on Valve’s platform that had garnered at least 3,000 reviews, while also including a randomized selection of games. Beyond the initial screenshots, each of these games included a few other pieces of information, like game tags and prices. All of this information was then fed to EfficientNet-B3, a neural network which excels at sorting visual data sets. Thus was born something Newbie calls The Gaming Map, an interactive showcase that can cluster video games together based on metadata.
So, what did he find? Perhaps this won’t be a shocker, but the games with the most reviews tend to be “action-packed” 3D games. Screenshots for these popular titles tended to focus on things like guns, people in motion, and war. Curiously, many of these games veered toward warmer color tones. Not so curiously, explosions were a common theme among this cohort of games. Fittingly, many — but not all — of the games under this umbrella were first-person shooters, military games, and tactical games.
Does this mean that action games are more likely to do well on Steam? Newbie is hesitant to make any definitive declarations about what a popular video game might look like, as there are always exceptions. Some of the most visually distinctive games were clustered far away from other titles regardless of how well they performed with PC gamers. The findings surrounding action-packed screenshots may have less to do with having the “right” genre than it might with effective visual messaging. What better way is there to make someone interested in your game than giving them a sense of what they’ll be doing — and making it feel thrilling?
Speaking to Polygon over email, Newbie stressed the axiom that correlation does not mean causation. He leaves it to viewers to conclude what they may about screenshots and popular games. Mostly, he was surprised his data displayed any correlation whatsoever between visuals and commercial success.
“I didn’t have high expectations going in,” Newbie said. “I fully expected the data to be too noisy to say anything interesting. So seeing that structure emerge was exciting.”
In the video, Newbie reminds viewers that action-heavy 3D games are typically the domain of AAA publishers who have the budget to make titles with complex visuals and promote them with extensive marketing campaigns. Unsurprisingly, though, Newbie said that a single screenshot can be a good predictor of how much a consumer might expect to pay for the title in question. Better graphics, higher cost.
Video game enthusiasts probably could have predicted many of these findings. But not everything was so cut-and-dried. You might think, for example, that popular games tend to be bright or highly saturated, like a good YouTube thumbnail tends to be. Instead, however, popular games tended to have darker, more muted aesthetics.
“I didn’t include this in the video, but I did color the map by [the] year of release and expected to see a strong gradient, assuming that visual styles had shifted a lot over time,” Newbie told Polygon. “According to this pipeline, that pattern just isn’t very obvious. That doesn’t mean it’s not there. The methodology has many limitations, and there are many other ways to analyze this, but I was surprised it wasn’t more evident. In retrospect, it kind of makes sense. There are modern games mimicking retro styles, older games that were visually ahead of their time, and a huge diversity of aesthetics all mixed together.”
Newbie cautioned that much of what he found and concluded from his experiment is preliminary. He thinks information like this might be more useful when one goes in with a specific goal in mind, like potential applications in recommendation systems.
“When it comes specifically to images, my personal take is that this kind of tool is most useful for exploration, comparison, and inspiration, or to help reason about very specific questions during development, rather than as a source of clean, actionable advice,” Newbie said.
“The goal was to test the space, see what’s even possible, and have some fun with it,” he continued. “Turning this into precise tools or strong claims would require a much more focused and careful follow-up.”
