If you can’t understand something, you can’t fix it. That’s the motivation behind a new platform that aims to cut through the noise of Steam reviews and find out what players really feel about a game, what the developers need to address, and even what game they should think about making next.
“Years of building, pitching, and raising for games made the same failure painful to watch,” says Justin French, co-founder and CEO of the start-up GameDataCore in Brighton, UK. He tells GamesIndustry.biz that the problem is that studios, publishers, and investors often make decisions about the same game based on a “different version of reality,” through different assumptions and interpretations. What’s needed, he thinks, is a shared understanding.
The trouble is that player feedback can be difficult to analyse. We’ve all seen stories of players leaving a negative Steam review after spending more than a hundred hours playing a game – yet clearly it couldn’t have been all bad. How did they really feel about it?
“What people say and what they mean are different things, and the useful bit is in the gap,” says Simon Sparks, co-founder and COO of GameDataCore. “One review usually holds three or four different things: something praised, something complained about, a comparison, a bug, a claim. A thumbs-up flattens all of that into one signal.”
There are already tools available that can sum up player sentiment towards a game based on reviews, like SteamSummarize. But Sparks says that whereas SteamSummarize is aimed at helping consumers to decide whether to buy a game, GameDataCore’s platform is aimed at developers, and it’s far more thorough.
“We break every review into snippets and classify each one across topic, emotion, tone and motivation, then weight it by who wrote it, using profile signals like library, playtime, and what else they play,” he says. “That’s how you find out a racing game’s reviewers don’t own another racing game, which changes what every complaint means. It also lets us look at the silent players: the people who played but never left a review, which is most of them.”
The idea is that similar complaints – such as about combat mechanics – can have different motivations behind them and be weighted accordingly. One player, for example, might be frustrated because they have never played a game in the same genre before, whereas another, veteran player might simply be bored after 100 hours.
The launch of the platform is deliberately being timed to coincide with Steam Next Fest, when hundreds of demos for new games will be made available on Steam between October 19 and 26. Sparks says that it’s an opportunity for developers to discover whether the feelings they intended to evoke in their new game are the same ones felt by their audience. “Next Fest is the week you find out.”
Laundry list
As well as interpreting what players feel, the GameDataCore platform provides a list of potential issues to address that are weighted by the potential impact on the game, along with the profiles of the players reporting the issue and what they said about it. The user can review the evidence for each one, looking at the player reviews that raise the issue, and then share the problem with their team via Slack or Discord.
Sparks points out that although it’s easy to spot things like game-breaking bugs by reading reviews, it’s harder to spot the “difference between the loudest complaint and the most important one” or the “praise sitting quietly behind a negative review”.
The interpretation is done by a series of underlying LLM models. These models are all based in Europe rather than being US-based frontier models, and GameDataCore says that no customer data will be sent to a third-party model provider.
“A frontier model gives a different answer each time it sees the same input”
Sparks says there are three reasons for choosing European-based models: “First, data sovereignty,” he says, pointing out that studios are “rightly sceptical about where their players’ words end up.” The second is cost and consistency. “A frontier model gives a different answer each time it sees the same input, and on a large data set it’s slow and expensive,” he says. By contrast, GameDataCore runs around ten small models, each trained for one job, and together they are small enough to run on a laptop.
The final reason is dependency. “Nobody knows where the frontier labs are going, and we didn’t want a product sitting on someone else’s roadmap and pricing.”
A pilot version of the platform has been in testing with 24 studios and publishers since March 2026, and GameDataCore says the plan is to fold in more sources of player feedback beyond just Steam reviews, including Discord communities, forums, Reddit, TikTok, YouTube, and Twitch. The company also plans to enable studios to bring in their own Steamworks data to the platform, such as wishlists, sales, and activations.
The platform could also help studios and publishers to plan for the future by discovering underexploited niches. “If you have a hunch about a niche, we tell you who’s in it,” says Sparks. He notes that by analysing several games in a genre, it’s possible to dig down into the review data to isolate a specific audience and discover what other things they play and what they’re asking for. “Where a cohort shares a complaint nobody has answered, that’s the gap.”
French hopes that by having greater understanding of what players liked, what they didn’t, and what needs to be fixed, the cycle of failure won’t be repeated. “The part the industry is bad at is learning from its own decisions,” he says. “Very few studios or publishers run a proper retrospective on what they chose to do and what it did. We make that the default.”