Google has put a new model behind its search box. Gemini 4 Hydrogen is now powering Google Search for every user worldwide, and the company calls it the biggest model upgrade Search has ever had. This isn’t a beta or a limited Labs test. It’s live for everyone at once, which shows how confident Google is that its AI answers are now central to Search.
Google is pitching four main improvements. The first is better understanding of messy, real-world queries. Search should now cope better with long questions that have several parts, poor spelling or very specific conditions, and address each part rather than answering only the easiest one. Google credits stronger instruction following and tool calling for this, the same agent-style skills AI labs have been prioritising across their models this year.
The second improvement is what Google calls query fan-out. Instead of running a single search, the model breaks a question into subtopics, searches them in parallel and combines the results. Google says Hydrogen’s reasoning makes this more thorough and surfaces a wider range of websites, including ones people might not otherwise find.
That claim deserves some scrutiny. Since AI Overviews launched in 2024 and AI Mode followed, publishers have argued that AI-generated answers keep readers on Google instead of sending them to the sites the information came from. Saying Hydrogen surfaces more links is clearly meant to answer that criticism. Whether those links actually get clicked is a separate question, and Google hasn’t shared any traffic data to support it.
Third, answers are getting more visual. Google promises more images, structured layouts, comparison tables and custom interactive tools built for specific questions. That’s useful if you’re comparing laptops or planning a trip. It also means more of the answer happens on Google’s own page.
Finally, Google says responses will be more nuanced across every language AI Mode supports. That matters a lot in regions like the Middle East, where Arabic-language AI search has generally lagged behind English in quality.
What Google hasn’t said matters too. There are no published accuracy figures for Search, no details on how often Hydrogen gets facts wrong compared with earlier models, and no explanation of how it chooses which sources to cite. AI answers in Search have made embarrassing mistakes before, and a more capable model doesn’t remove that risk. It mainly makes confident mistakes sound more convincing.
The bigger picture is that Google is no longer treating AI answers as an optional extra on top of its links. With Hydrogen live for everyone, the real test will be whether users trust its answers and whether publishers see any real benefit from the extra links Google says it’s surfacing.