Google Put Gemini 4 Hydrogen Behind Search for Everyone on Nov. 12, and Published No Click Figures
A.I. / news
Google Put Gemini 4 Hydrogen Behind Search for Everyone on Nov. 12, and Published No Click Figures
The one randomised study of AI Overviews found a 39.8% drop in outbound clicks on early-2026 Search.

Google made Gemini 4 Hydrogen available to all Search users worldwide on Nov. 12, and its announcement contains no benchmark, usage or traffic figure. The post, by Robby Stein, Google's VP of Product for Search, calls it "the biggest model upgrade to Search yet" and says the model now sits behind AI Mode and the harder questions in AI Overviews.
Whether it sends more visits to publishers is the open question, and the only controlled data on AI Overviews points the other way.
What Google says changed in Search
Stein's post lists four areas. Search handles longer, multi-part questions and follows instructions more precisely through improved tool calling. Query fan-out, the practice of running simultaneous searches across subtopics, links to a wider range of sites. Answers carry more images, structured layouts, comparison tables and custom tools. Responses are more detailed across the languages AI Mode supports, which the post does not list.
Barry Schwartz of Search Engine Roundtable reported the same day that the release coincided with a core update and that Gemini 4 Carbon had reached paid AI Mode subscribers earlier that week. He wrote that he suspects Hydrogen affects AI Overviews and AI Mode but not core rankings, and added that the post offers no independent count of links or their placement.

The click data Google has not answered
Google says fan-out links to sites users might not have found before. It has not said how many outbound clicks Hydrogen produces per search.
The best public number comes from a vendor reading of an academic paper. RevScope, which sells analytics software, summarised a working paper by Saharsh Agarwal and Ananya Sen. In a randomised experiment, 1,065 US desktop Chrome users ran a browser extension that hid AI Overviews for one group over two weeks in January and February 2026. Searches that would have shown an Overview produced 0.37 outbound clicks with it and 0.62 without, a 39.8% reduction.
- Overviews hidden0.62 clicks
- Overviews shown0.37 clicks
Source: RevScope's summary of Agarwal and Sen working paper, accessed 2026-11-18
RevScope states the limits. The paper has not been peer reviewed, and RevScope read a June 2026 draft. It covers US desktop Chrome only and a two-week window. The effect was statistically detectable for informational searches, while navigational and transactional samples were too small to read. Across all searches the authors estimate a loss of about 18.5% of organic clicks. Ad clicks did not change.
The experiment predates Hydrogen by nine months, so it measures the Overviews of early 2026, not this model. It is the baseline against which a Hydrogen claim would be tested.
What would settle it
Google's wording promises link breadth, not link volume, and the two can move in opposite directions. A page can be linked more often in an answer that makes a click less necessary.
The check is a repeat of the randomised design on Hydrogen-era Search, or Google publishing outbound clicks per AI Mode session. Neither exists as of Nov. 18. Google has said Hydrogen is live for everyone, so any publisher traffic change after Nov. 12 now overlaps with the core update Schwartz flagged, and the two cannot be separated from outside.
For another account of how model choice changes cost and behaviour, see Runta's 27,280-run harness evaluation, and for how Google's models are used against defenders, Google's exploit_agent finding.
Sources
More in A.I.
- 01Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of AstraArtificial Analysis puts Carbon at 36.0 minutes per Intelligence Index task against a 12.4-minute median, and a Kernel study traces most of the gap to extra steps.
- 02Researcher Says Claude Fable 5.2 Learned Protein Sequences; Test Shows a 45-Point GapXinming Tu's probe of 668 obscure bacterial proteins finds the newest Claude models predict real sequences far better than shuffled ones. GPT, Gemini and open-weight models do not.
- 03Tencent's Hy4-flash Weights Carry a License That Requires a Separate Deal Above $50 Million in RevenueThe 12B-active model is ungated on Hugging Face, but it ships under the Tencent Hy License, not Apache 2.0.
- 04OpenAI's GPT-6.2 Astra Lists at $8 per Million Input Tokens, 20% Under Its Predecessor, While Coding Scores Stay FlatArtificial Analysis ranks Astra second at 62 behind Claude Fable 5.2, and Sol's coding-agent score fell 3 points.