Google began rolling out Gemini 3.8 Live to Search Live on the Google app on September 15, replacing the audio model that previously ran the feature’s real-time voice conversations. Rajan Patel, Google’s VP of engineering for search, announced the swap on X, framing it as a move toward more natural, free-flowing spoken exchanges that surface web links mid-conversation.
Search Live lets a user tap a “Live” icon inside the Google app, ask a question aloud, and hear a synthesized response while related links appear on screen for follow-up reading. A transcript view lets users switch to typed follow-up questions, and past sessions remain accessible through AI Mode history.
Patel listed three specific gains from the new model: more helpful responses paired with links for deeper reading, fluid multilingual support that lets a user change languages inside one conversation, and more natural interaction patterns overall. Google did not publish comparative metrics showing how much the new model improves response quality over its predecessor within Search Live specifically.
The rollout follows Google’s own announcement of two related audio models, Gemini 3.8 Live and a variant called 3.8 Live Extended Thinking, which are also reaching the Gemini API, AI Studio, Gemini Enterprise, Gemini Live, and Workspace. Artificial Analysis, a third-party benchmark index, ranked the Extended Thinking variant first on its Speech-to-Speech Quality Index at 82.6 and credited it with handling 97 languages within a single call. That score comes from an outside evaluator, not from Google, and measures speech quality in isolation rather than the accuracy of the links or answers Search Live actually surfaces.
This publication reported on September 3 that Google had already swapped the model behind AI Mode to Gemini 3.8 Flash, a separate rollout from the audio model now running Search Live.
Google’s own line on the update, that the models “continue to get better and the results will change as these models change,” describes a direction, not a measured outcome. The company has not said how it evaluates whether a voice answer pointed a user to the right source or the wrong one.
For search teams, a voice-first answer that reads a link aloud while also displaying it on screen changes the click surface entirely. The relevant question is no longer only which page ranks, but which page’s link the model chooses to speak and show mid-conversation, an interaction pattern current rank-tracking tools were not built to capture.
Search Live launched inside the Google app in 2025 as a companion to AI Mode, giving Google a growing footprint of conversational search surfaces beyond the traditional results page. Each model swap so far has arrived with limited public detail on ranking or citation behavior, leaving practitioners to infer impact from usage patterns rather than documentation.
Teams that rely on Search Console impression data should start segmenting Search Live-referred sessions where the data allows it. A voice interface that names sources aloud may reward differently than a results page that displays ten links at once, and that difference will not show up in standard ranking reports.
Search Engine Land, in a report by Barry Schwartz published September 15, 2026, first reported the Gemini 3.8 Live rollout to Search Live.
