Google Search Live now runs on Gemini 3.8 Live, the company’s latest real-time voice AI model. Announced on September 15, 2026 by Rajan Patel, VP of Engineering for Search and co-founder of Google Lens, the upgrade delivers responses that include web links, automatic language switching across 97 supported languages mid-conversation, and more natural dialogue flow to all Search Live users globally.
What changed when Search Live moved to Gemini 3.8 Live?
Patel announced the model change on X on September 15, 2026. According to his post, users now get responses that include embedded web links for deeper exploration, fluid multilingual support with mid-conversation switching, and more natural free-flowing interactions. The upgrade applies to all users and requires no opt-in.
According to Search Engine Roundtable, Search Live has existed for roughly two years, first in audio-only format and later with video support. It rolled out globally in March 2026, following an earlier removal of the opt-in requirement.
What are Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking?
Google launched two models simultaneously, according to the Google blog (September 2026). They share a real-time voice architecture but serve different use cases.
Gemini 3.8 Live is built for scale and cost efficiency. It processes visual inputs in near real-time, automatically detects and transitions between 97 supported languages within a single conversation, and executes tools and API calls (connections to external services or data sources) in the background while the conversation continues uninterrupted. This is the model now powering Search Live.
Gemini 3.8 Live Extended Thinking is built for high-complexity tasks. It reasons and speaks simultaneously, using verbal cues like “Let me check that…” to acknowledge a request naturally, then narrates progress through multi-step background tasks as they run. Google has made it available inside Google Docs, Gmail, and Keep via Workspace integration.
How do these models perform on independent benchmarks?
According to Google’s blog (September 2026), Gemini 3.8 Live Extended Thinking ranked first overall on Artificial Analysis’ Speech to Speech Quality Index with a score of 82.6. It led in agentic task completion (the ability of an AI to independently execute sequences of actions) at 68.6% on the τ-Voice benchmark and 35.1% on Sierra’s τ-Voice-banking benchmark, which evaluates voice agents in banking scenarios. It scored 97.7% on Big Bench Audio, a standard benchmark for audio-based reasoning.
Gemini 3.8 Live, the model now in Search Live, placed second in the Speech Agent Arena, a community-run ranking of voice AI performance. Google describes it as “highly cost-effective” for developers and enterprises deploying at scale.
On ServiceNow’s EVA-Bench, a benchmark for evaluating voice agents in enterprise workflows, both models pushed what Google calls the Pareto Frontier, meaning they improved both accuracy and conversational quality at the same time, rather than trading one for the other.
Which developer platforms and enterprise partners are involved?
As of the announcement, both models are available through the Gemini API, Google Workspace, and the Gemini app, according to Google’s blog. Developer platforms including Agora, Fishjam, LangChain, LiveKit, Pipecat, Vercel, and Vision Agents have integrated the Gemini Live API to let developers build voice-driven interfaces without managing real-time media streaming infrastructure themselves.
Enterprise partners Salesforce, Genspark, and Lumeris are listed by Google as early adopters, citing low latency and tool-calling capabilities as key factors. All audio generated by Search Live and other Google AI products is watermarked with SynthID, an imperceptible signal embedded in the audio output to support transparency and attribution.
What this means for AI-search visibility
Voice search used to be a brand-awareness channel with no clear path back to your site. That is changing.
The most significant detail in this upgrade is not the benchmark scores. It is the inclusion of web links in spoken responses. Until now, when Search Live answered a query out loud, brands appeared as unnamed sources. The AI spoke the answer; no one knew where it came from. With Gemini 3.8 Live, those responses now carry links. That makes the citation layer audible and clickable, not just implicit. A brand cited in a voice answer now has a measurable referral path, which was not guaranteed in earlier versions of Search Live.
The 97-language mid-conversation switching also changes what multilingual content optimization means in practice. Previously, a brand might optimize one English page for voice search. With seamless language switching within a single session, Search Live may pull from sources in multiple languages to construct one answer. Content that exists only in one language is potentially invisible for portions of a multilingual conversation, even when the user starts in English.
Background tool execution raises a separate question for businesses in sectors where AI agents take action on search results, not just summarize them. Gemini 3.8 Live can call external APIs while speaking. Whether that action reaches your brand depends partly on whether your data is accessible via structured formats or APIs, not just web pages designed for human reading.
One implication the sources do not address: when a real-time voice model must deliver a single spoken answer, it cannot hedge by listing three sources the way a text-based AI overview might. That selection pressure likely increases the weight given to sources that are consistently structured, factually stable, and clearly attributed. Brands with scattered, contradictory, or poorly marked-up information may find themselves systematically excluded from voice answers, without any visible signal that it is happening.
Hingewise’s assessment: the link inclusion is the variable most worth tracking. If Search Live reliably surfaces web links alongside spoken answers at scale, voice search moves from an untraceable brand impression into a measurable traffic source. That is a structural change in the channel, not just a model upgrade. Whether Google maintains that behavior consistently or rolls it back based on user engagement data is the open question that will determine whether voice becomes a serious referral source or remains ambient background noise for most brands.
Quick checklist: what to verify before this change affects your visibility
- Run your top 5 target queries in Search Live and check whether your brand or content appears as a cited source.
- Audit your structured data markup (schema.org annotations that help AI parse your content accurately) on pages that answer common questions in your category.
- Verify that your most important content exists in the primary languages your audience speaks, not only English.
- Check page speed on mobile, since voice queries are predominantly mobile and slow-loading pages may reduce crawlability.
- Confirm whether your APIs or data feeds are accessible to external agents if you operate in a sector where AI tool-calling is relevant, such as travel, banking, or healthcare.
- Search for your brand name in Search Live when asking questions you should own, and note whether a competitor is cited instead.
The broader pattern to watch: as Search Live becomes a mainstream interface rather than an early-adopter feature, the transparency of its citation and attribution logic will come under greater scrutiny from both publishers and regulators. The link inclusion signals Google is aware of that pressure. How consistently it holds across query types and languages is a question the data will answer over the next few months.
Lam Nguyen · Hingewise
Sources
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.