Google announced Gemini 3.8 Live on September 15, 2026, with a post updated September 17 that lists Search Live among its rollout destinations. Google describes the model as combining voice conversation with visual context and background tool execution. Its separate Extended Thinking model is aimed at more complex tasks; the announcement does not identify that variant as Search Live's model.

The update follows Google's March expansion of Search Live to more than 200 countries and territories where AI Mode is available. That earlier announcement describes spoken follow-up questions, optional camera context and links to supporting information. September's development is an upgrade to that conversational experience, rather than the launch of voice search itself.

A different unit of measurement

Our assessment is that the useful unit for a visibility audit is a complete interaction. A typed question asked in a fresh session is one condition. A spoken question about an object on camera is another. A follow-up that adds a budget, a location or a compatibility requirement changes the task again. Treating all three as interchangeable would hide the context the tester supplied.

Consider a hypothetical shopper asking whether a replacement component fits a device. The initial question might be vague, the camera might expose a model number, and the next question might introduce an installation constraint. A brand could be relevant at one stage and irrelevant at another. An audit should preserve that sequence before drawing conclusions about whether the brand appeared.

This is a recommendation for better observation, not a claim that spoken prompts have a special ranking formula. Google's announcement establishes product capabilities. It does not disclose how publishers are selected for every spoken answer or provide a controlled test of any content tactic.

What a brand can test now

Build a short list of real customer tasks from support questions, product documentation and sales conversations. Select tasks that require a decision, clarification or explanation. For each, write a starting question and a consistent follow-up sequence. Where visual context is essential, use the same object or image across repeated checks.

Record the product surface, account condition, date, language and region. Keep a transcript of what was asked and answered, then distinguish a spoken brand mention from an exposed source link and a visit to the website. If a source is unavailable to inspect, mark it as unavailable rather than guessing which page supplied the answer.

Repeat the interaction on another day and preserve the differences. A test plan should be small enough to review manually and specific enough that a colleague can reproduce its conditions. Broad prompts such as which brand is best are less diagnostic than tasks with explicit customer constraints.

For an initial review, we would prioritize these questions:

  • Did the answer identify the correct product or service?
  • Did the follow-up preserve important qualifications?
  • Were the displayed links relevant to the task?
  • Could a visitor complete the next step on the linked page?

These checks describe the proposed audit. They are not results from a test conducted by this publication.

CiteCue provides a complementary site-usability workflow. Its Agent Usability documentation describes an AI agent attempting tasks such as locating pricing, understanding an offering and finding contact information, with reports showing where each task succeeds or stalls. That task-level approach can help teams turn the last question above into concrete page improvements. It examines what an agent can do on a website; the spoken Search Live interaction should still be observed separately under the conditions described here.

Improve the evidence available to customers

Google's website guidance recommends accessible textual information supported by useful images and video, and says AI Overviews and AI Mode require no special schema. Those principles are a sensible starting point for an editorial review, while remaining distinct from a documented Search Live optimization recipe.

We would review pages for explicit model names, compatible versions, limitations and clear instructions. Illustrations should help a reader identify the relevant part or step. A support answer should explain when the advice applies and when to seek another route. These changes are defensible because they improve the underlying customer information, even when their effect on assistant visibility remains unmeasured.

The next evidence to watch is how spoken answers expose sources and how consistently a brand survives follow-up questions. Until there are observations from the relevant surface, teams should label conversational-search plans as hypotheses. The September upgrade gives them a reason to investigate a richer customer journey, without assuming that a dashboard built for typed prompts already measures it.