Voice Search Is Back—But This Time It's AI-Powered

The numbers point the same way. In its one-year AI Mode report, Google said the average AI Mode search is triple the length of a traditional search, and its US insights deck adds that more than one in six AI Mode searches are non-text: voice, images, video or a Live conversation. Follow-up queries grew by more than 40% a month. Those figures cover all of AI Mode in the US, not voice alone, but they describe the habits voice brings.

Most voice search advice was written for a different machine. It tells you to win a featured snippet, add FAQ markup and target "near me" phrases, because an old assistant read one short answer aloud. This post explains what changed, which assistants matter as of September 2026 and where each gets its answers, how voice prompts differ from typed ones, and a spoken-answer format you can apply to any page. For the step-by-step setup work, see our voice search optimization checklist for 2026.

Why Voice Search Is Back, and What Changed

The first voice assistants treated speech as a remote control. You said a short command or question, and the assistant ran one lookup and read one answer. That shaped a decade of voice search advice: keep answers short, win the snippet the assistant reads, and optimize for "near me."

The new assistants work more like a person on the phone. They listen while they talk, let you interrupt and keep the thread across turns. OpenAI says ChatGPT's Live voice can listen and speak at the same time and can use web search and memory. Amazon says you can talk to Alexa+ naturally, change topics and skip the wake word after the first turn. Apple says users can turn almost any Siri response into a rich conversation and ask follow-up questions.

The table sums up the shift.

The third row is the one that matters most for your content. Google defines query fan-out in its AI optimization guide as "a set of concurrent, related queries generated by the model." OpenAI says ChatGPT search rewrites your question into one or more targeted queries for its search partners. Microsoft says Copilot sends the Bing search service a query of a few words informed by the user's prompt. A spoken question now fans out into several typed searches you never see.

The AI Assistants Behind Voice Search in September 2026

Six assistants cover most AI voice search today. They differ in where they live and how much they say about their sources. The table records what each vendor states on its own pages, checked on 29 September 2026.

Google: Search Live and Gemini Live

Google runs two voice experiences, and they're easy to mix up. Search Live is voice inside Search. Google expanded it on 26 March 2026 to every language and location where AI Mode is offered. You tap Live in the Google app, ask out loud, hear an audio answer, then "continue the conversation with follow-up questions or dive deeper with helpful web links." Its help page adds that you can reopen a Live session from your AI Mode history. Since mid-September it runs on Gemini 3.8 Live.

Gemini Live is voice inside the Gemini app. Google's help page says it can use connected apps such as Google Flights, Hotels and Shopping, and "can automatically use public information from Google Maps in your Live chats." On answer sources, Google's crawler documentation says Gemini Apps ground answers in content "from the Google Search index." Our Gemini guide covers what that means for citations.

Gemini now reaches far more people by voice than it did a year ago. A Google community manager wrote that on 4 September 2026 most users could no longer switch back to Google Assistant on Android phones. On speakers, Gemini for Home is in early access in 20 countries and 10 languages. Gemini Live on a speaker needs a Google Home Premium plan.

ChatGPT Voice

ChatGPT's voice lives inside a normal chat, so you can listen, read along or type. The ChatGPT Voice help page, updated in late September 2026, lists three options. Live runs on GPT-Live-1 or GPT-Live-1 mini and "can use web search and memory, show visual results through supported widgets." Advanced is the older real-time mode, and Standard transcribes your speech before replying. Free users get limited Live access, and Plus includes three hours in any 24-hour period.

The search side is spelled out on OpenAI's search help page: "During a voice conversation, you can ask ChatGPT to search the web." Restaurant reservation search isn't available in voice. The same page lists Microsoft among its search providers and says that to be eligible, a site should allow OAI-SearchBot. Our ChatGPT SEO guide covers that crawler in depth.

Siri AI and Apple Intelligence

Apple unveiled Siri AI at WWDC in June and began rolling it out in beta in English on 14 September 2026, with French, Japanese, Korean, Portuguese and Spanish due next month. It isn't yet available on iPhone, iPad or Apple Watch in the EU, or anywhere in China. Apple's example of a web question is telling: "the best local coffee spot in a new city." It says Siri AI will draw on information from across the web to answer.

Apple doesn't name a search partner. Its models are "custom-built in collaboration with Google and its Gemini models," but that describes the model, not where web facts come from. The clearest statement is on Apple's Applebot page, published 4 September 2026. It says Applebot's crawl may be used for "answering broad world knowledge questions in Siri and Search that may include links to sources." Siri can also hand a request to ChatGPT if the user turns on the ChatGPT extension.

Alexa+

Alexa+ is free with Prime in the US and $19.99 a month without it, with a limited free text chat at Alexa.com. Amazon lists 10 countries beyond the US, from the UK to Australia, and India joined early access on 16 September 2026.

Amazon describes three kinds of source. It partnered with news outlets including the Associated Press, Reuters and more than 200 others. It connects to services such as Yelp, Tripadvisor, OpenTable and Thumbtack. And its crawler page says allowing Amzn-SearchBot makes your content "eligible to appear in search experiences such as Alexa," while Amzn-User fetches live pages when an Alexa question needs current facts.

Copilot Voice

Microsoft's Copilot transparency note, dated 18 August 2026, says that for certain questions Copilot is grounded in web search results, and that it gives linked citations "in conversation transcripts in voice mode." Microsoft's web search docs say those results come from the Bing search service. That makes Bing's index, and Bing Places for local facts, the path into Copilot's spoken answers. Our Copilot guide covers Bing's side.

How Voice Search Prompts Differ From Typed Search

Put the vendor data and the research together and five differences stand out. Each one changes what a page needs before it can win a voice search answer.

1. Questions are longer and phrased like speech

Google's deck lists the top first words in AI Mode queries as "What," "How," "I," "Is" and "Can." People ask the question on their mind instead of guessing keywords. This isn't new to AI. Ido Guy's 2018 study of a commercial search app's logs, published in ACM Transactions on Information Systems, found that spoken queries were closer to natural language than typed ones and leaned toward question answering. AI assistants simply stopped punishing that habit.

2. One question becomes a conversation

A typed search often ends with a click. A voice search often ends with "and what about...?" Google said in May 2025 that Gemini Live conversations run five times longer than text chats on average. It gave no method, so treat that as a vendor claim. The 40% monthly growth in AI Mode follow-ups points the same way.

3. Context comes along for free

The listener doesn't repeat themselves, so the assistant fills the gaps. OpenAI says ChatGPT may use saved memories and rough location when it rewrites a search query: a vegan in San Francisco asking for "restaurants you might like" becomes a search for vegan restaurants in San Francisco. Our post on how AI search uses user intent and context covers these signals in detail. For your page, it means the facts that match a condition (price for a team size, a feature for a use case) must be written down, or the assistant can't match them.

4. People ask for a pick, not a list

Voice users want someone to choose. Google's US data shows searches starting with "where to," "where should I" and "ideas for" growing, and planning queries growing 80% faster than AI Mode overall in six months. A list of ten links doesn't work out loud. The assistant has to name one or two options and say why. Pages that state who a product suits, and who it doesn't, give it that reason. Our model of the four conversational buyer stages maps these turns from first question to handoff.

5. Names have to survive the microphone

A typed brand name arrives spelled right. A spoken one has to be recognized first. In a SIGIR 2023 paper, Apple researcher Christophe Van Gysel gives the example of "what is borrelia" being misheard as "what is gorilla" when the speech model rates the rare word as unlikely. Rare names are exactly what brands have. Say your product name the way customers say it, and spell it the same way everywhere, so the assistant links the spoken name to the right entity.

The Spoken-Answer Format

When an assistant answers out loud, it reads a few sentences and moves on. Google's own guidance for its text-to-speech markup recommends 20 to 30 seconds of content, or roughly two to three sentences, per spoken section. That's a useful budget for any passage you want quoted in a voice search answer.

We call the structure that fits this budget the Spoken-Answer Format. It has four blocks and one supporting layer. You can apply it to a product page, a comparison page or a guide.

Why each block earns its place

The say-it definition gives the assistant a sentence it can repeat word for word. Put the name and category together ("Tallyfold is an invoicing app for agencies"), so the sentence stands alone when lifted out of the page. If you need the general case for answer-first content, our passage-level guide covers it.

The decisive recommendation answers the question people actually ask by voice: which one? Hedged copy ("it depends on your needs") gives an assistant nothing to say. A condition plus a reason gives it a full spoken sentence and tells the listener whether it applies to them.

The comparison table might seem pointless in a spoken answer. It isn't. An assistant can pull a row from it when the follow-up comes, and phone assistants show results on screen: ChatGPT's Live mode can "show visual results through supported widgets." Each row should answer a likely follow-up, and each cell should make sense spoken on its own: "$60 a month for five users," not "60/5." Our comparison page formula covers fair tables in depth.

The transparent caveat does two jobs. It tells the listener when the pick is wrong for them, which makes the recommendation believable. And a date ("prices checked on 29 September 2026") lets the assistant and the listener judge how fresh the fact is.

The follow-up ladder exists because voice users keep talking. After "which one," they ask "how much," "does it work with my accounting tool" and "is it cheaper than the other one." Answer each under its own question-shaped H3, with the answer in the first sentence. Our reverse prompt playbook has a matching kit of Answer Units for pricing and fit.

The 30-second read-aloud test

Once a section is written, read it out loud with a timer. Score one point for each check it passes.

1. Answer first. The first sentence answers the question on its own.
2. Under 30 seconds. The answer block takes 30 seconds or less to say.
3. No pointing. It never says "above," "below," "this table" or "click here."
4. Named, not implied. The product and category are named in the answer, not referred to as "it" or "our tool."
5. A pick with a reason. It recommends something for a stated situation.
6. Numbers you can hear. Each sentence carries one or two numbers at most, with units said out loud.
7. A dated caveat. It says what the option isn't for and when the facts were checked.
8. Easy to say. A stranger can pronounce the brand name on the first try, and it's spelled the same way everywhere.

Seven or eight is ready. Five or six needs a trim. Anything lower needs the before-and-after treatment below.

Before and After: A Tallyfold Section Rewritten for Voice

Tallyfold is a fictional invoicing and payments app for agencies. Brindlework and Kestrelyn are fictional rivals. All names, prices and features here are made up for the example.

A buyer runs a voice search in Gemini Live: "What's a good invoicing app for a five-person design agency that bills clients a monthly retainer?" Here is the section on Tallyfold's comparison page that an assistant might read.

Before: 89 words, 1 of 8 checks

Tallyfold is the all-in-one platform trusted by forward-thinking agencies everywhere. Our powerful features help you streamline billing, get paid faster and scale with confidence. Unlike legacy tools, we were built for the way modern teams really work. See the table below for how we compare, and check out our pricing page for plans that fit every budget. Whether you're a solo freelancer or a growing studio, Tallyfold has you covered. Ready to transform your cash flow? Start your free trial today and see why agencies love working with us.

It passes only check 8: the name is easy to say. It never says what Tallyfold is for, never picks anything and points at a table the listener can't see. An assistant can't quote a single sentence of it in answer to the buyer's question.

After: 97 words and a table, 8 of 8 checks

Tallyfold is an invoicing and payments app for agencies of 2 to 50 people that bill clients on retainers or by project. If your agency bills most clients a fixed monthly retainer, Tallyfold is the stronger pick, because it sends retainer invoices on a schedule and follows up on late payments by email. If you bill mostly by the hour, choose Brindlework, whose built-in time tracking turns hours into invoices. Tallyfold doesn't run payroll or track time. Prices were checked on 29 September 2026 and can change, so confirm them on each pricing page before you buy.

What changed, and which check each change passes

The arithmetic on length: the definition and the recommendation total 70 words. At an assumed speaking pace of about 150 words a minute, that's 70 ÷ 150 × 60 = 28 seconds, inside the 30-second budget. The definition alone is 22 words, or about 9 seconds. The old version ran 89 words, about 36 seconds, and said nothing a buyer could use. The table and caveat are there for the follow-ups.

Now the follow-up ladder. Under the section, Tallyfold adds three H3s the buyer is likely to ask next: "How much does Tallyfold cost for a five-person agency?", "Does Tallyfold work with our accounting software?" and "Is Tallyfold cheaper than Brindlework?" Each opens with a one-sentence answer. When the buyer says "and how much would that be for us?", the assistant has a sentence ready.

What to Keep and Drop From Old Voice Search Advice

Most voice search guides still repeat tactics from the smart speaker era. Some still help. Some no longer do anything. Here's the sorting, with the evidence.

Markup built for the old assistant is going away piece by piece. Back in March 2024, Google dropped guided recipes from its docs because that Google Assistant feature was removed. And Google says structured data isn't required for its AI features. Use markup for the rich results it still supports, not as a voice search switch.

Local businesses have the most to gain from the listing work, because voice search prompts about places ("a dentist near the office that's open Saturday") lean on listings. Our guide to optimizing your business for AI search covers which listing feeds which engine.

How to Check Your Voice Search Visibility

No assistant tells site owners what people asked it by voice. Search Console's generative AI performance report counts impressions in AI Overviews and AI Mode, and doesn't mention Search Live or voice separately. So checking voice search is a manual job. It's easier than it sounds, because every assistant here leaves a written record.

1. Write 10 spoken prompts. Use full sentences with conditions, the way a buyer talks: team size, budget, location, a rival's name. Our free AI Visibility Prompt Kit gives you 30 buyer prompts to start from.
2. Ask them out loud in Search Live, Gemini, ChatGPT Voice and Copilot, plus Siri AI or Alexa+ if your buyers use them.
3. Ask two follow-ups each time: "how much would that cost for us?" and "is it better than [rival]?"
4. Read the transcript. Search Live keeps sessions in AI Mode history, ChatGPT adds a transcript to the chat, and Copilot shows citations in the voice transcript.
5. Log three things: whether you were named, whether your facts were right, and which pages or listings were cited.

Voice search answers vary from run to run, so repeat the panel monthly and read rates, not single answers. Our guide to measuring GEO explains sample sizes and a control test.

Where Rankbox Fits in Voice Search

Rankbox writes new articles; it doesn't track voice answers or AI citations today, and it doesn't refresh or re-date your existing pages. Where it helps is the writing. The Citation-Ready Writer researches the live web and drafts 2,000 to 3,500-word articles with sources linked, and Brand Voice applies the tone, audience and product details you give it. Brief it with the Spoken-Answer Format (a say-it definition, a decisive pick and a dated caveat), and you get drafts built for voice search answers.

Articles reach your site through Rankbox's API, which a developer connects. The Business plan is $49.50 a month with a 7-day trial. See plans and pricing.