On 1 September 2026, ZTE and ByteDance's Nubia NaviX Ultra received MIIT network approval, planned for sale in September. The so-called first global AI-agent phone ships with Doubao assistant, moving AI search and agent entry from apps onto hardware (Sina Finance, Netease Tech, 1 September 2026). It is the second-generation "Doubao phone": the first generation (engineering model Nubia M153), co-released with the Doubao on-device model in December 2025 at 3,499 RMB, sold out at launch and was resold for up to 12,999 RMB on secondary markets (Sohu timeline, 2026). Unlike conventional AI phones that mainly answer questions and generate content, NaviX Ultra is built around agent capability: with user authorization, the assistant understands natural-language instructions and completes operations across multiple apps (TechNode, 3 September 2026).
For brands doing GEO, this is not just consumer electronics: traffic allocation is being rewritten. You used to need to be recommended in Doubao app answers; soon you also need to be cited first by the phone's voice assistant. Background: CNNIC's 57th report put generative AI users at 602 million by December 2025 (42.8% penetration). QuestMobile's 2026 AI H1 report put Q1 Doubao MAU at 340 million, Qwen 170 million, DeepSeek 130 million. Hardware is scaling in parallel: IDC forecasts 2026 China new-generation AI phone shipments of 147 million units, up 31.6% year on year and accounting for 53% of the market (IDC, December 2025), while globally IDC projects 432 million Gen AI phones in 2026, 39.7% of shipments. When hundreds of millions move questions onto assistants that come preinstalled, the "citation slot" extends from SERPs to system-level voice.
Entry and form: from app search to phone voice
App-era path: open → type → read. Agent phones put Q&A at the system layer: wake the assistant, never open a browser. Competition upgrades from SERP rank to source priority when the assistant answers. Doubao already rebuilt source weights toward official sites, official media and verified accounts (see Doubao indexing); hardware will amplify that. The practical test is simple: say your own buying question out loud to the phone assistant, and see which source the answer cites first.
Spoken answers need structured, short, extractable copy. To be cited on "which Shanghai website company is reliable" or "how much does GEO cost," organize core facts as Q&A — the job of Schema and FAQs (see Structured data and LLM inclusion). Voice queries are longer and more colloquial, so cover them as long-tail content: "哪家上海建站公司靠谱" is not the same keyword as "web design company." Build a question–answer layer on the site and a brand knowledge base that can be read aloud without awkward phrasing.
Multimodal and how the phone cites brands
Feixiang.net reported Volcano Engine's Shenzhen AI tour on 24 September, where Doubao would ship a new video model. Combined with the phone's multimodal UI, answers move from text to image+video. Text-only GEO will miss that window. Stock product demos, location footage and how-tos, crawlable and identifiable (see Doubao multimodal GEO).
ITHome reported that NaviX Ultra will not read the screen or simulate clicks without user authorization (the so-called GUI technique); instead, apps are expected to expose MCP (Model Context Protocol) services that open specific data and controls to the assistant, and the on-device Doubao model has completed CAC on-device generative-AI filing. For brands, the implication is concrete: to be "operable" and "citable" by the phone assistant, you need both machine-readable content and, increasingly, structured service interfaces. This is where GEO starts to meet API and MCP strategy — an extension of the same logic as structured data for LLM inclusion. Brands that publish stable, well-typed business facts now will be the ones agents can act on later.
A four-step layout for the AI hardware era
| Entry | User behavior | Content fit | Action |
|---|---|---|---|
| In-app search | Type keywords | Structured text + authority | Site FAQ, Schema, knowledge base |
| Phone voice | Wake and speak | Short, speakable Q&A | Spoken long-tail, voice-friendly layout |
| Multimodal | Mixed image/video | Recognizable media + metadata | Real video, demos, alt/titles |
| Mini program / web | Convert after recommend | Fast, complete, trust signals | Mobile UX, credentials, cases |
More entries need measurement. Log core terms on Doubao, Kimi, DeepSeek and Yuanbao: frequency, position, description accuracy — monitor → diagnose → optimize → retest. Keep site, mini program and commerce pages consistent so multi-entry information does not conflict. Add one hardware-specific check: test the same question through the phone's voice assistant, not only the app, because the two can cite different sources. For help scoping an AI-phone visibility program, contact Zheming Digital Communication Research Institute (+86 18917757529, jaysun@widesight.cn) — engagement terms are subject to our quotation.
FAQ
Does the Doubao phone matter for ordinary companies? Yes. A preinstalled assistant means users ask by voice. If you are not cited, you are absent at the moment they speak. Local services, ecommerce and B2B are all in scope.
How does content get cited by the phone assistant? Same RAG logic as the app: authoritative, structured, question-relevant sources. Official site, official media, verified accounts and high-quality Q&A lead.
Do we need a separate "voice SEO" project? No, but fold spoken Q&A into GEO. Voice queries are longer and more colloquial — cover them as long-tail.
What is MCP and should we care? MCP (Model Context Protocol) is how apps can expose data and controls to agents without the phone simulating clicks. For most brands, publishing clean structured data now is the prerequisite; MCP services matter later, mainly for software and SaaS businesses.
Is it too late to layout hardware entries? No — it is the window. The phone ships in September; the video model was not yet out; most brands have not systematized multimodal and voice. Structured knowledge bases first.
GEO vs SEO in the AI-phone era? SEO optimizes page rank; GEO optimizes citation and recommendation in answers. Phones amplify the gap: users do not browse ten links; they trust the assistant. Quality and source authority outweigh keyword density.
Network approval: Sina Finance, Netease Tech (1 September 2026), TechNode (3 September 2026), ITHome, Beijing Daily and PANews. User scale: CNNIC 57th report. App MAU: QuestMobile 2026 AI H1. Phone shipments: IDC (December 2025). Volcano tour: Feixiang.net.
Related reading
- AI Search 2026 Trends: From Conversational Tools to Decision Gateways
- Brand Strategy for the AI Agent Era
This article was written by Zheming Digital Communication Research Institute. Data updated to 2026. AI hardware-era GEO consultation: +86 18917757529 ｜ jaysun@widesight.cn.