I went down a rabbit hole this morning reading the Google 2025年度最热门AI应用 recaps side by side with the late-2025 Juejin picking roundups, and what crystallized for me is that Google's own annual ranking of most-popular AI products does not include a chatbot at all. The top eight categories Google itself published for 2025 were 学习理解, NotebookLM的应用, 旅行游玩, 相册应用与图像编辑, 办公生产力, 购物, 硬件设备, and 个性化定制. Every one of those is either an output-transformation surface or an AI-embedded-in-existing-product surface, not a chat surface. The reader-jobs Google says are most popular inside its own product line — turn a PDF into a podcast, turn a math problem into a guided-learning tutorial, turn a spreadsheet cell into a semantic category, turn a search query into a working mini-app, turn a holiday cookie photo into a structured recipe book — do not show up on the Juejin picking scorecard at all, because the picking scorecard is still measuring how good the typed-prompt reply is.
The Juejin picking roundups I read this morning are still committed to ranking raw chat capability. The late-2025 picking piece puts Cursor at S档 for IDE workflow and Claude Code at A档 for code generation, with Codex at A档 and a long tail of CodeBuddy at 9.6, Cody at 8.2, Ghostwriter at 8.0, and Codeium at 7.8 on a five-axis decimal scorecard. The 2025年度盘点 piece names Gemini as 首选 for 多模态 plus 超长上下文 plus 原生搜索, ChatGPT as 首选 for 通用性强, and Claude as 其他 for 代码能力强 and 长文写作. Both roundups ask how well this AI responds to a typed prompt and answer in either S/A/B/D tier letters or 9.6-out-of-10 decimal scores. To be fair I would take the exact decimals with a grain of salt because the test corpus is never disclosed, but the structural tell is that neither roundup has a column for what artifact the AI produces. NotebookLM turning a research paper into a podcast is not on the picking scorecard. Gemini 3 turning a mortgage-rate question into an interactive calculator is not on the picking scorecard. The picking scorecard measures how smart the text response is and the Google recap measures what new thing exists in the world because the AI ran.
The meta-pattern I want to put down is that the 2026 search results page has at least three formats that all claim to answer what AI to use in 2026, but they are actually answering three different questions and the cross-format bridge is invisible. The picking roundups ask which chat is most capable and answer with tier letters and decimal scores. The Google annual recap asks which AI products users actually engage with inside Google's own ecosystem and answers with output-transformation and product-integration tools. The October GitHub trending recap asks which open-source project got the most stars last month and answers with prompt-eng-interactive-tutorial, Agent-S, claude-cookbooks, supermemory, and TradingAgents-CN — none of which is a chat surface either. Honestly I am a little skeptical of any 2026 roundup workflow that pulls a single piece off the search results page and treats it as the answer to what AI to use, because each piece was written for a different reader-job and the cross-format integration is left to the engineer.
The practical takeaway I want to write down is that the picking roundups are still useful for the within-chat anchor and the Google recaps are useful for the within-transformation anchor, but neither format is useful for the cross-format re-rank most engineers are quietly trying to do this quarter. The picking roundup did name Cursor, Claude Code, Codex, CodeBuddy, Cody, Ghostwriter, Codeium, Tabnine, CodeWhisperer, and Blackbox as the editor-loop shortlist. The Google recap did name 学习理解, NotebookLM, the Gemini Sheets AI function, Gemini in Search, and the AI mode mini-apps as the transformation shortlist. Neither list is good at telling the engineer which transformation tool to add on top of which chat-capable tool, because the picking scorecard never asks what artifact each row produces and the Google recap never asks how each row compares on raw chat quality. My gut says the reader has to do the cross-format multiplication by hand, and the multiplication is in none of the roundups.
I will reassess in three months. For now I am still mostly on Cursor Pro plus Claude Code for coding, ChatGPT Plus for general chat, Gemini AI Pro for the Sheets AI and Drive features I already use, and NotebookLM for the occasional PDF-to-podcast experiment. What has changed is that I now read the Google recaps as a separate category — output-transformation tools and AI-in-product integrations — rather than as a chat-ranking I should fold into the picking scorecard. Give it six months and I expect either the picking roundups to add a what-artifact-does-this-AI-produce column or the Google-style transformation recaps to get their own dedicated ranking niche, and whichever moves first will tell me whether the format has finally noticed that the reader-job has forked.
Top comments (0)