Last week was the week generative search outside Google stopped being a demo and started producing numbers, some of them flattering and some of them brutal. Baidu handed webmasters an AI traffic dashboard and the first screenshots were humbling. Yandex gave its best Alice models away for free. Naver kept tightening the loop around its own properties. DeepSeek and Alibaba spent the week arguing about speed, chips and what counts as a launch. Here are the stories from 28 September to 4 October that actually matter if you run search outside the Google bubble, with our take on each.
Baidu: lots of citations, hardly any clicks
The single most useful data point of the week came from China. One brand site’s Search Resource Platform screenshots showed roughly 1,004,742 AI citations against about 120 AI impressions and 39 AI clicks over 30 days. We picked it apart in Baidu AI Traffic Analysis Reality Check, and the forum chatter in our Baidu Webmaster Recap confirms that plenty of dashboards are still empty because the data service is rolling out in phases.
Our take: anyone selling “get cited by Baidu AI” as a traffic plan should be made to show this screenshot to the client first. Being cited and being visited are now different metrics, and Baidu is the first engine to say so in its own tooling. It gets worse for the citation chasers: an academic audit found Baidu’s Chinese AI overviews are far more source-concentrated than Google’s, so the pie is small and a few sites already own it.
Baidu is also quietly rewriting what a submission is. Search Widgets lets verified sites submit interactive tools rather than URLs, and the invite-only Authority Provenance trial looks like the old news-source whitelist rebuilt for models. If you work in China, those two are the long-term stories; the dashboard is just the one people are screenshotting.
Yandex: Alice is the results page now
Yandex says Alice AI quick answers reached 49.5 million monthly users and 42% of queries in Q2. Then it unlocked its flagship Alice chat models for free, with live search citations, added Expert mode and automatic model routing, and open-sourced the base of AliceAI-Foundation-80B under Apache 2.0.
Our take: when a generative layer answers four in ten queries, it isn’t an experiment any more. Russian-language SEOs are running the AI-SERP economics test at national scale, a long way ahead of most English-language markets. On the practical side, Wordstat moving into the Cloud Search API is the quiet winner. It has already spawned agent wrappers for demand research, and that will change CIS keyword research workflows faster than any Alice headline.
Naver: AI everywhere, citations at home
Naver rolled its conversational AI Tab out to all users, put ads inside AI Briefing, and turned Whale into an AI browser with sidebar AI Chat. The number to remember comes from our Naver community recap: a summer source study put Naver-owned properties at about 82% of AI Briefing citations.
Our take: Naver is monetising generative answers faster than most Western engines dared, and it’s keeping the sources in-house while it does it. For brands in Korea, Blog, Cafe and Smart Store presence isn’t a nice-to-have alongside your website. It’s the citation pool. Also check which query types trigger Briefing at all before you rewrite a single page.
DeepSeek and Qwen: speed, silicon and slideware
DeepSeek open-sourced Ascend tooling built with Huawei, a bigger strategic move than any benchmark because it targets CUDA lock-in directly. It also folded V4 Pro traffic into V4.1 Flash, and independent timing showed Flash beating Qwen3.8 Flash on speed, with caveats about thinking mode. Alibaba, meanwhile, put Qwen 4 on a keynote slide without shipping it.
Our take: for search people, the Qwen story that matters isn’t Qwen 4. It’s Model Studio’s first-class web search tool, which means more Chinese apps will be grounding answers in live web results and citing publishers. And with QuestMobile putting Qwen past 160 million users, Alibaba’s consumer AI reach is no longer theoretical.
Europe and the independents: who actually owns an index?
Ecosia and Qwant brought their European index live in Germany and are selling Staan to AI developers. That is a welcome counterweight to the uncomfortable truth that most “European” search engines still rent Bing or Google. Swisscows admitted that fully anonymous search is now a paid product, and Kagi walked back its Translate paywall after a backlash.
Our take: both stories come back to who owns the index. If your results come from someone else, so do your terms. Keep the short list of real crawlers handy the next time someone pitches “multi-engine SEO”. Elsewhere, Yahoo’s CEO argued that Scout keeps the blue links because micropayments will never pay publishers, and Brave is using its new Chrome listing to sell Brave Browser back to Chrome users.
The long tail: Shenma, Weibo and Cốc Cốc
We spent part of the week on markets most decks skip. Start with our Shenma for business guide, then read the YisouSpider robots.txt trap, because a lot of “Shenma doesn’t work” post-mortems turn out to be crawl blocks. On Weibo, we argued it’s a discovery engine, not a small Baidu. In Vietnam, Cốc Cốc’s Q2 trends showed jobs queries up 27%.
One more thing
If you missed it, International Search News is now International Search Insights (ISI). Same focus on the engines that Google-first coverage ignores, with a name that better fits the analysis we’re publishing.
The thread running through the whole week is simple. Generative answers are now the default surface in Moscow, Seoul and Beijing, and the engines are being unusually honest in their own tools about how little traffic those answers send. Measure citations and clicks separately, know who owns the index you’re optimising for, and stop treating non-Google search as a footnote. See you next Monday.