This is my attempt to say what happens next in the search engines Google doesn’t run, written down so I can be scored on it. These are my forecasts for 2027 across Baidu, Yandex, Naver, Cốc Cốc, DeepSeek and Qwen, and each one rests on something that happened in 2026.
My rules are simple. I take what I’ve learned from watching global search engines and apply a high evidence threshold. A prediction needs a dated 2026 signal, a number I can check, and a confidence level I’m prepared to be embarrassed by. Where a company hasn’t said something, I won’t invent it. I wrote this on 12 November 2026, before Baidu’s Q3 results on 16 November and before the 25 November deadlines for the Qwen4-Max preview and ERNIE 5.5 Flash weights, so those could move a number below.
On confidence: 50% is a coin flip with a mild lean, 70% means I’d be surprised but not shocked to be wrong, and 85% means I think the evidence is close to settled. These are my forecasts, not company statements.
Baidu: the search box is shrinking, so the index is being sold
The 2026 signal is money. In Q2, Baidu’s online marketing revenue fell 19% year on year to RMB13.1bn, while AI-native marketing services were flat at RMB2.6bn after growing 36% in Q1, according to Drillr’s compilation of Baidu’s results. Baidu App monthly users fell from 679 million in December 2025 to 644 million in June. Management said it was delaying AI search monetisation on purpose. Meanwhile, Baidu has been turning search into something others can rent: a promotional ¥4.9 per 1,000 calls search API (list price ¥36 or more) and, at Baidu World, ERNIE Lightning, a model it says was built to serve Baidu Search.
1. Baidu will name a format or date for ads inside AI answers by its Q2 2027 results
Deferred revenue doesn’t stay deferred while ad revenue falls by a fifth. Morgan Stanley expects the Q3 decline to widen to 22%, according to 36Kr. Confidence: 65%. That the revenue will actually show up in 2027 is a harder call, and I’d put it at 40%.
2. Baidu App will end 2027 below its December 2025 level of 679 million monthly users
The app lost 35 million users in six months, and the assistants taking them belong to companies with more distribution. QuestMobile counted 419 million monthly users for Doubao and 158 million for Qwen in September. Baidu’s AI answer above, doctor below health layout shows where it can still win: high-stakes queries. Confidence: 75%.
3. No new named Baidu spam algorithm will be announced in 2027
Look at the Baidu algorithm updates page. Named releases such as Green Vine, Qingfeng, Jufeng and Jinglei largely stop after autumn 2021. What Baidu publishes now is tooling: AI traffic dashboards (ReadIsi reported one site logging 1M citations and 39 clicks), Search Widgets, and an Index Volume tool it admitted was unreliable for ten days in October. Attention has moved to skills, widgets and agents. Confidence: 85%. The model side is on the ERNIE model updates page, and the Base weights licence excludes the UK, EU and South Korea.
Yandex: Alice is winning the page while the plumbing is rebuilt
In Q3, Alice AI answered 44% of Yandex search queries, up from 42% in Q2, and the monthly audience for answers under the search bar rose 11% quarter on quarter to 54.8 million. Mediascope put Alice AI at 48% of Russians aged 12 and over in September. Then drones hit three data centres in October. Yandex withdrew guidance and wrote off 2.84% of Cloud customer data, and the Search API now runs only at night, at 10–20% of its old capacity.
1. Alice AI will answer 50% or more of Yandex queries by its Q3 2027 results
Two points a quarter is the current pace, slower than earlier in the year. At that rate 50% arrives in Q2 2027; I allow a quarter’s slippage because the easy queries are taken. What remains is commercial and local, where Ashmanov found Alice quick answers on roughly 15% of ecommerce queries. Confidence: 55%.
2. Yandex will launch an advertising product aimed at AI agents
At the National Advertising Forum on 11 November, Yandex Advertising’s Alexei Shtokolov said brands must soon advertise to agents as well as people, citing Yandex figures: 72% of online shoppers aged 18–54 would let an agent search and order, and Alice AI traffic to online shops grew 4.6 times from January to August. A day later, Alice began phoning businesses for users in a paid test, with a mass launch planned for the end of 2026. Confidence: 60% on a named agent-facing ad or catalogue format during 2027, and less on timing. Your Yandex Business card is the instruction sheet the agent works from.
3. Full daytime Search API capacity will not return before April 2027
Yandex has given no date. It says it won’t raise capex this year because it had already bought reserve capacity, and it is moving to a “more distributed” model. Pixel Tools says its switch to live SERP parsing may last one to two months, which I think is optimistic, especially while indexing logs show new pages ignored and old ones recrawled. Confidence: 60%. Treat October 2026 as a break in every Yandex series you keep. The Yandex algorithm updates history shows it has rebuilt ranking often, and data centres rarely. Note that Chinese advertisers spent over ₽10bn on Yandex in nine months.
Naver: AI Tab turns searches into transactions, and the blog is the raw material
Naver’s Q3 results on 12 November showed revenue up 10.7% to KRW3.47tn and operating profit down 20.9% on AI spending. AI Tab reached 10 million users by July, AI Briefing has more than 30 million monthly users, and the CEO said the flow from AI Tab into shopping and local services has nearly doubled since launch. Naver’s own AI Tab brief said 23.7% of queries lead to action, with click-through into Shopping and Place above 40%.
1. Naver will sell dedicated ad placements inside AI Tab by the end of Q2 2027
The Q3 slides talk about improving AI Tab revisits and conversion “to drive AI monetization in search”. ADVoost ads have run inside AI Briefing since July, and a personal AI agent app is due by the end of 2026. I’d be surprised if AI Tab stayed ad-light through the first half. Confidence: 75%.
2. Naver Blog will remain the largest single cited source in Korean AI answers all year
Acrosse’s GenRank index found 46.0% of cited-source AI answers cited a Naver Blog post in its 1 November edition. A separate study found 46.5% for AI Briefing and none in over 42,000 ChatGPT and Gemini citations. Naver blocks outside AI crawlers from its blog, café and Knowledge iN and pays Clip creators when its AI cites them, so the moat is deliberate. I expect blog share in GenRank to stay between 35% and 50%. Confidence: 70%. Both studies are vendor indexes, so I’m forecasting direction, not decimals.
3. The freshness war will keep producing unannounced shake-ups
Late on 29 October, community threads reported new blog and café posts jumping straight to the top of integrated search, and Naver said nothing. By 12 November there were reports of an indexing stall. Naver publishes models such as C-Rank and D.I.A. but no calendar, as the Naver algorithm updates page shows. I expect at least one unconfirmed, community-reported shake-up every quarter. Confidence: 85%. It’s a low bar, but many Western SEOs still expect an announcement. Local is moving too, with clinics matched to symptom and procedure searches from 12 November, and Korea’s two search-share trackers disagree by more than 20 points.
Cốc Cốc: small, regional and cautious about AI
StatCounter had Cốc Cốc at 5.6% of Vietnamese search in October 2026, against Google’s 93.38%. In August it held 7.15% of desktop searches and 3.49% of mobile. Cốc Cốc’s own 2026 credentials claim 30.5 million monthly users, 25.7 million of them on PC and 4.8 million on mobile. StatCounter’s monthly share bounced between about 3.4% in July and 6.0% in September, so treat panel data as weather, not climate.
1. Cốc Cốc’s AI answers will stay mostly informational
In our 100-query test in October, an AI answer feature appeared on 40 of 50 informational queries and only 3 of 50 commercial ones. It is still labelled experimental. If we repeat the test in Q3 2027, I expect AI features on fewer than 20% of commercial queries. Confidence: 70%. Its Q3 trends show why: searches for “giá iPhone 18” rose 2,802% quarter on quarter. Price and second-hand queries are where users still want a list of shops.
2. Ask AI will reach Cốc Cốc’s mobile apps
The sidebar assistant launched in April, runs on ChatGPT or Gemini, and gained YouTube summaries in July and translation in August. Cốc Cốc’s own page says it works only on Windows and macOS. Mobile is where its search share is thinnest, and where it has 4.8 million users. Ask AI answers the page you were about to read, which is awkward for publishers on any device. Confidence: 50%. Nothing public mentions a mobile version, so I won’t go higher.
3. Cốc Cốc will average 4–6% of Vietnamese search in 2027 and stay stronger on desktop
The 2026 range is wide, but the desktop-over-mobile gap has held. Its credentials deck sets a 2030 target of 40 million users, which would need a mobile breakthrough I can’t see in the data. Confidence: 75%. Whatever the share, Cốc Cốc still runs a real, national web index, and it is where a Vietnamese-language page can beat a global one.
DeepSeek: from chatbot to agent plumbing, and a listed company
QuestMobile counted 123 million monthly users for the DeepSeek app in September, against 419 million for Doubao and 158 million for Qwen. The share of AI-native app users who opened DeepSeek each month fell from 50.1% in August 2025 to 22.8% in August 2026. Yet its agent app, Harness, reached 1.54 million daily users on 4 November, at about 92 trillion tokens a day, and its funding round was reported at ¥105.5bn from nearly 30 investors. WSJ reported an annualised revenue run rate of $1bn in September.
1. DeepSeek will list on Shanghai’s STAR Market by the end of 2027
It hired CITIC Securities, which also put ¥8bn into the round, and the Wall Street Journal reported it was accelerating plans for a Shanghai listing in early 2027. Listings slip. Confidence: 65%. A prospectus would be the first audited look at what search and fetch earns.
2. DeepSeek will tier or surge-price its Search and Fetch API
The Search and Fetch API launched on 17 October at $0.75 per 1,000 successful calls, with an authority tier and a filter for “GEO manipulation”. Billing for agent harness calls started on 1 November. DeepSeek already introduced peak and off-peak model pricing in August, with off-peak at half the peak rate (see the DeepSeek model updates page). If that logic reaches search, it changes what agents can afford to read. Confidence: 55%.
3. Harness will pass 5 million daily users at some point in 2027
It went from about 271,000 daily users at the end of September to 1.54 million on 4 November, per Bloomberg. About 35,000 third-party plugins exist, an official marketplace is promised, and outages on 3 and 7 November show the strain. I’m guessing at the curve, not reading it. Confidence: 55%. Outside China, the US DHS order reaches third-party products built on DeepSeek.
Qwen: the weights are open, the app is bought, and Apple is the wild card
Nathan Lambert and Florian Brand found about 942 million cumulative Qwen downloads by March 2026, and 69% of new derivative models in February built on Qwen. The consumer side costs more. QuestMobile data analysed by Hello China Tech shows Qwen’s daily users peaked at 73.5 million in early February after a ¥3bn Lunar New Year promotion, then sat near 31 million by month’s end. Qwen took 77.7% of media ad spend among three tracked assistants in June. September’s monthly count was 158 million, down from about 167 million in June.
1. Qwen4 will get a full release and model card by 31 January 2027, with at least one open-weight model by 31 March
Alibaba posted a Zhuangzi fish teaser on 6 November and put Qwen4-Max into preview on 11 November with no model card, with the Token Plan preview closing on 25 November. It has shipped a major release roughly every quarter this year, as the Qwen model updates page records. Confidence: 70% on the full release and 60% on open weights, since the rumoured 27B tier comes from event recaps, not a model card. Qwen3.8-27B went out under Apache 2.0 in August, but the 2.4T model carries display and revenue conditions, so the licence matters more than the weights.
2. Alibaba will run another large holiday promotion through Qwen, and the spike will fade again
Lunar New Year falls in early February 2027, and the 2026 playbook worked well enough to repeat. The number to watch is where daily users settle afterwards. Confidence: 80%. The real test is whether the base holds if Alibaba cuts its ad budget.
3. Qwen will become a visible answer layer on Chinese iPhones
Apple confirmed Qwen in Siri, Visual Intelligence and Writing Tools for China from the 4 November beta, with general release in iOS 27.3 in December. Baidu isn’t named in Apple’s release. If Qwen takes “certain requests” in Siri, more questions are answered without a click, and I expect Chinese SEOs to be asking in H1 2027 how to appear in Qwen answers, where citations and publishers are already on the hook. Confidence: 70%. Agents like Qwen Go! add to that. New to Qwen? Start with the beginner’s guide.
What this means for international SEOs
The engine is becoming an index you rent. Baidu sells search calls, DeepSeek sells its index, Tencent opened Yuanbao Search, and Yandex and Naver are building agents on their own data. Visibility means being in the index agents buy, not only on page one.
The answer is eating informational queries, not commercial ones. Cốc Cốc’s test, Naver’s action-query data and Yandex’s 44% point the same way. Price, place and “near me” intent still earns the click.
Local entities beat websites. Yandex Business cards, Naver Place listings and Baidu widgets are what agents read. Keep them accurate.
Annotate your data. October 2026 is a break in Yandex series, 14 to 23 October is bad data in Baidu’s Index Volume tool, and StatCounter’s China numbers swung twelve points in a month. When a number moves faster than behaviour can, suspect the measurement first. And don’t over-correct on one source: GenRank is a vendor index, the AI Tab brief is Naver grading its own homework, and Yandex’s figures are company-reported.
I’ll revisit this in December 2027 and score myself publicly, including the 85% calls if they’re wrong. If you think one of my confidence levels is too high, tell me what evidence I’ve missed.
