Yandex Transformed Its Search Engine into an AI-Powered Ecosystem

Over two years, Yandex transformed from a classic search engine into a comprehensive AI-powered ecosystem, with its own YandexGPT language model (six generations), the Neuro product handling 20% of queries, and 43 million subscribers to the Yandex Plus service. The company chose a hybrid approach — AI answers above traditional results rather than replacing the search engine — thereby preserving its advertising model and reducing hallucinations thanks to its own web index covering 130 billion pages. Total revenue in 2024 reached 1,095 billion rubles (+37%), with the long-term monetization of AI answers remaining the key question.

In two years, Yandex transformed from a classic search engine into an AI-driven platform generating over one trillion rubles a year. Six generations of its own language model, AI search handling a fifth of all queries, 43 million subscribers, and a cloud platform growing 50% annually — that is the balance sheet of the transformation the Russian technology giant has carried out since its corporate restructuring in July 2024.

Yandex belongs to a relatively narrow group of companies worldwide with a completely proprietary web index. Alongside Google and Bing, a proprietary index is operated by, among others, Brave, Mojeek, China's Baidu and Sogou, South Korea's Naver, and the Czech Seznam. Unlike DuckDuckGo, Ecosia, or Yahoo, however, Yandex neither resells nor purchases results from another provider. This independence has proven crucial: the proprietary index serves as a verification layer for AI-generated answers, thereby reducing hallucinations and sustaining the advertising model.

The following analysis maps what Yandex has built over two years — from the perspective of the user, the technologist, and the financial analyst. It deliberately makes no comparison with any other company.

The most visible change is the Neuro product, launched in April 2024. Instead of a classic list of links, the user receives a summary answer assembled from current web sources, always with links to the references. Since October 2024, Neuro has processed approximately 20% of all searches automatically; for the rest, the user can activate it manually.

The page architecture is layered: the AI-assembled answer is displayed above the traditional list of links, which is retained. The user additionally has a dialog mode available — they can ask follow-up questions in the context of the original query. Yandex thus chose an overlay model rather than a replacement of classic results.

In Turkey, where Yandex operates under the Yazeka brand, the rollout is more aggressive — AI assists with more than 40% of searches, and the user base grew by 75% over 2025. In December 2025, Yandex expanded AI search globally to the yandex.com domain. It is one of the few search engines combining a proprietary web index with a proprietary language model and global availability.

The most tangible AI innovation is the shopping agent — a system that autonomously browses offers across e-shops and compares prices. Yandex Search uses an extensive database of product offers for this. The company is also testing direct purchasing from search results and is introducing specialized verticals for goods, real estate, and financial products.

The voice assistant Alice, with 66 million monthly users, is among the first virtual assistants fully powered by a large language model (YandexGPT 3). It runs on smart speakers, mobile devices, in the car, and in the smart home.

Since December 2025, the Yandex Browser has included an AI-powered editor for creating and editing texts, a translator with contextual selection of specialized terminology, and a tool for summarizing documents in PDF and DOCX formats as well as entire books.

Yandex Plus is a bundled subscription combining video (Kinopoisk — 4.7 billion hours of viewing annually, according to Yandex), music, free delivery from the Market department store, cashback, e-books, and sports. In the second quarter of 2025, Yandex Plus had 43.2 million subscribers, representing a year-on-year increase of 28%.

YandexART 2.0 generates images with text using a hybrid architecture combining convolutional and transformer networks. In Yandex's internal tests, it outperforms Midjourney V6.1 on complexity in 66% of cases and on aesthetics in 58% (no independent comparison is available). According to Yandex, 11% of advertisers on the Yandex Direct advertising platform already use images generated with YandexART, with the combination of generated and proprietary visuals increasing campaign effectiveness by 10–15%. The Shedevrum app then generates complete posts — text and illustration at the same time.

At the core of the transformation is the YandexGPT family of language models. Since the first announcement in February 2023, it has gone through six generations.

YaLM 2.0 / YandexGPT (February–May 2023) was the first generation integrated into the Alice assistant. YandexGPT 2 (September 2023) was rated better than the first version in 67% of cases. YandexGPT 3 (April 2024) became the foundation of the Neuro product. YandexGPT 4 (autumn 2024) offered a fourfold larger context. YandexGPT 5 Pro (February 2025), the main fifth-generation model, was preferred over GPT-4o in 54% of cases in Yandex's internal blind test — meaning approximate comparability rather than clear superiority. YandexGPT 5.1 Pro (July 2025) is the current main model.

The fifth-generation model was pretrained on 15 trillion tokens (70% Russian, 30% English) and supports a context window of 32,000 tokens in two variants — Pro (deep reasoning for corporate customers) and Lite (optimized for low latency). The API is compatible with the OpenAI libraries in both Python and JavaScript and supports integration with LangChain and AutoGPT.

A technical article by Andrey Syutkin, head of ML development for the Neuro product (Medium, October 2024), describes a five-step procedure.

First, the user poses a query in the context of the conversation. Second, query reformulation occurs — a compact version of YandexGPT 3 analyzes the last 16 messages of the conversation and generates a search query. The model is trained via supervised learning and repeated cycles of active learning on more than 10,000 manually created reformulations.

Third, the search engine returns the most relevant documents from its proprietary index covering over 130 billion pages with a 400-millisecond latency. Fourth, 5 documents are selected from them. Pages are split into overlapping blocks of 256 tokens, ranked by a compact BERT model (8 layers) according to relevance to the query, and at most 1,500 tokens per document are selected. CatBoost (on search factors) and BERT (on page texts) predict which document the generative model can best answer from.

Fifth, a generative model from the YandexGPT 3 family — not the general version, but a specially trained model that must not use facts from internal memory, only from the supplied sources — assembles the answer with links.

Why exactly 5 documents and a limit of 8,000 input tokens? According to Syutkin, the difference in answer quality between one and two documents is large, between five and six already negligible, but each additional document significantly increases GPU consumption. It is an empirically verified balance between quality, latency, and cost.

The generative model's training path: supervised learning on thousands of hand-written answers, generation of answer pairs, evaluation by assessors, relevance annotation with hallucination detection, and training of reward models.

Yandex selected 200,000 anonymized queries from the search stream, filtered out the irrelevant ones, and kept only those where users were seeking information — excluding queries with the intent to watch a video, shop, or visit a social network. Neuro thus activates primarily on informational queries requiring synthesis from multiple sources. Navigational, transactional, and media queries remain on classic results.

Yandex explicitly rejects the model of fully replacing the search engine with a conversational bot. Pyotr Yermakov of Yandex Search's management stated in an interview for Turing Post (April 2025) that the optimal solution is a balanced combination of traditional search and a large language model — for two reasons: the computational demands of full processing by a language model and the risk of hallucinations.

Yandex defines two extreme approaches. First: find the page with the answer and have the language model merely summarize it — verifiable, but incomplete. Second: query the model from its internal memory — it answers anything, but hallucinates, has outdated data, and cannot handle rare facts. Yandex chose a middle path: the model receives the query and the relevant pages, must not add its own facts, but may draw logical conclusions, and every part of the answer must have a verifiable source.

Two-stage retrieval of relevant results (fast coarse filtering, then precise reranking) recurs across Yandex's products. The same pattern is used by the geocoder in Yandex Maps (dual BERT encoder → HNSW index → reranking by a BERT cross-encoder), search, and Neuro. Everywhere, they work with BERT models fine-tuned to the specific task rather than general ones, with active-learning cycles for iterative improvement, and with small specialized teams — the geocoder for the entire world is run by five machine-learning engineers.

Total revenue for 2024 reached 1,094.6 billion rubles (~$12.3 billion), representing year-on-year growth of 37%. Adjusted EBITDA was 188.6 billion rubles with a margin of 17.2%. In the second quarter of 2025, revenue reached 332.5 billion rubles (+33% year-on-year), adjusted net profit 30.4 billion rubles (+34%), and the EBITDA margin improved to 19.9%.

The largest segment is e-commerce, mobility, and delivery, with revenue of roughly 592 billion rubles for 2024. The second is search and portal, with approximately 440 billion rubles (+30%). The advertising segment (including real estate) brought in 34.1 billion rubles (+41%). Yandex Cloud reported 19.8 billion rubles (+50%). The entire group together reached roughly 1,095 billion rubles (+37%).

Yandex Direct serves hundreds of thousands of advertisers and delivers billions of ads daily across more than 55,000 partner platforms. Approximately 25 different neural networks are involved in delivering each ad impression.

AI increases the effectiveness of the advertising business in three ways: neural networks optimize targeting and bids; YandexART generates visual materials directly within the Direct platform; and higher user engagement thanks to AI features leads to longer visits and more ad impressions.

Monetization of AI search is, for now, indirect. Neuro is free for users. Yandex's management admits it does not yet have a direct model for monetizing AI answers — strategically it relies on increased user engagement leading to more ad impressions from the classic results below the AI answer. Yandex does not display advertising within the AI answer itself.

The entertainment segment generates 29.7 billion rubles per quarter (+34% year-on-year). The Yandex Plus model functions as an ecosystem bond: a single payment ties the user to streamed video, music, e-commerce, and other services. The more services a user uses, the higher the cost of switching to a competitor.

Yandex Cloud reported revenue of 19.8 billion rubles in 2024 (approximately +49% year-on-year), serving more than 44,000 customers, with corporate customers accounting for 51% of consumption. Internal consumption by the Yandex group represents less than 5% of capacity — the cloud is therefore primarily an external business.

The market for corporate technology services in Russia is estimated at 240 billion rubles (2024), with expected growth to 510 billion rubles by 2028.

The Yandex Cloud AI Studio service (launched in September 2025) offers no-code creation of AI agents, multi-agent collaboration systems, and a catalog of more than 24 models including proprietary ones (YandexGPT) and openly available ones (Qwen3 235B, Gemma3, DeepSeek VL2). The API pricing model: YandexGPT Pro 5.1 at $0.0067 / 1,000 tokens (synchronous mode), Lite at $0.00167, asynchronous mode at a 50% discount. The Boost AI grant program provides up to 1 million rubles.

AI assistants are integrated into the cloud platform — a console assistant that designs infrastructure based on natural language, WebSQL AI for database queries, YDB AI for self-diagnostics, and DataLens for analyzing tabular data. According to Yandex, they reduce customers' routine tasks by up to 30%.

The two-year course of the transformation yields six principles that Yandex has verified in live operation.

AI as an extension, not a replacement. Yandex did not throw away 27 years of search-engine development. Traditional search serves as a verification layer for the language model — it reduces hallucinations, sustains monetization, and enables a gradual transition for users. The architecture goes through search first, only then through the language model.

A proprietary index as a prerequisite. Neuro runs on top of the proprietary index, not as a separate product. Without a proprietary index, the verification layer is missing and dependence on an external provider arises.

The ecosystem bundle. Yandex Plus's 43 million subscribers create a stable revenue base resilient to fluctuations in the advertising market.

Corporate cloud as a second engine. The cloud is growing by ~50% year-on-year versus 37% for the entire group. AI Studio, with no-code agent creation and an OpenAI-compatible API, lowers the barrier to entry for corporate customers.

Localization as a differentiator. In its expansion (Turkey +75%, the global domain), Yandex invests primarily in the quality of AI in the local language. Language specialization works better than a general approach.

Selective open source. CatBoost, Perforator, and YandexGPT 5 Lite Pretrain (openly available weights since February 2025) build a developer community without endangering the core business.

First, Yandex openly admits it does not yet have a direct model for monetizing AI-generated answers. As the share of these answers grows (20% today, potentially significantly more), the pressure for direct monetization will intensify — but inserting ads into AI answers risks degrading the user experience.

Second, the computational costs of AI answers are orders of magnitude higher than those of classic results. Yandex addresses this with selective query routing (an AI answer only for informational queries) and the limit of 5 documents / 8,000 tokens, but scaling to a higher share will require further optimization.

Third, financial data is reported in rubles, and Russia's geopolitical situation complicates international comparison as well as the assessment of the sustainability of growth. The July 2024 restructuring (sale of the Russian assets to a consortium of domestic investors for approximately $5.4 billion, renaming of the international entity to Nebius Group) created a complex corporate structure: Yandex LLC in Russia, Nebius Group in Amsterdam (AI cloud infrastructure), Avride in Austin (autonomous vehicles), Yango Group in Dubai (delivery, robotics), and Toloka (data annotation).

Group revenue: in 2023 approximately 800 billion rubles, in 2024 a total of 1,095 billion rubles (+37%), and in the second quarter of 2025, 332.5 billion rubles for the quarter (+33% year-on-year).

EBITDA margin: in 2024 it was 17.2%, in the second quarter of 2025 it improved to 19.9%.

Yandex Plus subscribers: roughly 30 million at the end of 2023, about 39 million at the end of 2024, and 43.2 million in the second quarter of 2025.

Cloud revenue: in 2024 a total of 19.8 billion rubles, with over 44,000 customers.

Search market share in Russia: steadily around 66%.

Yandex has demonstrated a model in which a regionally dominant technology company can successfully deploy AI from the user interface (Neuro, Alice, the browser) through the technical backend (25 neural networks in the advertising system) all the way to the corporate platform (Cloud AI Studio). With its key decision — hybrid AI search instead of replacing the search engine — the company preserved its existing monetization and opened up new revenue streams.

The open question remains how long Yandex can distribute AI answers for free. Indirect monetization (higher user engagement = more advertising) works as long as the AI answer generates a sufficient volume of interactions with the classic results. Once the share of AI answers exceeds a critical threshold — and in Turkey, at 40%, it is already approaching — a new model will be needed. Autonomous shopping agents and the offering of AI as a service for businesses indicate where the path might lead.

The article is based on the following primary and secondary sources, verified as of March 2026:

Yandex financial reports for 2024 and the first half of 2025 (Statista, TASS, Online Marketplaces, Moscow Times, Reuters). Yandex official press releases (yandex.com/company/news). AI Studio documentation and pricing (yandex.cloud/en/docs/ai-studio/pricing). Andrey Syutkin: "Yandex's Neuro: Here's How It Works" — Medium, October 2024, technical description of answer processing. Pyotr Yermakov: interview for Turing Post (April 2025) — AI search strategy. Intelligent CIO ME (May 2024) — interview with Yandex management about advertising AI. TAdviser — profiles of YandexGPT 5.1, AI Studio, Yandex Cloud. International Search News (2024–2025) — analyses of the Neuro product and global expansion. Anadolu Agency, Türkiye Today — Yandex metrics in Turkey. CanvasBusinessModel.com — analysis of the growth strategy. Wikipedia: Yandex, YandexGPT, Alice AI.

Transparency of creation

The concept, structure, and editorial line of the article are the work of the author, who prepared the content outline, established the key theses, and directed the entire creation process. Generative AI (Claude Opus 4.6, Anthropic) was used as a tool for research, fact-checking, and fleshing out the author's draft.

The author verified the key findings and approved the final wording. No part of the text was published without conscious authorial control. The factual data was verified against the publicly available sources cited in the text.

The procedure complies with the transparency principles of EU Regulation 2024/1689 (AI Act). #poweredByAI