Yandex has published named search platforms since 2007. This page lists the releases the company, or its webmaster team, actually described, plus the one commercial-link change that was a policy switch rather than a city name. Unofficial nicknames that Russian SEO blogs invented to keep the alphabet chain going (Nachalovo, Odessa, Amsterdam, Kirov and the like) are left out. Dates below are announcement or documented launch dates. Yandex also ships unnamed formula tweaks every week; those are not rows.
| Date | Update | What changed |
|---|---|---|
| 28 October 2025 | Alice AI | New Alice generation in Search, on the YandexGPT 5 family. Generative answers are a normal result, not a side experiment. |
| 22 May 2025 | Alice in Search | Reasoning mode, text and image generation, and longer answers with images, video and source links. Neuro is folded into Alice. |
| 3 October 2024 | Search with Neuro | Generative Neuro answers inside ordinary results. Source pages are still chosen by the classic ranking stack. |
| 28 November 2022 | Y2 | Annual search release: object search, “Big Topics” personalisation, translated video, with CS YATI in the ranking stack. |
| 20 September 2022 | CS YATI | Transformer trained on IT and machine-learning documents plus expert judgements, for programming queries. |
| 10 June 2021 | Y1 | More than 2,000 improvements on YATI and YALM: many more quick answers, in-video search, and review summaries. Proxima is the commercial quality metric. |
| 25 November 2020 | YATI | Transformer ranking, announced at YaC 2020. Yandex called it the largest quality jump since MatrixNet. Volatility showed up from September. |
| 17 December 2019 | Vega | Semantic index clusters, pre-rendered results, expert assessors on specialist topics, and hyperlocal ranking. |
| 19 November 2018 | Andromeda | More weight on site quality, richer quick answers, official-site marks, and tighter links to other Yandex services. |
| 22 August 2017 | Korolyov | Neural match of the query to the whole page, and to other queries that reach that page. Palekh only read the title. |
| 23 March 2017 | Baden-Baden | Over-optimised text is demoted inside the main formula, not only by a post-filter. |
| 2 November 2016 | Palekh | Neural ranking for rare and natural-language queries, comparing the query with the page title. |
| 2 February 2016 | Vladivostok | Mobile-friendly pages promoted in the mobile results. |
| 15 May 2015 | Minusinsk | Sites with too many SEO links demoted. Link factors then counted again more broadly, including commercial Moscow queries. |
| 12 March 2014 | Commercial link switch-off | Link factors dropped for groups of commercial queries in Moscow. Descriptive name, not a Yandex city brand. |
| 30 May 2013 | Dublin | Personalisation from short-term interests inside the same search session. |
| 12 December 2012 | Kaliningrad | Personalisation from long-term interests and sites the user already prefers. |
| 17 August 2011 | Reykjavik | Language preferences. The first real step towards personal results. |
| 15 December 2010 | Krasnodar (Spectrum) | A query is split into intents so the page is not one meaning only. Stronger city-level formulas. |
| 13 September 2010 | Obninsk | Re-tuned ranking for geo-independent queries, most of the stream. |
| 17 November 2009 | Snezhinsk (MatrixNet) | MatrixNet machine learning, far more ranking factors, and separate formulas for major regions. |
| 10 April 2009 | Arzamas | Regional ranking for major Russian cities, with point releases through 2009. Yandex’s company history dates this only to 2009. |
| 11 September 2008 | Nakhodka | Stop-words kept, a thesaurus, and a new machine-learning pass. |
| 16 May 2008 | Magadan | Faster candidate selection, plus broader synonyms and abbreviations. Magadan 2.0 followed on 2 July 2008. |
| 2 July 2007 | Ranking formula 7 | New formula and more factors. Announced to webmasters, not branded as a city. |
Key algorithm events
Yandex’s public ranking story starts later than the company. The engine itself is from 1997. Named platforms start in 2007, once the formula was large enough that webmasters noticed a single cutover. Formula 7 in July 2007, then Formula 8 that winter, were still “a new formula”: more factors, more trust for established sites, and the first serious filtering of link runs. Magadan (May 2008) and Nakhodka (September 2008) are the bridge. Magadan is about speed and language coverage. Nakhodka is about keeping function words and learning from a thesaurus instead of raw tokens. None of that is a penalty brand. It is Yandex teaching the formula Russian.
The break is November 2009. Snezhinsk shipped MatrixNet, the gradient-boosted tree model Yandex had built to handle far more factors than a hand-tuned formula. The company history still treats MatrixNet as the machine-learning breakthrough, and later staff compared YATI only with that moment. Arzamas, earlier in 2009, had already made the user’s region a first-class input, with separate formulas for large cities. Obninsk (2010) retuned the geo-independent majority. Krasnodar, the same December, added Spectrum: one query can mean several intents, so the results should not collapse onto a single reading. That is the ancestor of every “diversity” argument in Russian SEO.
Personalisation is a three-step staircase, and the names are Yandex’s. Reykjavik (August 2011) starts with language. Kaliningrad (December 2012) uses long-term interests and sites the person already trusts. Dublin (May 2013) shortens the window to the session, so the results can move while you are still searching. Y2’s “Big Topics” block in 2022 is the same idea pulled into a visible module. If a ranking report ignores personalisation, it is describing a logged-out average that many users never see.
Links are the part people still argue about, and the dates are specific. On 12 March 2014 Yandex stopped using link factors for groups of commercial queries in Moscow. That was a live experiment, not a city-branded platform, which is why this table does not call it Nachalovo. Minusinsk, on 15 May 2015, went the other way in public: sites whose link profile was dominated by SEO links were demoted, and link factors were allowed to count again, including on those commercial queries, once the obvious spam was priced in. Buying links did not “stop working” in 2014 and stay dead. The abuse did.
Text and neural ranking are the next cluster. Vladivostok (February 2016) is the mobile-friendly boost and is easy to underestimate because it is narrow. Palekh (November 2016) is the first neural ranker aimed at the long tail, matching a query to a title when the words do not overlap. Baden-Baden (March 2017) is the over-optimisation penalty moved into the core formula. Korolyov (August 2017) reads the whole document and the other queries that lead to it. Yandex’s own history describes that hand-off in those terms. Andromeda (2018) and Vega (2019) are bundles: quality, quick answers, official marks, a larger index via semantic clusters, expert assessors, hyperlocal results. They are real releases. They are not single-factor “algorithms” in the Minusinsk sense.
YATI is the one Yandex itself ranks next to MatrixNet. Announced on 25 November 2020 at YaC, it is a transformer scored at search latency. Independent SERP tracking in the Russian market saw the break earlier, around late September; the table uses the announcement date and says so. Y1 (10 June 2021) is the product layer on top: quick answers above organic and ads when the page actually answers, video opened at the relevant moment, review summaries for venues. In the webmaster post, Yandex staff also pointed at Proxima, the commercial quality metric whose job is to order relevant pages by the chance of a completed task and a satisfied user. Read that carefully. It is not a published “who pays ranks higher” switch. It is an admission that commercial ranking is scored on outcomes, and outcomes in Yandex’s ecosystem often run through its own marketplaces. CS YATI (20 September 2022) is the specialist cousin for programming queries. Y2 (press reports 28 November 2022; the webmaster write-up is the same release) packages a year of that work.
The generative layer does not replace the table above. Neuro inside ordinary Search is 3 October 2024 (company news). Alice takes that surface on 22 May 2025, with reasoning, writing and image generation (22 May note). Alice AI, on 28 October 2025, is the next model generation (Alice AI announcement). Yandex has been explicit that Neuro’s sources are chosen by the existing search algorithms. Ranking quality still decides who is eligible to be summarised. For a working view of how thin the public measurement still is, see the Yandex Webmaster recap on the Alice visibility metric and the note on AliceAI-Foundation-80B.
What this page will not pretend to be is a complete anti-spam log. AGS, behavioural-factor penalties, adult filtering, affiliation grouping and the “low-quality business” demotion run beside these named releases. Yandex documents many of them in Webmaster help, without giving each one a city and a single day. Pre-2007 changes are product history, not a public ranking diary. Between branded releases the formula moves quietly. That is the gap, and it is a real one.
Same kind of record for Baidu and Naver. Last reviewed 3 October 2026.