Someone Scraped All 295,000 Answers in Yandex’s Neuro Library — Dzen Gets Cited 2.5x More Than Wikipedia

If you want to know whom Yandex’s AI actually trusts, a Russian SEO has just done the counting for you. In a vc.ru analysis published on 4 October, Sergey Paraev says he downloaded the whole of Neurum, Yandex’s public library of Alice/Neuro answers at ya.ru/neurum. That is 295,345 questions with their answers, tags and sources. His headline finding: Dzen is cited 2.5 times more often than Russian Wikipedia.

It is one practitioner’s scrape, not a Yandex disclosure. But it is the biggest public look yet at the source mix behind Yandex’s generative answers, and it doesn’t flatter brand websites.

What Neurum is

Yandex launched the Neuro question-and-answer library in September 2024 as a showcase of popular, anonymised questions answered by its AI. Paraev describes the pipeline as RAG. Yandex takes generalised user queries to Search with Alice, YandexGPT turns trending topics into questions, Alice finds sources and writes the answer, and the card gets navigation tags. One caveat runs through everything below: the questions are machine-generalised, not verbatim queries, so the data reflects both demand and the generator’s habits.

The numbers

  • Sources per answer: 5.2 on average, so the “five sources” pattern isn’t a hard rule. There are 1.54 million source links in total, across 138,647 unique sources.
  • Dzen first: 82,103 mentions, 1.7 times VKontakte and 2.5 times Russian Wikipedia.
  • UGC heavy: Dzen, VK, Otvety Mail.ru and Bolshoy Vopros together account for about 193,000 mentions. That is roughly 12.5% of all links, or one in eight citations pointing at user-written content.
  • Yandex cites Yandex: yandex.ru is eighth with 14,758 mentions (Market, the business directory and other services).
  • Academia punches above its weight: CyberLeninka ranks sixth, above YouTube. Educational sites together top 52,000 mentions. Habr and vc.ru both make the top 15.
  • Skewed categories: Science & Education, Other and Technology make up 57% of the base. “Other” alone holds 62,194 questions (21%). There is no Health category even though 5,110 cards carry a “Health” tag. Paraev suspects deliberate YMYL caution.

Why Dzen winning is not the self-preference story it looks like

The lazy reading is “Yandex boosts its own platform”. But Dzen hasn’t been Yandex’s since VK bought it in 2022. So Alice citing Dzen 82,000 times is Yandex leaning on a rival’s UGC platform because it is huge, fresh and easy to index. As Paraev puts it, expert articles and authorless rewrites sit side by side there. The genuine self-preference signal is the smaller one: yandex.ru in eighth place, with Yandex’s own services guaranteed visibility.

Compare that with Korea, where Naver keeps its UGC for itself and AI Briefing cites mostly Naver-owned pages. Russia’s answer engine cites the open UGC web. Korea’s cites a walled garden.

What it means for Yandex GEO

Paraev’s takeaways are blunt. Syndicate expertise to Dzen. Answer questions on Q&A platforms. Publish on Habr and vc.ru for B2B and IT. Write for specific long-tail questions. Use Neurum’s questions and tags as a free map of topic clusters. And check which domains keep appearing in your niche, because they are your real competitors in AI answers even if you never see them in classic results.

That squares with earlier evidence. A Habr experiment in July ran 60 generative Search API calls and found citations came overwhelmingly from the organic top. Brand sites lost out to media, blogs and aggregators writing about those brands. The same post cites Ashmanov & Partners’ finding that the share of non-top-10 sites in neuro answers fell from 40% to 10% in six months.

The scale is the reason to care. Alice quick answers reach 49.5 million monthly users. On 6 October TKS.ru reported that Yandex’s Alice apps ranked 13th by monthly audience among mobile gen-AI products in a16z’s latest Top 100, built on August Sensor Tower data. TKS says it was the only Russian entry in the top 15.

Our take

Treat this as a well-documented scrape with stated limits. One sub-finding about Turkey-related questions comes from a 140-question sample, and Paraev says so himself. Neurum is a curated showcase, not a log of live SERPs. Still, the direction is consistent with everything else we’ve seen. Yandex has just launched a course telling publishers to write original, useful material to get cited. Neurum suggests that original material had better also exist on Dzen, a Q&A site or Habr, not just on your own domain. In Yandex GEO, distribution decides who gets cited.