How to Get Cited by Russian AI Search: Alice AI, Yandex Neuro and GigaChat

Russia’s answer layer runs on different rails

The AI answer competing for your Russian-speaking buyer is not written by ChatGPT. It is written by Alice AI inside Yandex, which held roughly 73% of Russian search across all devices in March 2026 against Google’s 26%. In the second quarter of 2026, Alice AI’s quick answers reached 49.5 million monthly users, and short AI answers now appear on about 42% of queries.

That last number is the one to sit with. Nearly half of Russian search sessions now open with a machine-written paragraph, and the links under it are the new first page.

The mechanism differs from the one Western GEO advice assumes. ChatGPT and Perplexity retrieve from their own crawls and partner indexes, which is why a page ranking twelfth on Google can still be cited. Alice AI works the other way around: it builds its answer from pages Yandex has already found and ranked, then picks the fragments that best close the query. In Russia, generative visibility is not a parallel channel to search. It sits directly on top of it.

Two practical consequences follow. First, a Russian page that ranks nowhere is invisible to the answer layer, no matter how well written it is. Second, ranking alone no longer pays out, because the click that used to follow position three now often ends at the answer box. You need the position and the quotable paragraph.

v1be is a company that builds GEO and SEO focused AI growth software; our flagship engine, Conty, researches the live web and writes long-form articles that carry their citations with them. Everything below is the Russian-market version of that discipline, and most of it you can run yourself this week.

How Alice AI picks the sources it links to

Alice AI selects source fragments from pages that already rank for the query, evaluating them on meaning, structure and usefulness rather than on keyword density. It assembles a generative answer from more than 40 block types (links, media, weather, quotes) and regenerates it each time, which is why the source list under an answer shifts from week to week.

Read that regeneration behaviour carefully, because it changes how you measure. A citation is not a position you win and hold. It is a probability you raise. Checking one query once tells you almost nothing; the same query asked next Tuesday can return a different set of sources with your page in or out of it.

Yandex’s own documentation is unusually direct about what raises that probability:

• Publish expert, useful, original and substantive material, the phrase used verbatim in the Webmaster help.
• Improve your position on the query through ordinary optimisation, since the tool only tracks queries where the site already ranks high.
• Extend coverage to the material relevant to users’ primary and secondary intent, not just the head query.

Practitioners add one field observation that matches the mechanism: sites in the top five positions supply the bulk of cited fragments, and results from position six down are pulled in far less often. Treat the top five as the entry ticket, and the quality of your paragraphs as what decides whether the ticket gets used.

Share of Voice: GEO finally has a meter

On 7 April 2026, Yandex added a «Site visibility in Alice AI» section to Yandex Webmaster, which reports how often your site appears among the sources in Alice AI answers. Its core metric, Share of Voice (SoV), is the ratio of queries mentioning your site to the total number of queries that produced an Alice AI answer, expressed as a percentage. The report covers the last three months, updates weekly, and shows example queries plus a list of other sites acting as sources in your topic.

This is a bigger deal than the release note suggests. Everywhere else, GEO measurement is inference: you run a panel of prompts, log which brands come back, and estimate. Russia is the first major market where the engine itself hands you the number.

Three ways to use it that go beyond watching a line move:

1. Set a baseline before you change anything. Note this week’s SoV and the query examples attached to it. Without the baseline, a month of work produces an opinion instead of a delta.
2. Read the neighbouring sites list as a citation gap report. Those are the domains the model reaches for in your topic. Open the queries where they appear and you are looking at the specific answers you failed to close.
3. Segment by intent, not by page. Group the sample queries into “what is”, “how to”, “which is better”, “how much”. SoV usually collapses in one of those groups, and that group is your next content sprint, not a rewrite of everything.

One caveat the documentation states plainly: the tool only observes queries where your site already ranks well. A flat SoV can therefore mean your paragraphs are unquotable, or simply that you are not ranking on the queries that trigger AI answers. Check position data before you rewrite a single page. If you want the general version of this measurement problem, our guide to measuring AI brand visibility covers the prompt-panel method used in markets without an official metric.

GigaChat: the second door into Russian AI answers

GigaChat, Sber’s model, is the other surface worth optimising for, and it behaves less like a search engine and more like a researcher. By default it answers from what it knows; switch on its search mode and it goes to the live web, and its research mode (available since June 2025) goes deeper, visits more pages, and returns an answer with links to the material it used.

The retrieval path being different changes what gets picked up. A research-mode run reads several pages on a topic and synthesises across them, so it rewards depth on a single URL over a thin page multiplied by fifteen near-duplicates. Long, well-sectioned pages tend to survive that synthesis; landing pages built from slogans do not.

The Russian source base is also its own ecosystem. Where an English-language model reaches for Wikipedia, Reddit and trade press, Russian-language models lean on the platforms where Russian expertise actually lives: Habr, VC.ru, DZen, industry catalogues and marketplace listings. A brand with no presence on those platforms has no off-site corroboration, and corroboration is what turns a claim on your own domain into a fact the model is willing to repeat.

So the practical move is unglamorous. Publish the substantive version of your expertise on your own domain, then make sure at least two independent Russian-language sources say something consistent about you. That is the same “entity consistency” work behind getting cited anywhere, described in our guide on becoming a source for AI answer engines.

The technical floor: what actually blocks you

Nothing on the content side matters if the page is not cleanly crawlable by Yandex, and the failures here are boring and common: a robots.txt rule that blocks Yandex, pages that render only after JavaScript, slow mobile loads, and duplicate content spread across parameters.

Work through this floor before you touch the copy:

• Open the door in robots.txt. Yandex reads a file up to 500 KB, and a file that does not meet requirements makes the site count as open for indexing. Run it through the robots.txt analyser inside Yandex Webmaster rather than assuming.
• Register in Yandex Webmaster. Beyond the Alice AI report, this is where indexing problems become visible instead of theoretical.
• Turn on IndexNow. Yandex and Bing both support the protocol, so a published or updated URL can be reported in the same minute instead of waiting for the next crawl. For a content programme shipping weekly, that is the difference between competing on this month’s query and next quarter’s.
• Serve content in the HTML. If the main text arrives via client-side rendering, treat it as content that may never be read.
• Kill the duplicates. Parameter variants, print versions and near-identical landing pages split the signal and give the model three mediocre candidates instead of one strong one.
• Add a Yandex Business profile if you have a physical location, so contact details, hours and category resolve to a real entity.

Write a fragment a machine can lift

The unit of GEO is not the article. It is the paragraph that survives being cut out of it.

The quotable unit is a two to four sentence block that answers one question and stays true with zero surrounding context. Alice AI lifts fragments, not pages, so the page’s job is to offer a dozen clean fragments rather than one long argument that only works read end to end.

Four properties separate a liftable paragraph from filler:

• It repeats the subject. “The service costs from 4 900 ₽ per month” survives extraction; “It costs from 4 900 ₽” does not, because the pronoun loses its antecedent the moment the sentence is moved.
• It carries a number, a date or a name. Concrete tokens are what a model can check against other sources, and checkable claims get repeated.
• It sits directly under the heading it answers. The first paragraph after an H2 is the highest-value real estate on a Russian page, because it is where the extractor looks first.
• It is written in Russian, by someone who thinks in Russian. Machine-translated copy reads as generic to the model and as foreign to the buyer, and both readings cost you.

Freshness deserves its own line. Russian practitioners consistently report that AI answers favour recently updated material, and Yandex’s guidance points the same direction. An article with 2024 pricing is not merely stale; it teaches the model a wrong fact about you, and it will keep repeating that fact until the page changes.

The same rules govern the FAQ block. Write the questions the way people actually type them into Yandex (“сколько стоит”, “как подключить”, “чем отличается”), and answer each in a self-contained paragraph. Our practical GEO checklist covers the answer-first structure in more depth.

What does not work

There is no markup, meta tag or hidden prompt that gets a site into Alice AI answers, and every “neural network optimisation” package built on that premise is selling a mechanism that does not exist. Yandex states outright that no direct manipulation of inclusion is possible.

Four more approaches that waste a quarter:

• Volume without expertise. Fifty machine-written articles on the same cluster produce fifty pages with nothing to extract. The model has no reason to prefer any of them.
• Corporate abstraction. “Solutions for business of any scale” answers nothing. Prices, timelines, capacities and named limitations do.
• Keyword-first writing. The evaluation runs on meaning, structure and usefulness. Density optimisation is aimed at an algorithm that stopped mattering.
• Copying the top result. Reproducing what already exists gives the model a second copy of an answer it already has. Original data (your own numbers, your own case) is the only thing it cannot get elsewhere.

A 30-day plan for the Russian answer layer

Four weeks, one measurable delta: your Share of Voice against last month’s baseline.

Start by fixing measurement, then supply, then structure. In order:

1. Days 1 to 3. Connect Yandex Webmaster, open «Site visibility in Alice AI», record the current SoV and export the sample queries. This is the baseline you will be judged against.
2. Days 4 to 7. Clear the technical floor: robots.txt through the analyser, IndexNow enabled, mobile speed checked, duplicates canonicalised.
3. Days 8 to 14. Take the 20 queries where you rank in the top ten but hold no citation. Rewrite the first paragraph under each relevant heading into a self-contained answer with a number in it.
4. Days 15 to 21. Publish two new pages against the intent group where SoV is weakest, at real depth, with original data on the page.
5. Days 22 to 26. Fix the entity: refresh your Yandex Business profile, update the platform profiles that carry your description, and make sure the facts about your company match across all of them.
6. Days 27 to 30. Re-read SoV, compare against the baseline, and list which of the 20 queries now show your domain. That list is the brief for month two.

One month will not move a domain from invisible to dominant. It will tell you which of the three constraints (ranking, technical access, quotability) is actually binding for you, which is the answer most brands are missing when they start paying for “AI promotion”.

The Russian-speaking market is one of the few places where the answer engine publishes your score. That makes it the cheapest place in the world to learn what generative visibility actually costs, and the most embarrassing one to be absent from, because the number is right there in the panel. Conty writes this class of article, in Russian, with the sources attached and a human approving before anything publishes.