GEO for Runet 2026: how to get into the answers of Alice, Neuro and GigaChat

A second layer of search results has appeared in RuNet, which is not visible in Ya.Metrica and is almost invisible in positions. A person asks Alice or opens Neuro, receives a ready-made answer of three paragraphs and a link to the source below it. For some requests, he will no longer go to the site. The question for 2026 is not “how to rise to first place,” but “how to become the paragraph that AI pulls out in response.” This is GEO, and it is configured differently for Russian AI search than for Western ones.

The article is for a practitioner who manages the site’s content himself, and not for a ranking researcher. No analysis of the architecture of Transformers. Only what changes the chance of getting a reply from Neuro, Alice and GigaChat, and how to check it manually.

I’ve been in digital for 9 years, for the last year and a half I’ve been heavily involved in AI search and testing hypotheses directly on this blog. The numbers “we were cited N times” will not be here - there are no such counters in RuNet yet. There will be observations, working principles, and what actually plays out from page to page.

GEO is not the “new SEO.” Classic SEO fights for page position. GEO fights for one paragraph within it. If you don't have an extractable passage, you're not optimizing for AI—you're just hoping.

1. How GEO differs from classic SEO

The main substitution of concepts is to think that GEO is SEO with a different name. The unit of measurement is different. In SEO, I fight for URL position by key. In GEO, I fight for the model to pull out a specific passage and put me as the source under the generated answer.

This leads to something unpleasant. You can be in the top 3 organics and never get into Neuro - because the text is not broken into self-sufficient pieces. And vice versa: the page in position five is cited in the answer because it has a clean definition and a table with numbers. Position and citation are related, but they are not the same thing.

The base remains. Neuro and Alice collect the answer from the top Yandex results. If the page is not in the index or it is not ranked for the query, there is nothing to cite. Therefore GEO does not cancel basic SEO, and is built on top. First get into the index and the top, then make the text extractable.

2. How does AI search work in RuNet?

Three sources, three different mechanics, and each is optimized in its own way.

• Yandex Neuro - a generative answer right in the search results. He collects it from the top organics plus his index, and puts source cards under the answer. The main channel for RuNet.
• Alice is the voice and text in Yandex search, works on the same source logic as Neuro. The wording of questions is often conversational.
• GigaChat and YandexGPT - in search mode, pull up fresh pages through their retriever. This is already about domestic neural networks and their issuance in responses.

The Western circuit works similarly, but this is not your case for RuNet. ChatGPT and Perplexity are pulled from their index, and optimized for them separately - I have an analysis of this in my article about AI Overview and Perplexity. You cannot mix recipes: the systems have different sources.

3. What AI takes in response and what it ignores

Over a year and a half of testing on different pages, a stable set of formats emerged that the model most readily pulls out. Not magic, but simple logic: the parser needs a piece that is read outside the context of the page.

What increases the chance of citation:

• Direct answer in the first paragraph under the question heading. Not a three-sentence rush, but straight to the point.
• Definitions of the "X is Y" format. The model loves a clear definition; it is convenient to insert it into the answer.
• Tables with numbers. When comparisons or specific values ​​are needed, a table is a prime candidate for citation.
• Numbered steps and lists. The request “how to do X” almost always pulls up a step-by-step block.
• FAQ section. The ready-made question-answer pair fits into the generative answer almost verbatim.
• Fact with source and date. “According to Y. Webmaster for May 2026” is quoted more readily than a bare statement.

What the model wastes: watery introductions “in this article we will look at,” paragraphs without a single supporting figure, abstract reasoning without definition, text where the answer is spread out on half a page. If a person needs to read the entire section to understand the idea, the passage is not retrievable.

4. Page structure for extraction

I rearranged several articles on this blog under one scheme, and it gave the most noticeable effect. I put the title in the format of a question or a short thesis. The first sentence below it is a ready-made answer that can be torn out and pasted into the search results without edits. Next - disclosure, numbers, table.

The principle is simple: each section should begin as if it is the only thing a person will read. Not “prelude - development - conclusion”, but “conclusion - proof - details”. An inverted pyramid, like on the news. The AI ​​pulls out the first paragraph, and the person who finished reading gets the rest.

One paragraph - one thought. When there are three topics in a paragraph, the model doesn't know which part to quote and doesn't take anything. I chop it into pieces. A short sentence works too. The long one follows so that the rhythm is not mechanical. By the way, the naturalness of the text here is important not only for people: about how Yandex and Google detect AI content, I have a separate parsing, and the text re-optimized for extraction is easily rolled into a robot.

5. Schema markup and technical base

Honestly: no one has publicly shown a direct confirmed boost from schema in Neuro. Therefore, I put it not as a “magic button”, but as a hygiene feature. FAQPage, HowTo and Article describe on the page where the question is, where the answer is, where the steps are - there is less work for the parser to guess.

What I put on the information pages:

• FAQPage - a question and answer for each block. Coincides with how AI searches for “question → answer” pairs.
• HowTo - step-by-step instructions with explicit steps.
• Article with publication date and author - a signal of freshness and authorship.

The technical basis behind this is ordinary: the page is in the Yandex index, open to robots, loads quickly, pure HTML without rendering the response through heavy JS. If the content is collected on the client and the retriever sees a blank page, there is nothing to quote, no markup can save it.

6. Semantics for questions, not for keys

I collect keys for classic SEO. I collect questions under GEO. These are different entities. “Apartment price Moscow” is the key. “How much does a one-room apartment cost in Moscow in 2026” is a question, and it is in this form that it is asked to Alice in her voice.

Therefore semantic core For AI search, I add a layer of questions: what, how, why, how much, what’s different, what to choose. Long conversational language rather than short commercial keys. Each question has a separate heading and a direct answer. Essentially, the core turns into a “question → extractable paragraph” map.

A convenient side effect: the text for the questions is better readable by humans. You don’t have to choose between “for search” and “for people”—colloquial wording works for both.

7. How to measure hits in AI responses

I'll disappoint you here. There is no dashboard “you have been cited 47 times” in RuNet. The metric shows traffic, but does not show whether your paragraph was included in Neuro. Therefore, the measurement is manual, and I treat it as a regular task, and not as a report from the system.

The working procedure that I use myself:

1. I take a list of 10–15 targeted questions per page.
2. Once every two weeks I run each one through Alice, Neuro and GigaChat.
3. I record in the table whether I got into the sources or not, which paragraph was pulled out, verbatim or paraphrased.
4. I look at the dynamics. After restructuring the text for extraction, the share of “hits” increases.

I also consider indirect signals. If the positions for an information request are stable, but organic traffic has dropped, some of the answers are taken by AI directly in the search results. This is not a reason to panic, it is a reason to check whether you are being quoted. Getting to the source under the answer is almost the same asset as a click, because the link and brand remain under the card.

8. Tables: factors and formats

The first table shows how the same factor weighs in classic SEO and GEO. The ratings are qualitative, according to my observations on this blog and projects, not from someone else’s rating.

The second table is content formats and the likelihood that the AI will pull them out in response. This is what I keep in mind when I'm laying out the page.

Conclusion

GEO does not cancel SEO or replace it with a buzzword. First, the page gets into the index and into the top of Yandex - this is the entrance. Then the text is divided into extractable passages - this is GEO. Without the first, the second is meaningless; without the second, the first gives the answer to a competitor.

I’m not selling a ready-made figure “we’ll grow by such and such a percentage in citations” - no one in RuNet honestly measures it yet. What is reproduced: a direct answer in the first paragraph, definitions, tables with numbers, FAQ and facts with a source increase the chance of getting into Neuro, Alice and GigaChat. I’m testing this on my blog and transferring it to projects. Then it’s a routine: send questions through the AI ​​once every two weeks and see who it quotes.

If you want to analyze your site specifically - which pages are already cited and which AI is skipping past - write to me at Telegram or leave a request via form. Starting consultation - 0 ₽: on it I will show 3 pages for restructuring and priority, where to start.