On April 7, 2026, Yandex launched in Webmaster the tool “Visibility of the site in Alice AI”. It shows how often the site appears in Alice’s answers among the sources. The data is collected over three months, updated once a week, and takes into account only the queries on which the site already ranks high in search. Yandex itself separately stresses that the tool does not affect ranking.
The last detail matters more than it seems. Neural-network answers in search rest on the same pages that search has already found and rated. So a conversation about GEO starts not with secret techniques, but with how the site itself is built. Below I will sort out how AI search chooses sources, what here depends on the site, what does not, and which pieces of advice can safely be skipped.
What GEO is, briefly, and why it is still about search
GEO (generative engine optimization) is work so that a site appears in the answers of generative search: in the short retellings from AI in Google, in AI Mode, in Alice’s answers and other assistants with search.
It sounds like a new discipline, but the official sources describe it more modestly. Google, in the documentation on AI features in search, writes that for showing in short retellings from AI and in AI Mode there are no additional requirements: the page must be indexed and suitable for showing in the ordinary results with a snippet. Special files or special markup are not needed. In the guide on optimization for generative AI, Google repeats the same thought: the SEO recommendations are still applicable.
Yandex, judging by the new tool, looks at the same picture from the other side: visibility in Alice is counted only for queries where the site is already shown well in search. That is, the foundation is the same.
How AI search chooses whom to cite
If you simplify it, generative search works in two steps. First the system finds pages that fit the query — in essence, by ordinary search. Then the model reads what was found and formulates an answer, pointing to some of the pages as sources.
Two things follow from this. The first: a page that is not in the index will not get into the answer. The second: at the formulation stage it is easier for the model to rely on a page where the answer to the question is stated clearly, and the facts are easy to find and check.
Google in the documentation describes one more detail: short retellings from AI and AI Mode can use the query fan-out technique, that is, run several related searches on subtopics and different data sources in order to assemble an answer. According to Google, that is why such an answer can contain a wider and more varied set of links than the classic results.
For a site this means a simple thing. A person asks a long question, the system breaks it into several short ones, and the source can be a page that answers well only one of the parts. You do not have to be “the main page on the topic”: it is enough to answer your own concrete question clearly and fully.
Search engines do not publish the exact rules by which the model chooses whom to cite. Everything that can be done from the site’s side is to remove obstacles on the first step and make the content understandable on the second.
What depends on how the site is built
Robot access and indexing
The basic condition: the search engine’s robot must have access to the page, and the page must be in the index. Google directly lists which ordinary tools control showing in AI features: robots.txt, noindex, nosnippet, data-nosnippet, max-snippet. If a page is closed from indexing or from showing a snippet, it will not be in the short retellings either.
Separately there is the Google-Extended token: it controls whether the site’s materials are used to train Google’s models. According to Google’s documentation, it does not affect showing in search. These settings should not be mixed up: a ban on training and a ban on showing are different things.
Content without JavaScript
Google can process JavaScript, but in the optimization guide it notes that SEO for sites on JS frameworks is harder. Other systems’ capabilities may be different. It is more reliable when the main text of the page, the headings and the links are in the HTML right away, and do not appear only after scripts run.
This is checked in the page’s source code: if the key paragraph is not in it, the robot will have to run scripts to see it.
HTML structure
One H1 heading, a logical hierarchy of H2 and H3, paragraphs, lists and tables where they are needed by meaning. This is ordinary careful markup, without any “layout for AI”. It helps both people and search engines understand what the page is about and where the answer to a concrete question is.
A useful habit: at the start of a section give a short direct answer, and then the details. It is faster for a person that way, and it is simpler to pick the answer out of such a text.
One more detail of structure is tables and lists where the information really is tabular: a comparison of options, characteristics, steps. Such information is simpler to read both for a person and for a system that assembles an answer. Turning everything in a row into tables is not needed: a text with reasoning stays a text.
Author, dates and sources
When a page shows who wrote it, when it was published and updated, and where the facts came from, it is easier to assess. For articles that is the signature of the author or the company, the dates of publication and update, links to the original sources for numbers and statements. For service pages it is clear who provides the service and how to get in touch.
This is not a guarantee of being cited, but the absence of these details makes the page less reliable in the eyes of any reader, including algorithms.
Structured data
Structured data (schema.org) helps search engines understand more precisely the type of the page and the key facts: an article, a product, an organization, contacts. Google notes at the same time that special markup for AI features is not needed. Structured data is useful as part of ordinary SEO, but a separate “AI markup” does not exist.
| What is on the site | How it affects | How to check |
|---|---|---|
| Robot access, indexing | Without an index the page is not in the answers | robots.txt, Search Console, Webmaster |
| Snippet bans | They limit showing in AI features | The nosnippet and max-snippet meta tags |
| Text in HTML | The robot does not have to run scripts | The page’s source code |
| Headings and structure | It is simpler to find the answer to a question | One H1, an H2–H3 hierarchy |
| Author, dates, sources | The page is simpler to assess | Whether they are visible on the page |
| Structured data | The page type and the facts are understood more precisely | Markup checking tools |
What does not depend on the site
A lot in generative search does not depend on the site owner. Which model the search engine uses, how it formulates the answer, how many sources it shows and in what order — all of that is decided by the platform. How often an answer from a neural network is shown for your queries at all is also decided by it.
What does not depend on the site either is that the user could get an answer without following a link. That is part of how generative search is built, not an optimization mistake.
So promises to “get into a neural network’s answer for sure” are worth taking skeptically. From the site’s side it is possible to make it so that nothing prevents it from being a source. The decision is still made by the system.
Myths: llms.txt, “layout for AI”, AI markup
A lot of advice quickly appeared around GEO. Google directly refutes part of it in its guide on optimization for generative AI.
llms.txt. A file that supposedly helps neural networks read the site. Google writes that it is not needed for search and can be ignored.
“Layout for AI”. Advice to break text into short pieces so it is easier for models to “slice” it. According to Google, there is no need to specially split content into fragments. An ordinary understandable structure is enough.
AI markup. According to Google’s documentation, special markup for AI features is not required. Standard structured data stays useful, but there are no new mandatory types for AI search.
The general conclusion from the myths is simple: if a piece of advice requires doing something “only for neural networks” and does not make the site better for people, it is worth treating it with caution.
How to check whether neural networks see the site
The first is an ordinary check of indexing. In Google Search Console and Yandex Webmaster you can see which pages are in the index and whether there are problems with access.
The second is special reports. In Yandex Webmaster there is the tool “Visibility of the site in Alice AI” with a visibility-share metric. To use it, the site must be added and confirmed in Webmaster. Google in the guide mentions a report on performance in generative AI in Search Console.
The third is a manual check: an assistant with search is asked the questions your pages answer, and you look at whom it cites. This is not statistics, but it shows well which sites and pages become sources in your topic and how they differ from yours.
In a manual check it is useful to write down not only “yes or no”, but which exact pages get into the sources. Often it is not the home page and not the service page, but a detailed article or a reference section. That is a hint about which materials the site is missing.
The fourth is a technical check of the pages: the source code, the headings, snippet bans, structured data. If the site is being built or rebuilt, these things are simpler to lay in at once when developing a corporate website or a landing page than to fix later.
If work on the site’s visibility in AI search as a whole is needed, more about GEO is on the site of our agency «Горячий ветер»: hot-wind.com/geo.