GEO optimization: what it is and how to get into AI answers
GEO is the work of getting an AI assistant to name you as a source. How it differs from SEO, why an answer holds one to three slots, where to start.
GEO (generative engine optimization) is the work of getting an AI assistant to take a fact from your page and name you as the source when someone asks ChatGPT, Claude, Perplexity or Alice. SEO (search engine optimization) is about your position in the list of links in Google or Yandex. The difference is the number of slots: search has about ten on the first page, and the eighth still sends people your way, while an AI answer fits one to three sources and has no eighth slot at all.
What to do first: open your-site.com/robots.txt and make sure OAI-SearchBot, Claude-SearchBot, PerplexityBot and YandexBot are not blocked. Those are the crawlers pages travel through into answers. If even one of them is blocked, writing articles is pointless for now: the door is locked. The check takes five minutes.
GEO does not replace SEO. They share a foundation: a site a crawler can read, with your own facts on it. That is why you do both at once instead of one after the other.
The rules and timings below were verified on 10 August 2026 against the documentation of Google, Yandex, OpenAI, Anthropic and Perplexity, with a link next to every number. Conditions change quietly, so keep the verification date in mind.
How does GEO actually differ from SEO?
The first difference is the unit of result. In SEO you compete for a position, and a position can be anything from first to tenth. In GEO the result is binary: your fact is either inside the answer paragraph or it is not.
The second is what gets taken. A search engine evaluates the whole page: title, structure, links, load speed. A model pulls out a fragment it can paste into an answer: a specific number, a deadline, a sequence of steps. General reasoning does not get taken, because the model already has it.
The third is where the person ends up. From search they click through to you. From an AI answer they usually do not: they got the answer and closed the tab. What you get is a mention of your name, not a visit, so name yourself in the text directly instead of hiding behind an impersonal "experts recommend".
Why does an AI answer have one to three slots while search has ten?
Because they are different output formats. Search hands you a list of links and you pick. A model hands you a finished paragraph, and ten sources do not fit into five sentences without turning back into a list.
Here is where GEO breaks for most people. In SEO an average result still counts: eighth place brings less than first, but it brings something. In GEO there is no average result, so competing with generic wording is useless: on a generic question the model answers by itself or names someone far better known. The slot goes to what the model does not have: your case and your number.
Google says the same from its side in its guide to optimizing for generative AI search: content people find unique, compelling and useful influences your presence in AI answers more than anything else, while chopping text into tiny pieces and adding AI-specific files such as llms.txt are not needed for its results.
Where do you start so the first month is not wasted?
Order matters more than speed here: every step is pointless until the previous one is done. Five steps, and the first two fit into one evening.
- Open the door to crawlers. They fall into three groups. One surfaces your site inside answers: OAI-SearchBot, Claude-SearchBot, PerplexityBot, YandexBot. Another trains on content: GPTBot, ClaudeBot, Google-Extended. The third arrives on a specific person's request: ChatGPT-User, Claude-User, Perplexity-User. Blocking the first group is an emergency, blocking the second is your call. Which names you must not touch and how it looks in the file: robots.txt for AI crawlers.
- Check that the site is readable at all. Is the text visible without JavaScript, is there exactly one h1 per page, is there a sitemap.xml, are Yandex Webmaster and Google Search Console set up. The whole checklist as a single prompt: how to check in five minutes whether AI assistants can see your site.
- Collect real queries instead of inventing them. Volumes come from a keyword tool (in the Russian-speaking market that is Wordstat, free with a Yandex account), direction comes from Google Trends. One thing you must not do: ask a model for search volume. It will produce a plausible number that does not exist, and your publishing plan will follow it off a cliff. How to collect and sort by intent: how to find out what your customers search for.
- Write something worth quoting. The customer's question as the headline, the answer in the first two or three sentences with no run-up, your own number with its period and its source, a questions block at the bottom. A walkthrough with sample paragraphs: how to write an article that AI assistants cite.
- Publish on your own domain first. Your page on your own site, and only then a rebuild for external platforms. Why that order and what gets rejected there: where to publish articles.
This article follows the same rules: questions as headings, the answer in the first paragraph, numbers with a link to the source, questions and answers at the bottom. It takes a minute to verify.
What is different for Yandex and Alice?
English-language GEO guides skip the Russian-language layer entirely, and it decides half the outcome for anyone selling in that market.
YandexBot sits in the same group as OAI-SearchBot and PerplexityBot: it is responsible for surfacing pages, and Alice answers on top of Yandex search. It sometimes gets blocked in robots.txt to reduce server load, which costs search positions and AI answers in one move.
Yandex Webmaster is not a formality, that is where the sitemap goes. Per Yandex documentation the crawler loads an added file within two weeks; the refresh button shortens that to three days, works up to 10 times per host, and the next attempt unlocks 30 days after the first. Google works differently: repeat recrawl requests for the same URL speed up nothing, and a sitemap is not required at all if the site is around 500 pages or fewer and every page is linked from inside the site.
External platforms (Habr, VC and Dzen in that market) work as mentions rather than link weight: outbound links in Habr and VC articles carry the nofollow attribute, but your project name and your figures stay in the text, and models read text. Posting a copy of your own article there is a bad idea: VC bans it in clause 10 of its rules, which forbids publishing new copies of previously published material without approval, and Habr warns that material below 75 percent uniqueness will most likely fail moderation.
Why doesn't this replace ordinary SEO?
Models pull sources from the same index that search crawlers fill. A page blocked in robots.txt or invisible without JavaScript does not exist for the results list or for the answer.
An example from our own site. Until 31 July 2026 the academy's English pages sat in the sitemap only as hreflang alternates, with no entries of their own. Crawlers that read only <loc> tags never saw the English tree, and no amount of writing would have fixed that. After the fix the map holds 184 addresses, 80 of them English: academy.agineai.com/sitemap.xml opens without a password, count them yourself.
The reverse is only partly true. Classic SEO can push a page up with volume and links, and a model does not react to that: it needs an extractable fact. GEO is a layer on top, and it makes sense to start it once the technical side is in order.
How long before it starts working?
Crawling a new page takes Google from a few days to a few weeks. With Yandex, indexed pages appear in search within two weeks.
After that comes the part where a nice-sounding deadline usually gets quoted. There is no verifiable data on how many weeks it takes before you get cited inside answers: neither Google nor OpenAI nor Perplexity publishes that statistic. "Two months and you are in the answers" is somebody's impression sold as a norm, and we are not going to name our own number either. You measure it yourself: once a month, the same question in ChatGPT, Claude, Perplexity and Alice, and a note of whose names came up. You do not switch the ads off during that time.
What does it cost if you do it yourself?
Zero, plus one evening of setup and one evening a week per article after that. The keyword tools, Yandex Webmaster and Google Search Console are free, and you most likely already pay for an AI subscription.
In the Russian-speaking web GEO is written about mostly by agencies that sell the work on a monthly retainer. They do exactly what is described above, and they do it properly. The question is whether you need a contractor at the stage where the whole job is "check robots.txt, collect forty queries, write four articles a month". Once you are publishing twenty pieces a month, that conversation changes.
One thing not to do at any budget: generating articles in bulk. Google calls this scaled content abuse and warns in its spam policies that a violating site may rank lower or not appear in results at all, and the measure can reach wider than a single page.
A detailed walkthrough with ready-made prompts (site check, query collection, the interview your article gets built from) is in the Echo block of the Academy, three lessons open for free: academy.agineai.com/block/b12-eho. The course is taught in Russian.
Questions
They are the same work under three names: getting into an AI answer as a named source. GEO stands for generative engine optimization, AEO for answer engine optimization, LLMO for large language model optimization. The acronym you pick changes nothing about what you do.
With a site. Models assemble answers from web pages, while social networks and messengers are visible to crawlers only partly or not at all. Until a page exists at a public address without a password, there is nothing to cite. External platforms work as an addition to your own domain, not a substitute: the ranking weight stays with the platform and you get the mention.
Open ChatGPT, Claude, Perplexity or Alice and ask the way your customer would: “recommend a [your kind of business] in [city], give me a few options and explain the choice”. See whose names come up. Repeat monthly with the exact same wording, otherwise there is nothing to compare against.
That is your call and it is not a fault either way. Just do not mix up the groups: GPTBot, ClaudeBot and Google-Extended train on content, while the ones that surface you inside answers are OAI-SearchBot, Claude-SearchBot, PerplexityBot and YandexBot. Block the second group along with the first and you disappear from answers.
The ones that crawl on their own do, and they follow it. The separate case is fetchers launched by a specific person's question: OpenAI writes that robots.txt rules may not apply to ChatGPT-User, Perplexity puts it more bluntly and says Perplexity-User generally ignores them, while Anthropic states that its bots, Claude-User included, honour robots.txt directives.
A workable rhythm is one article a week on topics from your collected query list. The first five topics cover a month. Regularity beats one-off volume: a single topic taken all the way to a verifiable number of your own is worth more than ten empty texts pushed out in a day.
Crawling a new page takes Google from a few days to a few weeks, and with Yandex indexed pages appear in search within two weeks; both numbers come from the official documentation. For the time until your first citation inside an AI answer there is no verifiable data: neither Google nor OpenAI nor Perplexity publishes that statistic, so we are not naming a number of our own either.