More and more people no longer scroll through search results — they ask an AI assistant: “recommend a dental clinic near me,” “which company should I choose for a renovation,” “how much does air conditioner installation cost.” In reply, Alice (Yandex's assistant), ChatGPT or GigaChat name two or three companies — and the customer goes to them. If your company is not in that answer, for that customer it does not exist.
Making sure AI assistants know and recommend your business is the job of GEO (Generative Engine Optimization). This guide covers how AI assistants decide whom to name, 10 steps that raise your chances of getting into the answers, the specifics of different assistants, common mistakes and ways to measure the result. All of it is based on our SEO practice since 2010 and brand visibility measurements in PromoLab AI Radar.
To get into AI answers, a company needs to become a source the model trusts. That means the website must be open to AI bots, the facts about the company must be unambiguous and confirmed on third-party platforms, and the site must have pages that directly answer customers' questions. The rest is regular measurement: whom the AI assistants name today and what changes after each improvement.
How AI assistants decide whom to recommend
An AI assistant has two sources of knowledge. The first is the data it was trained on: the model's “memory,” a snapshot of the internet as of some date. The second is search at the moment of answering: the assistant finds fresh pages, reads them and assembles an answer with links to sources. Most popular assistants can do both, and the answer is often a mix of the two.
- 1Question“Recommend a reliable dental clinic in Moscow”
- 2Searchfinds fresh pages or relies on training data
- 3Selectionkeeps the sources it trusts
- 4Cross-checkcompares facts across sources
- 5Answernames 2–3 companies and gives links
This leads to the main rule of GEO: the AI recommends not whoever “optimized” their site best, but whoever has clear, consistent and confirmed information in the sources the model considers reliable. Six things raise your chances of getting into the answer:
AI bots can open and read the pages: they are not blocked in robots.txt, and the text is not hidden in images and scripts.
Facts about the company are easy to extract: services, prices, addresses, timelines, structured data, llms.txt.
There are pages that directly answer customers' questions: “how much does it cost,” “how to choose,” “what is the difference.”
The company is mentioned on maps, in directories, rankings and industry media — independent sources confirm the facts.
Reviews and ratings on platforms: AI assistants use them to judge whether a company can be recommended.
Current prices, update dates, new content: all else being equal, models prefer fresh data.
Nobody can guarantee a place in an AI answer: answers change from query to query and from model to model. The job of GEO is to systematically raise the probability that you get named, and to measure how often it happens.
GEO and SEO: what is the difference
SEO fights for a site's position in a list of links; GEO fights to have the AI name the company in its answer and link to its site. But these are not competing approaches: assistants with search rely on pages that rank well, so GEO is built on an SEO foundation. For more on the differences, see the article “How to get visibility in AI with GEO”.
| GEO | SEO | |
|---|---|---|
| Where the result shows | in the answers of Alice, ChatGPT, GigaChat and other assistants | in Yandex and Google search results |
| What we optimize | the website and mentions of the company across the internet | mostly the website and links to it |
| Main signals | confirmed facts, direct answers, mentions, reviews | relevance, technical health, links, user behavior |
| How we measure | share of mentions, position in the answer, links, AI traffic | rankings, traffic, enquiries |
| When it kicks in | first changes after 1–2 months, sustainable after 3–6 | first changes usually after the second month |
How to get into AI answers: 10 steps
The order of the steps matters: without a “before” measurement you cannot prove the result, and without a technical foundation content and publications work at half strength.
Step 1. Collect the questions customers ask AI assistants
GEO starts not with keywords but with questions. People write to AI assistants at length and in plain language: “recommend,” “which company is better,” “how much does it cost,” “how to choose,” “what to do if.” Sources of questions:
- calls and correspondence with customers — what they ask before buying;
- search queries from Wordstat and Yandex Webmaster — especially long, question-type ones;
- reviews of you and your competitors — which selection criteria keep coming up;
- search suggestions and “related questions” blocks in the results.
To start, 30–50 questions closest to the purchase are enough. Group them by intent: recommendation (“recommend”), comparison (“who is better”), price (“how much does it cost”), choice (“how to choose”) and a specific task (“where to get it done urgently”).
Step 2. Measure whom the AI assistants name today
Ask each question in several AI assistants — for example, Alice, ChatGPT, GigaChat, DeepSeek and Perplexity — and record which companies are named, in what order and which sources are linked. Ask the questions in a new chat with no history and repeat them: answers change from one run to the next.
This measurement will show three things: whether you are in the answers, who your real competitors in AI assistants are (they often differ from your competitors in search) and which sites the AI assistants cite in your niche. That last list is a ready-made plan of where you need to appear. Doing this manually across dozens of questions takes days, so we do it in PromoLab AI Radar: the service regularly asks AI assistants the questions and shows whether you are mentioned, whether you are recommended and whom they name more often.
Step 3. Open the site to AI bots
An AI assistant will not recommend what it cannot read. Check your robots.txt: it must not block search engine crawlers or AI bots. Large AI companies run several bots with different jobs — at OpenAI, for example, these are GPTBot (collecting training data), OAI-SearchBot (search in ChatGPT) and ChatGPT-User (visits on a user's request). The minimum set of permissions looks like this:
User-agent: YandexBot
Allow: /
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: GPTBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: ClaudeBot
Allow: /Whether to allow the bots that collect training data is a business decision. But if you block the AI assistants' search bots, the assistants will not be able to cite your pages in answers with links. Besides robots.txt, check:
- bot protection — a firewall or CDN may block AI bots even if robots.txt allows everything;
- text in the HTML — prices, services and contacts must be text on the page, not only in images, PDFs or widgets loaded by a script;
- indexing — the pages are in Yandex and Google, sitemap.xml contains no junk, speed and the mobile version are in order.
You can quickly check access for AI bots, robots.txt, the sitemap, llms.txt and structured data in our free express check — it takes about 30 seconds.
Step 4. Make the facts about your company unambiguous
AI assistants cross-check data from different sources. If the site shows one phone number, the maps another, and a directory lists an old address, the model is unsure of the facts and will more readily name a competitor with clear data. Bring the following into line:
- company and brand name — spelled identically everywhere;
- addresses, phone numbers, opening hours;
- the list of services or product categories and prices — at least “from”;
- year founded, licenses, certificates, key specialists.
Check the website, your listings in Yandex Business, 2GIS (a Russian maps and directory service) and Google Business Profile, industry directories and social media. On the site, gather the facts on the “About us” page — as text, not images.
Step 5. Add Schema.org structured data
Structured data describes the page's data in a language search engines and AI assistants understand: who you are, what you sell, where you are located, what is written about you. The basic set for most businesses:
- Organization or LocalBusiness — name, address, phone, logo, links to the company's profiles;
- Service or Product with Offer — services and products with prices;
- FAQPage — question-and-answer blocks;
- Article with an author — for articles and expert content;
- BreadcrumbList — the structure of sections.
An example of organization markup: the sameAs field links the site to the company's profiles on other platforms and helps establish that it is the same company.
{
"@context": "https://schema.org",
"@type": "LocalBusiness",
"name": "Company name",
"url": "https://example.ru/",
"telephone": "+7 495 000-00-00",
"address": {
"@type": "PostalAddress",
"addressLocality": "Moscow",
"streetAddress": "1 Example St"
},
"sameAs": [
"https://yandex.ru/maps/org/...",
"https://2gis.ru/moscow/firm/..."
]
}Mark up only what is actually on the page. Ratings and reviews — only real ones: search engines penalize markup that does not match the content.
Step 6. Add an llms.txt file
llms.txt is a text file in the root of the site that briefly describes, in Markdown, what the company does and lists links to key pages. It is a proposed standard: not all AI assistants officially use it, so it will not be a magic button. But it is easy to make, and it helps language models quickly understand what the site is about.
# Company name
> In brief: what you do, for whom, in which city, since what year.
## Services
- [Service 1](https://example.ru/service-1): what is included, price from …
- [Service 2](https://example.ru/service-2): what is included, price from …
## About the company
- [Contacts](https://example.ru/contacts): address, phone, opening hoursStep 7. Write content that is easy to quote
An AI assistant assembles its answer from fragments. The easier it is to extract a precise answer from a page, the higher the chance it gets cited. A page that performs well in AI answers is built like this:
- Heading = the customer's question. The wording a person would type into an AI assistant.
- The answer in the first lines. One or two sentences that can be quoted in full: figures, timelines, terms.
- A table of facts. Prices, specifications, a comparison of options — structured data is the easiest for models to extract.
- Criteria and steps as a list. “How to choose,” “what is included,” “where to start” — in short bullet points.
- Questions and answers. Frequent questions with FAQPage markup.
- Author and date. Who the expert is and how fresh the information is — trust signals.
Which materials to create first:
- service pages with prices — at least a “from … to …” range and what makes up the cost;
- answers to questions about choosing — “how to choose,” “what to look for,” “how A differs from B”;
- comparisons and roundups — honest, with criteria, not promotional;
- case studies with figures — task, solution, result;
- expert content from specialists with their experience stated: for healthcare, finance and law this is especially important, as the E-E-A-T principles say.
Be specific. The phrase “an individual approach and affordable prices” gives an AI assistant nothing. The phrase “air conditioner installation — from 7,000 ₽, same-day” gives a fact that can be passed on to the user.
Step 8. Show up where AI assistants get their data
AI assistants are more willing to recommend what several independent sources confirm. So a website alone is not enough — the company needs mentions on the platforms that models read and cite:
- maps and directories — Yandex Business, 2GIS, Google Business Profile, industry directories;
- rankings and roundups — “best clinics in the district,” “top contractors”;
- industry media and blogs — expert articles, comments, interviews, research;
- review platforms — niche review sites and reviews on maps;
- marketplaces and aggregators — if you sell products or work through service aggregators.
The main source of ideas is the list of sites the AI assistants cited during the measurement in step two. If the same ranking keeps popping up in answers about clinics in your district, getting into it is more useful than getting into ten random directories.
Step 9. Work with reviews
Reviews are one of the sources AI assistants use to judge a company: when answering “whom to choose” questions, assistants with search often rely on maps and review sites. What helps: a regular flow of real reviews on the platforms where people read them, replies to reviews — especially negative ones — and recency: three-year-old reviews say less than fresh ones. Purchased reviews are a bad idea: platforms detect them, remove them and demote the listings.
Step 10. Measure and update
After a month, repeat the measurement on the same list of questions and compare it with the results from step two. See which questions you now appear for, which pages started getting cited and where competitors are still ahead. Update the plan based on this data: write answers to the questions where you are absent, update prices and dates, add new external sources. GEO is not a one-time setup but a “measure — improve — measure” cycle.
Collecting questions, baseline visibility measurement, access for AI bots, consistent facts about the company, structured data.
Answer pages, llms.txt, listings on maps and in directories, the first external publications. The first changes in AI answers typically come 1–2 months after technical preparation and publishing content.
Reviews, rankings, industry media, regular measurements and improvements. A sustainable effect comes after 3–6 months of regular work.
How to get into the answers of Alice, ChatGPT, GigaChat, DeepSeek and Perplexity
The basic steps are the same for all AI assistants, but their data sources differ. So visibility has to be checked in each one separately.
Answers with search are built on Yandex results. So for Alice the most important things are SEO in Yandex, indexing in Yandex Webmaster, a Yandex Business listing and reviews on the maps.
With search enabled it finds fresh pages and shows links; without search it answers from training data. Make sure the site is open to OAI-SearchBot and ChatGPT-User, and work on mentions on authoritative platforms.
Answers from training data and may rely on pages found online. Unambiguous facts about the company and mentions on authoritative Russian-language resources matter.
In search mode it searches the internet and shows sources; without it, it answers from training data. Check the answers in both modes.
Always searches the internet and shows sources. Handy for analysis: you immediately see which sites are cited in your niche. Access for PerplexityBot matters.
AI Overviews in Google rely on Google's search index. You need indexing, quality content and good rankings in Google.
Mistakes that keep AI assistants from seeing you
How to check whether AI assistants mention you
The simplest way is to ask. Put 5–10 questions your customer would ask to the AI assistants, in a new chat and without mentioning your company. Note whether you are in the answer, in what position, what is said about you and whether there is a link to your site. Then ask directly: “What do you know about [company name]?” — this shows which facts about you the model considers true.
For ongoing control, manual checks are not enough. We look at five indicators:
- Share of mentions — in what percentage of answers to your list of questions the company is named, compared with competitors;
- position and context — recommended first, mentioned in a list or only on a direct question;
- sentiment — what exactly the AI says about the company;
- links to the site — whether the assistant cites your pages as a source;
- AI traffic — visits to the site from AI assistants.
AI traffic is visible in Yandex Metrica: in the sources report, look for visits from chatgpt.com, perplexity.ai and other AI services. Links from ChatGPT usually carry the utm_source=chatgpt.com tag. Share of mentions, positions and cited sources across your list of questions are shown by PromoLab AI Radar.
Checklist: is your business ready for AI answers
Tick off what is already done — we will calculate your readiness. The ticks are saved only in your browser.
Tick the items that are already done.
How much time and money it will take
The first changes in AI answers typically come 1–2 months after technical preparation and publishing content; a sustainable effect comes after 3–6 months of regular work. Some of the steps — access for bots, consistent facts, structured data, llms.txt, map listings — you can do yourself. The hardest parts are measuring visibility systematically, creating content and getting mentions on external platforms.
We broke down what GEO costs on the market and what makes up the budget in the article “How much GEO costs in Russia”. You can estimate a budget for your business in the pricing calculator.
We will check for free how AI assistants answer your customers' questions, show you whom they name today and suggest where to start.
