GEO in 2026: how to get cited by ChatGPT, Alice and AI Overviews

GEO, AEO and SEO: precise definitions and the fact most posts bury

SEO optimises for the classic results page: the goal is to sit higher and get the click. AEO, answer engine optimisation, targets direct answers: featured snippets, People Also Ask blocks, the spoken replies Alice gives. The goal there is to have your paragraph lifted as the ready answer. GEO, generative engine optimisation, targets engines that do not pick one source but synthesise an answer from several and attach citations. The goal is to be among those sources and to be named inside the answer text.

Now the fact most articles bury halfway down: GEO is an add-on to SEO, not an alternative. Yandex neural search does not maintain a separate index, it builds the answer on top of the regular results. Google AI Overviews draws its sources mostly from documents already ranking near the top for the query or its paraphrases. Which means something simple: a site that is not indexed, has no trust and does not reach the top 20 will not appear in a generative answer no matter how much neural optimisation is bought.

That dictates the order of work. Baseline SEO first: indexing, mobile speed, relevance, commercial factors, domain trust. Then the GEO layer: answer structure, factual density, structured data, crawler access, external mentions. A contractor offering GEO instead of SEO, or selling a GEO package for a site that is not even indexed, is selling air. The check takes a minute: if the site is not in the Yandex top 20 for its own commercial queries, GEO is not where to start.

So why bother with GEO at all. Because the share of queries resolved inside a generative interface keeps growing, and the choice of a contractor or a product is increasingly made from a short list of three to five names the model produced. Getting onto that list is the new visibility problem. It is solved by content structure and presence on sources the engines trust, not by budget size.

Platform by platform: where each engine takes its sources

A common error is treating optimising for AI as one uniform task. The engines are built differently and the work only partly overlaps. AI Overviews and Yandex Neuro sit on their own search indexes, so they inherit SEO results directly. ChatGPT and Perplexity use web search plus their own crawlers, which means they also depend on whether you let those crawlers through in robots.txt. GigaChat leans more heavily on large Russian-language sources.

The practical takeaway from the table below: the base layer is the same everywhere, be indexed, rank well, answer directly and structure the page. The details diverge from there. Western engines hinge on crawler access and mentions on authoritative sites; the Russian ones hinge on ordinary Yandex SEO quality and presence in sources the Russian-language web trusts.

• For a Russian audience the measurement priority is Alice and Neuro, GigaChat, ChatGPT and Perplexity, with AI Overviews added as Google share in your niche grows.
• The same article can be cited by Perplexity and ignored by Neuro. That is normal, the engines sample different slices of the web.

The on-page checklist that actually causes citation

A generative engine does not read the page end to end, it extracts passages. What gets cited is never a good article in the abstract but a specific paragraph that survives being lifted out of context. That leads to the core structural rule: every section opens with a direct answer in the first 40 to 60 words, and everything after that is elaboration, detail and caveats. Warm-up intros are not merely useless, they occupy the exact slot the model quotes from.

The second rule is question-shaped H2s. The model matches the query wording against document headings, so How much does SEO cost outperforms Pricing. The third is factual density: numbers, ranges, dates, units. Quickly does not get quoted, 2 to 3 months does. The fourth is tables, which engines lift especially readily because the structure is already done for them.

Here is what that looks like in practice. Before: In the modern world, website promotion plays a key role in business development. Many companies wonder how to increase their visibility in search engines. In this article we will examine the main aspects of this question in detail and give useful recommendations. Forty-odd words, zero facts, nothing worth quoting.

After: SEO in Yandex starts at ₽40,000 a month. The first position changes appear in 2 to 3 months and stable traffic growth in 4 to 6. Most of the weight in commercial results comes from commercial factors: prices, contacts, delivery terms and reviews. Thirty-eight words, four extractable facts, and a ready answer to two questions at once, how much and how long. That is the paragraph an engine quotes and attributes.

• A direct answer in the first 40 to 60 words under every H2, before any elaboration.
• H2s phrased as questions: how much, how to choose, which is better, how long it takes.
• Numbers, ranges and dates instead of adjectives: from ₽40,000, 2 to 3 months, 2026 data.
• At least one comparison table per article, the single most-cited format.
• An FAQ block with questions worded the way people actually ask them out loud.
• No in the modern world, in this article we will look at, or as everyone knows.
• Attribute numbers to a source and date them. Models prefer to cite what can be verified.

The technical layer: markup, crawlers and the truth about llms.txt

JSON-LD is the cheapest way to tell an engine what it is looking at. The minimum set: Organization sitewide with name, logo, contacts and profile links; Article on blog material with author and publication and update dates; FAQPage on question blocks; HowTo on step-by-step instructions; Service or Product on commercial pages with a price. Markup does not guarantee citation, but it removes ambiguity so the engine does not have to guess who wrote this and which company stands behind it.

Heading hierarchy matters more than it looks. One H1, sequential H2s with no skipped levels, H3s only inside their parent H2. That structure is exactly how an engine slices the document into passages. A site where headings were chosen by font size rather than meaning extracts badly: the boundaries between ideas blur and the model gets mush instead of a clean answer.

Crawler access is the item that most often breaks the entire effort. If robots.txt blocks GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot or Google-Extended, those engines physically cannot see the page no matter how well it is optimised. Check the file now: plenty of CMS defaults and content-protection setups block these bots out of the box. Separately confirm that YandexBot and Googlebot are unrestricted and that no server rule or CDN filter drops them by user agent.

Now llms.txt, where most Russian-language guides get it wrong. It is a community-proposed file at the site root that maps content for language models. It is optional and it is not a standard: Google has stated plainly that Search does not use it, and no other major engine has confirmed official support. Adding it is fine, it costs half an hour and does no harm, but selling llms.txt implementation as the foundation of GEO is not honest. The real technical gains come from crawler access, structured data and heading structure.

• Audit robots.txt for GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended and YandexBot.
• Implement Organization, Article, FAQPage, HowTo and Service or Product, then run them through a validator.
• One H1, sequential H2s and H3s chosen by meaning rather than font size.
• Publication and update dates in both the markup and the visible text, since recency affects source selection.
• llms.txt is optional, half an hour of work and no promises. Google Search ignores it.

Off-page: where engines form their opinion of your brand

Generative models lean on what is written about you elsewhere far more than classic search does. The reason is mechanical: your own site is an interested party, an outside mention is corroboration. Which means a company that exists only on its own domain will barely appear in generative answers, no matter how well its pages are structured.

In the Russian-language segment the heaviest weight sits with large, well-indexed, high-trust platforms: vc.ru, Habr, trade media, vertical publications and directories. Models cite those domains noticeably more than average, so one substantial piece on vc.ru often does more for AI visibility than ten posts on your own blog. Format matters: an analysis with numbers and conclusions gets cited, promotional copy does not.

The second source is reviews and directory listings. Yandex Maps, 2GIS, Zoon and vertical aggregators form the factual base about a company: what it does, where it is, how clients rate it. Models treat that as checkable data. The third is roundups and comparisons. Pages titled ten development agencies in Moscow or contractor comparison feed straight into context when a user asks a model to recommend someone.

What does not work here: bulk press releases on aggregators, paid mention packages, doorway-style roundups. Models follow sources that search already trusts, and junk platforms are not in that pool. A realistic plan is one or two solid publications per quarter plus systematic review collection. It is slow, and it is exactly what builds durable presence in generative answers.

How to measure AI visibility: the method that actually works

The weakest part of the GEO market is measurement. Contractors sell optimisation for neural networks and report a screenshot of one lucky ChatGPT answer. That is not measurement: generative output is non-deterministic and the same prompt returns different results run to run. What you measure is frequency, not occurrence. Below is a method you can run in an ordinary spreadsheet that produces numbers comparable week over week.

Step one, fix the prompt set. Assemble 30 to 50 phrasings in three groups: commercial (which agency should I hire to build a website in Moscow), comparative (Tilda or a custom build), and problem-shaped (why is my site not growing in Yandex). The set stays unchanged for months, otherwise the numbers are not comparable. Adding prompts later is fine, deleting old ones is not.

Step two, control the conditions. Clean sessions only: logged out, incognito, memory and personalisation off, the same region every time. Otherwise you are measuring your own query history rather than what the engine really returns. Step three, repeat: run each prompt at least three times and score the share of answers containing a mention rather than a single hit. Step four, log it: date, engine, prompt, brand mention as 0 or 1, link to your domain as 0 or 1, where in the answer the mention sits, which competitors are named, and which domains were cited as sources.

Step five, the metrics. Mention Rate is the share of answers naming your brand out of all answers. Citation Share is the share of answers linking to your domain. Share of Voice is your mentions against all brand mentions in the answers. And the most useful metric for planning is Source Overlap, the list of domains the engines cite most often across your prompt set. That list is your external publication plan, because those are the places to get into.

Step six, validate on the server. Check the logs for GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and YandexBot visits to confirm the crawlers actually reach the pages that matter. In your analytics create a referrer segment for arrivals from ChatGPT, Perplexity and Yandex neural answers. The volume is small but conversion is usually above average, because the visitor arrives already holding a recommendation. Cadence: weekly at the start, fortnightly once things stabilise. Realistically the first movement shows up in 4 to 8 weeks.

• Mention Rate equals answers naming the brand divided by all answers. The baseline visibility metric.
• Citation Share equals answers linking to your domain divided by all answers. It shows whether the engine trusts you as a source.
• Share of Voice equals your mentions divided by all brand mentions. Your standing against competitors.
• Source Overlap is the set of most-cited domains. That set is your publication target list.
• A realistic quarterly goal is not first place but lifting Mention Rate across your cluster from zero to 15 to 30 percent.

What GEO cannot do and where the market oversells

The first promise to ignore is a guaranteed first place in ChatGPT. Guarantees do not exist here by construction: the answer is generated afresh each time, the source mix shifts and the selection criteria are never fully published. The correct way to state a goal is a rise in mention and citation frequency across a fixed prompt set, measured by the method above.

The second is GEO without SEO. If the site is not indexed, is slow and does not reach the top 20 for its own commercial queries, no amount of paragraph restructuring will lift it into a generative answer. Yandex neural search pulls from the main index and AI Overviews mostly from documents already ranking well. Selling GEO as a replacement for SEO is either a misunderstanding or a deliberate simplification.

The third is mass-generating articles with a model in order to please models. It degrades how ordinary search assesses the domain, and through that the odds of citation, since engines draw sources from the same index that treats hundreds of near-identical texts as low-value. The opposite works: fewer pieces, more substance, each page closing one question completely.

The fourth is measuring with a single run in a logged-in account with memory enabled. A model that remembers your earlier questions will happily name your brand, and that means nothing. A report made of screenshots like that is not a report. Ask for a table with dates, prompts, repeat counts and calculated rates.

A 30-day implementation plan

Week one is the audit and the baseline. Check robots.txt for AI crawler access, inventory the existing structured data, build a prompt list of 30 to 50 phrasings and run the first measurement across every engine. Without a baseline there is nothing to compare against later, which is precisely what turns GEO into an unmeasurable service.

Week two is the content rewrite. Take the 10 priority pages, the ones already ranking in the top 20 and answering commercially meaningful questions. Rewrite the opening of every section as a direct answer, reshape H2s into questions, add numbers, dates and at least one table per page, and cut the warm-up intros. It is the heaviest week and it produces most of the effect.

Week three is the technical layer. Implement Organization, Article, FAQPage and Service or Product, validate them, fix the heading hierarchy, add visible update dates, refresh the sitemap, register the site in Bing Webmaster Tools and switch on IndexNow. Add llms.txt if you like, but do not treat it as billable work.

Week four is the external perimeter and the re-measurement. Prepare one substantial publication for vc.ru or Habr, bring the Yandex Business listing to full completion, start collecting reviews, check the logs for AI crawler visits and re-run the same prompt list. Comparing week one against week four gives you the first genuine GEO report. After that the cycle repeats monthly.

• Week 1: crawler access, markup audit, prompt list, baseline measurement.
• Week 2: rewrite 10 priority pages for passage extraction.
• Week 3: JSON-LD, heading hierarchy, dates, Bing and IndexNow.
• Week 4: external publication, reviews, Yandex Business listing, re-measurement.

Frequently asked questions

Related services

• GEO: getting your brand cited by AI searchMore and more queries end in an AI answer rather than a click. GEO is the work of getting named in that answer, with a measured mention share rather than a promise.
• SEO and GEO website promotionSearch traffic is the one acquisition channel that keeps working after the ad budget stops. We run two tracks in parallel: classic SEO for Yandex and Google, and GEO, the work of getting your pages quoted inside AI Overviews, Yandex Neuro, ChatGPT and Perplexity, where a growing share of queries now ends. It starts with an audit, never with buying links.
• Website development, end to endWe build sites that still hold up a year later: from a one-page CMS build to a store or a web application on React and Next.js. One team for design, code, integrations and launch.

Read next

• SEO in Yandex in 2026: what changed and what no longer worksLink buying is dead weight, templated copy pushes pages into the low-value bucket, and behaviour manipulation is no longer a grey tactic but a way to lose the site. Here is what actually drives Yandex rankings in 2026.
• Yandex and Google: how promotion differs and how to work with bothOne site, two search engines, two different ranking logics. Where the differences actually change the work, why regional settings matter so much in Yandex, and how to build a strategy without writing everything twice.
• B2B Content Marketing: Making Articles Produce Leads Instead of PageviewsMost corporate blogs optimise for reach and report pageviews. Here is how to pick topics by funnel stage, why comparisons and pricing breakdowns produce enquiries, why author pages matter and which metrics belong in a monthly report.