GEO for Real Estate: How to Get Your Brand into AI Answers

What GEO is

GEO, or Generative Engine Optimization, is optimising a website and its content to be cited in the answers of AI models: Yandex Alice, ChatGPT, Perplexity and AI Overviews in Google. The goal is clear from the name: when a buyer asks a model which residential complex to choose in their city and budget, your brand should be among those mentioned. The term appeared in 2023 alongside an academic study, and by 2026 optimisation for AI search had become a separate line in marketing budgets: Western brands already allocate to it a noticeable share of what used to go entirely into classic SEO. The Russian-speaking market is following the same route with a lag of a year to eighteen months: Yandex has built generative answers into its search, and the same citation logic now operates on the main source of traffic for real estate.

For a developer this is a question of presence on a new floor of the funnel. Choosing a flat stretches over three to six months and consists of dozens of questions: which district is growing, which developers to trust, what to pick between two specific projects. Some of these questions people already ask an AI model, and it answers with a short list of three to five names with explanations. If your project is not on that list, you lose the selection stage before the buyer even reaches the search results and your advertising. And you lose silently: the contact that never happened is invisible in the reports. Cost per lead in the ad account stays the same, reach stays the same, and the top of the funnel quietly drains into a channel where you are absent. That is exactly why GEO should be considered together with end-to-end analytics: both disciplines give marketing back its visibility where there used to be a blind spot.

Why buyers ask Alice and ChatGPT

Buyers turn to AI models for a ready answer: instead of ten links from the search results, Alice, ChatGPT or Perplexity compare districts, developers and specific projects right in the conversation. By 2026 ChatGPT's weekly audience is measured in hundreds of millions of people, and Alice lives in the speakers and phones of tens of millions of Russians. Promotion in Alice and ChatGPT has stopped being exotic: it is a response to where part of the questions about buying a home have physically moved. The person who used to open a search engine and three property portals now holds one long conversation and arrives on the website with a shortlist already in hand. The typical questions are easy to imagine: “is it worth buying a flat in this district”, “is developer N reliable”, “compare these two projects by price per square metre and completion dates”. To each of them the model replies with specific names and arguments it took from somewhere.

The share of AI sources in real estate website traffic is still modest: by various estimates one to three percent of visits, depending on the segment and region. What matters is the trend: year on year this share grows several times over, and such traffic converts noticeably better than search traffic, because the person arrives after comparing, with an opinion already formed. Marketing directors ask the models their own questions: whom to invite for an audit, which cost-per-lead benchmarks to consider normal, as we discussed in the article on enquiry cost by segment. A developer that AI models recommend wins both conversations: with the buyer and with the B2B client. The proportions differ by segment: in premium and overseas property the share of such clients is higher, because that audience adopts new tools earlier and compares options for longer before a deal.

How GEO differs from SEO and what they share

GEO differs from SEO in the unit of result: classic search fights for a position in the results and a click, while optimisation for AI search fights for a mention and a citation inside the model's answer. Yet they have more in common than not: both disciplines rest on quality content, the technical accessibility of the website and brand authority. A site that ranks well in Yandex and Google has already covered half the distance to being cited by AI models, because the models rely on the same indexes and the same quality signals. So setting the disciplines against each other is pointless: GEO is better described as a layer on top of SEO that adds a few new requirements for text structure and the site's technical wiring to the familiar work.

The difference shows in the details. The model assembles its answer from fragments, so the citability of an individual paragraph matters more than the position of the page as a whole. Some answers happen without a visit to the site, so a brand mention works even without a click or traffic. External platforms weigh more: an AI model is more willing to trust a developer written about by industry media, VC.ru and Habr than a website talking about itself. The horizon is different too: search positions can be moved over months of targeted edits, while a model's knowledge of a brand updates with crawler recrawls and the release of new versions, so the work has to run ahead. Measurement is also structured differently: instead of positions and clicks you track the share of answers where the brand is mentioned, the tone of the mention and the accuracy of the facts the model retells about you.

What gets a brand into AI answers

Content that answers the question directly in the first paragraph, contains specific figures and is backed by brand mentions on external platforms gets into AI answers more often. The Generative Engine Optimization study showed that adding statistics, expert quotes and source references lifts a page's visibility in model answers by 30 to 40%. Filler works the other way: the model will not retell abstract texts about a comfortable environment and unique architecture, there is nothing for it to take from them. A page with the phrase “flats from 12.5 million, completion in the second quarter of 2027, mortgage from six percent” has many times more chances in the answers than a page about the philosophy of life by the water.

The practical list of factors looks like this. The page title is phrased as a buyer's question, and the first paragraph gives a self-contained answer in two or three sentences. FAQ blocks and schema.org markup help the model extract facts: prices, completion dates, housing class, mortgage programmes. Tables and lists are cited more readily than solid text. A fresh update date raises the chance of appearing in an answer to a question with the word “now”. A separate layer: mentions on external platforms, from industry media to property portals, which the model treats as independent confirmation that the developer exists, builds and delivers on time. For a regional developer the achievable bar is three to five substantive publications a quarter, with project figures and a first-person comment, on platforms that are already in the models' indexes and training sets.

llms.txt, markdown versions and AI crawler access

llms.txt is a text file in the root of the website that describes the resource in a language convenient for language models: a short description of the project and a list of key pages in markdown format. Together with markdown versions of the pages themselves, it makes the site cheaper for a model to read: instead of heavy HTML with scripts and styles it gets clean structured text. For a residential project's website this means prices, dates and floor plans reach the model without the losses that are inevitable when parsing complex markup. The format was proposed in 2024, and its support has been expanding since: the file is read by agents and by the search crawlers of some models, and llms.txt generation tools are built into popular CMSs and frameworks.

The second part of the technical base is crawler access. Check robots.txt: GPTBot, ClaudeBot, PerplexityBot and YandexBot must not be blocked. Developer websites often block them along with aggressive price scrapers and then wonder why the brand is invisible in the answers. The third part is schema.org: Organization, FAQPage and project card markup that hands the model facts in structured form. An honest caveat: llms.txt is still a convention, and not every system reads it. But implementation takes hours, the cost is close to zero, and the usefulness grows with every new release of models and crawlers. In a niche where almost none of your competitors have done this yet, an early step gives a disproportionately large advantage.

How to check whether AI models see your project

The check starts with a direct question. Ask three or four AI models (Alice, ChatGPT with search enabled, Perplexity and Google with AI Overviews) which residential complexes they would recommend in your city and segment and what they know about your developer at all. If the brand is absent from the answers, if the model confuses you with a competitor or cites data two years old, the scope of work is defined. Such a measurement takes an hour and gives more than a week of theoretical discussion. It helps to phrase the questions the way a buyer does, without the brand name: “which business-class residential complex to choose near the metro”, “which developers in the city have never postponed completion”. Answers to branded and unbranded questions often diverge, and the work needed differs in each case.

Then step by step. First: crawler access in robots.txt, this is fixed in an hour and takes effect before anything else. Second: the first paragraphs of key pages, do they answer the buyer's question without scrolling. Third: FAQ blocks and schema.org on project pages. Fourth: llms.txt and markdown versions. Fifth: external mentions, at least three to five publications a quarter on platforms the models trust. Repeat the measurement once a month against the same list of 10 to 15 questions. The niche is young, specialised LLM visibility tracking services are only just appearing, so a manual check against a fixed list is for now more reliable than any dashboard. The priority is simple: technical accessibility first, then content structure, then the external layer. The reverse order does not work: media publications will help little if the model physically cannot read the site and verify the facts.

How Matveo implements GEO on its own site and on client projects

At Matveo we implement GEO on our own website and on client projects: matveo.com runs llms.txt, markdown mirrors of articles, schema.org markup and a text structure where the first paragraph of every section gives a direct answer to the question in its heading. We carry the same set over to developers' websites together with a crawler access check and a plan for external publications. We test the approach on ourselves first, then offer it to the client: it is more honest, and you see faster what works. GEO remains one layer of the system: alongside it run end-to-end analytics down to the deal and AI agents that manage campaigns daily, which we covered in the articles on MCP and the closed advertising loop.

It all starts with a two-week marketing audit: we look at how the brand appears in AI answers, what the site serves to crawlers, where traffic and enquiries are lost. The cost of the audit is credited towards the pilot, so testing the hypothesis does not mean paying twice. The logic of specifics works beyond AI search too: in a case with content for a narrow B2B audience, betting on figures and direct answers brought 240 qualified enquiries from top managers in three months. Material that honestly answers the reader's question gathers both people and model citations: it is the same content, it just has more readers. The same principle holds up the rest of our cases, from a premium residential complex in St Petersburg with a sixteenfold return on traffic to international projects: decisions are made on figures, and content is built around figures.

Frequently asked questions

What is llms.txt in plain terms?It is a text file in the root of the website, a map of the resource for language models: a brief description and a list of key pages in markdown format. It helps an AI model quickly understand what matters on the site without parsing heavy HTML. Implementation takes an hour or two and requires no changes to the rest of the site.

How do you get into Google AI Overviews?There is no special application: AI Overviews are assembled from pages Google has already indexed and considers authoritative. The same factors work as in all of GEO: a direct answer in the first paragraph, schema.org markup, specific figures and brand mentions on external platforms. For the Russian market the focus shifts to Alice and Yandex Neuro, where the mechanics are similar.

When should you expect an effect from GEO?Technical fixes (robots.txt access, llms.txt and schema.org) show up in model answers within a few weeks, as crawlers recrawl the site. Building up external mentions and content takes three to six months. Measure visibility monthly against a fixed list of questions, and you will see the trend before traffic growth appears in the reports.

Will GEO replace classic SEO?Not in the coming years: search traffic still provides the main share of visits, and AI sources a few percent. GEO builds on top of SEO and uses the same base: content, the technical health of the site, brand authority. A sensible strategy develops both layers at once, especially since the effort overlaps by around seventy percent.

How do you measure enquiries from AI models?Visits from ChatGPT, Perplexity and AI Overviews are visible in analytics by referrer, although some visits are disguised as direct traffic. A “how did you hear about us” question in the enquiry form and the first-call script helps: the share of “I asked an AI” answers is already measurable. Measurement accuracy in the niche is still low, and it is worth admitting this honestly, allowing a quarter for the experiment.