GEO (Generative Engine Optimization): visibility in AI search

1. Retrieval

The system searches the search engine index or the web for sources that match the user's query. A page the crawler cannot see drops out at this stage already.

2. Selection

The engine picks passages that match the intent of the question. Compact, unambiguous text wins: text that can be pasted into an answer without cutting it or adding context.

3. Generation

A large language model (LLM) assembles one answer from the selected passages. At this stage the system often cites its sources, with a link or just the brand name.

The unit of work is a passage, not a page

A model takes a piece of text into its answer, not a whole page. So we edit paragraph by paragraph: the question in the heading, the answer in the first two sentences under it, with no references to what stands next to it. In SEO the whole page counted, here every paragraph counts on its own.

The measure is share of answers, not position

There is no list of ten results, so there is nothing to number. We ask models a fixed set of questions, count how often your brand comes up and next to whom, and repeat the same measurement every week. We report the reading once a month, because a shorter window shows mostly noise.

Half the work happens off your site

When asked to choose a supplier, a model reaches for sources that are nobody's advertising: discussions, comparisons, round-ups. In classic SEO this part was link building, here it is presence in content that others write.

A consistent company entity, not just keywords

The model has to recognise that all the mentions are about one company. One name, one address, one scope of services in structured data and in directories. When sources say different things, the answer gets whichever version the model came across, and you fix that at the source, not on your site.

Content and its structure

A model does not read a page the way a person does: it splits it into passages and uses the ones that match the question. So every section opens with a concrete answer, headings follow a logical hierarchy, and definitions, tables, lists and FAQs answer questions asked in full sentences, not just keywords.

Expertise and E-E-A-T

Google's guidelines on helpful, people-first content describe the same signals we look after for models, so one piece of work counts in both places. In practice: signed articles with an author bio, original data and benchmarks, cited sources, update dates and deep coverage of a topic instead of shallow texts for every keyword.

Structured data and technical basics

A model will not cite a page its crawler cannot fetch. The foundation: Schema.org for articles, FAQs, products, authors and organisations, a correct robots.txt and XML sitemap, Core Web Vitals and the mobile version. On top of that we add an llms.txt file, which organises what the model sees first; it is not an official standard and does not earn citations on its own.

The brand as an entity

A model names a brand it can link unambiguously to a topic, a product and specific people. We build that through consistent company data across the web, presence in the Google Knowledge Graph, digital PR and mentions in trade media, reports and interviews.

Optimising for conversational queries

People ask AI differently from classic search: instead of "seo agency london" they type "which SEO agency in London is best at e-commerce?". Content has to answer detailed, contextual questions with a clear intent, so its structure changes, not its substance.

Link building for AI

In GEO links still count, but where they come from counts more. Measurement shows which sites models in your category really cite, and that is where we aim link building: expert portals, trade sites and media that already appear in those answers.

Three differences affect every decision. First, there is no full transparency: classic SEO has documentation and historical data, while nobody has publicly described how AI systems choose their sources. What counts is systematic prompt monitoring and testing, not one-off actions. Second, volatility: the same prompt gives different results depending on the model, the time, the location and the wording, so stable visibility needs real topical authority. Third, hallucinations: generative systems can describe a brand wrongly or cite it in the wrong context, so GEO also covers monitoring the quality of mentions, not just their number.

A model builds its answer from pieces of text that match the user's question, not from whole documents. In practice the advantage goes to a paragraph with the question in the heading and the answer in the first two sentences under it, even if the neighbouring page is better as a whole. The same format lifts classic SEO, because Google has been looking for exactly this since it introduced featured snippets.

It will not change a model's opinion of your brand, because that comes from what others write, and you work on it as long as you work on reputation. Nor can you delete false information from a model: you can cut off its source and supply the correct version where the model looks, then wait until it stops coming back. Anyone promising either within a month is promising something they do not control. It will not work on a site whose content only appears after scripts run either: the model's bot sees a skeleton, so the foundation has to be fixed first, a site that delivers its full content in the server's first response.

The narrower the category, the easier it is to be cited in it, because the model has fewer candidates to name. A small company with a clearly described specialism is often better placed than a big one with a general description of everything, and this is the one place where the discipline can be kinder to smaller players than classic SEO.

SEO, or search engine optimisation, means optimising for classic search results and has rested on the same principles for twenty years. SGE, or Search Generative Experience, was the name of Google's 2023 experiment with generative AI in the results, which grew into today's AI Overviews. GEO means optimising for generative search engines: it covers all engines at once and concerns AI-generated content, not a list of links. Good SEO strategies do not lose from this shift, because the same signals that earn rankings also earn citations.

Our office is in Kraków, and visibility in AI models is by nature independent of location, so we work with companies from all over Poland and from foreign markets. Language does matter, though: a model answers a question asked in Polish differently from one asked in English, so the measurement question set is built separately for each market where you want to be named.