AI SEO / GEO: so that assistants name your brand when someone asks

How work on AI visibility goes, step by step

1. Baseline. We check how the brand looks today: in which answers it appears, next to whom and with which sources. In parallel we assess the technical state of the site. We most often do this as part of an SEO and GEO audit.
2. Technical accessibility. We remove blocks for crawlers, fix indexing and check whether the key content is in the page code.
3. Question set and gaps. The measurement shows which questions the brand has no answer to on its site, or where it is absent from assistant answers even though competitors are present.
4. Content. We write or rebuild pages so that they answer those questions in a citable way: with a definition, data and a comparison.
5. Entities and data. We unify the company description, add structured data and correct discrepancies in directories and listings.
6. External sources. We identify the sites that assistants cite in your industry and plan the brand’s presence there, for example through expert publications, profile updates or inclusion in roundups.
7. Measurement and corrections. Each month we compare the indicators with the baseline, check which changes had an effect and adjust the plan.

The order matters. New content may not help your site get cited if the system cannot access the page, and work on external sources without consistent company data can scatter the message.

Who it makes sense for and what the scope depends on

Work on AI visibility makes sense where customers ask assistants about services, products and comparisons: in specialist services, B2B and e-commerce, and also in local businesses, if customers look for recommended providers. If independent sources rarely describe the brand and the site has technical problems, you first need the basics described in the section on the factors that affect presence.

The scope and cost depend on several things that we establish once we know the site:

• the technical state of the site and the number of problems to remove;
• the number of topics and questions to be covered by measurement;
• the number of competitors against which we calculate share of voice;
• the size of the site and the amount of content to rebuild or write;
• the brand’s presence in external sources and the need to increase it;
• the language versions and markets in which the brand is to be visible.

If you want to start with a diagnosis, it is described on the SEO and GEO audit page. Measurement alone, without changes to the site, is the scope of monitoring. For sites that also need work on classic visibility, the starting point is SEO services.

What cannot be promised

Nobody except the assistant’s provider controls its answers. So we do not promise a specific position for the brand in an answer, its being named for a given question, or a date by which that will happen.

• Answers change. The same sequence of questions can give a different result after a model update, a change of sources or ordinary variation in answers.
• The mechanisms are not fully public. Providers describe only some of the rules for choosing sources. They do not publish the full algorithms or weights, so we supplement their documentation with observation and tests.
• Our influence is limited. We can improve what depends on the site owner and on the brand’s presence in sources. We cannot affect the knowledge stored in the model or the provider’s decisions.
• Answers can be wrong. When an assistant gives false information about a company, we can identify the source of the error and correct the data where possible. We cannot guarantee that the assistant will change its answer.

For this reason we report visibility indicators: presence in answers, citations and share of voice, together with the change against the baseline. We do not present them as a promise of traffic or sales.

How we measure visibility in AI

Assistant answers change between queries, so a single check is not reliable. Our measurement rests on a fixed set of questions and regular repetition, described further in the section on AI visibility monitoring.

1. We build a question set from real search keywords and conversations with clients: category questions, comparisons, brand questions and problem questions.
2. We ask these questions in ChatGPT, Gemini, Perplexity and Claude and save the full answers together with their sources.
3. We calculate the indicators: in how many answers the brand appears, how many times its domain is cited, and what share of all the brands named it accounts for (share of voice against competitors).
4. We assess the context: whether the brand description is correct, whether the answer is positive, neutral or wrong, and which sources shape the description.
5. We repeat the measurement at fixed intervals and compare results between periods.

Code calculates the numbers, with no manual tallies, which makes results comparable between months. We attach the question set and the text of the answers to the report, so that every indicator can be checked against its source.

Besides AI answers, we track visibility in Google and the state of indexing. We compare the state of the site in the search engines a given assistant uses with the results of measuring its answers.

What AI SEO and GEO are

AI SEO is work to make AI assistants know a brand, name it in their answers and cite its pages as sources. A user asks ChatGPT, Gemini, Perplexity or Claude for a recommended company, a comparison of services or a solution to a problem. The answer often contains several brand names and links to the pages the information came from. The aim of the work is a place in that answer.

An AI answer often combines information from many pages into one text, sometimes with product cards and a list of sources. How many brands it names depends on the question and the system, so we measure the brand’s presence and its share against competitors.

It also helps to understand what this work does not cover. It does not involve buying a place in the answers or influencing a model with hidden instructions in your content. An ad in ChatGPT is a separate, paid format shown below the answer. Work on AI visibility covers what a site owner actually controls: the site’s accessibility to crawlers, the quality and structure of the content, the data about the company and the brand’s presence in the sources assistants use.

Why so many names: GEO, AEO, AISO, LLM SEO

The industry has not settled on one name, so in offers and articles you will meet several abbreviations for similar work.

• GEO (Generative Engine Optimization): optimisation for search engines that generate answers. We use this name most often, alongside AI SEO.
• AEO (Answer Engine Optimization): optimisation for answer engines, meaning any system that returns a ready answer instead of a list of links. The term was in use before assistants based on generative AI became popular, and it also covered featured snippets in Google.
• AISO (AI Search Optimization): optimisation for AI-assisted search.
• LLM SEO: optimisation for large language models, the technology the assistants run on.

The differences between these names are mostly marketing. The actual work is defined by the goal: the brand’s presence in the answers of AI assistants. When you compare offers, look at what the provider actually does and how they measure the result. The abbreviation matters less.

How an assistant builds an answer with web search

Web search mode works, roughly, in four steps. Individual systems differ in the details, and their providers do not disclose the full mechanism, so we describe the general outline without guessing at the algorithm.

1. Finding sources. The assistant turns the user’s question into one or more queries and fetches a list of pages from a search engine or from its own index.
2. Selection. From the pages found, the system picks those that look useful and reliable, and reads their content.
3. Synthesis. The model assembles the passages it has read into one answer in its own words. It also draws on what it learned during training.
4. Citation. Some systems attach links to sources to the answer. A brand can be named without a link, and a link does not mean the page was used in full.

Two practical conclusions follow. First, good indexing helps an assistant discover your page, although it can also fetch a URL the user gives it directly. Second, depending on the tool, the assistant processes selected passages or the full content of the page. Content that is understandable when a single paragraph is lifted out has a better chance of being used than content that needs the context of the whole article.

When a question does not need current data, the assistant may answer without searching, using only the knowledge stored in the model. Then the brand’s presence depends on how often and in what contexts it appeared in the material the model was trained on. That influence cannot be changed quickly, so we mainly work on the search-based part.

How AI SEO differs from classic SEO and what they share

Both fields rest on the same foundations: a site has to be accessible to crawlers, fast, understandable and credible. They differ in what the result is and how to read it.

• Result. In classic SEO it is a list of pages with positions. In AI SEO it is a text answer with a few brands and sources.
• Unit of work. In SEO it is a keyword and a page. In AI SEO it is a question, a topic and the brand as a whole.
• Metric. In SEO: positions, visibility and indexing. In AI SEO: mentions, citations and share of voice.
• Repeatability. Google positions can be tracked, although they too can change sharply. The same assistant answer can look different the next time you ask.
• Role of external sources. In SEO, links strengthen a page. In AI SEO, mentions in independent sources shape the description of the brand.

The shared elements show up in practice. A site that is blocked from crawlers, slow or full of duplicate URLs does harm in both. Content that answers a question clearly helps in Google featured snippets and in assistant answers alike. That is why we start work on AI SEO by checking the basics familiar from SEO services, and only then move to the elements specific to AI.

For its own AI features Google says so directly: SEO best practices remain relevant, and there are no additional requirements or special optimisations to appear in AI Overviews or AI Mode. A page has to be indexed and eligible to be shown in Google Search with a snippet.

The difference in repeatability affects measurement. A single answer proves nothing, because asked again the assistant may name other brands. So we calculate the result over a set of questions and repeat the checks over time, as described in the section on measurement.

What affects a brand’s presence in AI answers

Assistant providers describe some of the factors for choosing sources; OpenAI, for example, mentions user intent, relevance and recency. They do not publish the full list or the exact weights. The areas below follow from how search and answer generation work, and from observing answers in our measurements. We treat them as working hypotheses that we test on each brand’s data.

Site accessibility for crawlers

A fetching block limits that crawler’s access to the content. Its effect on citations depends on how the assistant obtains information. We check server response codes, blocks in robots.txt and in the firewall, noindex, and whether the most important content is in the page code and not loaded by scripts. Some AI crawlers do not execute JavaScript the way a browser does, so content that appears only after it runs may be invisible to them.

Content that is easy to cite

The system picks passages that can be used without further explanation. These help:

• a definition of the concept in the first sentence of a section, before the elaboration;
• concrete data with a named source and a date, for example dimensions, ranges, deadlines, conditions;
• comparisons in tables with clearly named columns;
• questions in headings and direct answers beneath them;
• step-by-step lists for how-to content.

Such passages are also easier for people to read, so this work does not require writing separate texts for machines.

Consistent company data

The name, address, scope of services, business description and contact details should read the same on your site, in your Google Business Profile, in industry directories and in social media profiles. Discrepancies make it harder for a system to establish that it is one company, and they are a source of wrong answers, for example with an old address or an offer that no longer exists.

Mentions and external sources

Assistants readily draw on sites they consider independent: industry portals, rankings, comparison sites, review sites, forums and encyclopaedias. A brand’s description in such places can end up in answers. We check which sources assistants cite in answers to questions from the client’s industry, and we assess where the brand is absent or described out of date.

Structured data

schema.org markup for organisations, services, products, articles and questions with answers describes content in a way machines can read without ambiguity. There is no public guarantee that a particular markup will increase the number of citations. Structured data does, however, organise information about the entity, meaning what the company is and what it offers, and it is also useful in classic search.

What content to prepare so that it can be cited

An assistant answers questions, so it is worth preparing pages that answer customers’ questions factually and checking in measurements whether assistants use them. We choose those questions from real search queries, from sales conversations and from measuring AI answers.

Service and category pages

The start of the page should briefly explain what the service is, who it is for and what it includes. Further down it is worth giving the scope, the terms, the stages of cooperation and the factors that affect the cost. An assistant looking for an answer about the scope of a service then finds a ready passage.

Comparisons and roundups

Questions like “what to choose” and “how does it differ” are worth answering with a comparison. A factual comparison of solutions, with the selection criteria and the situations where each one makes sense, gives the system material for synthesis. An honest comparison that admits the weaknesses of your own offer is more credible than praise alone.

Guides and answers to questions

Longer expert texts build the site’s topical credibility, and FAQ sections provide short, self-contained answers. Each answer should be understandable without reading the rest of the page.

About-the-company pages

Assistants may answer questions about what a company does and offers. A page that describes the business, the scope of services, the area of operation and the contact details can be one of the sources of that information. It should agree with the description in the other places on the web.

Updating

Content with out-of-date service prices, team details or deadlines misleads the systems. For each page it is worth deciding who reviews it and when, and marking the date of the last update.

AI bots in robots.txt

The robots.txt file defines which crawlers may fetch which parts of a site. Assistant providers use several crawlers with different tasks, so a single rule saying “block all AI bots” is usually too broad. Examples of names that the providers themselves publish:

• GPTBot (OpenAI): a crawler that collects content which may be used to train models.
• OAI-SearchBot (OpenAI): a crawler that indexes pages for search in ChatGPT.
• ChatGPT-User (OpenAI): fetches a page when a user asks for it in a conversation. Because a person triggers the visit, robots.txt rules may not apply to it.
• ClaudeBot (Anthropic): collects content which may be used to train models.
• Claude-SearchBot (Anthropic): indexes pages for search in Claude.
• Claude-User (Anthropic): fetches a page when a user asks a question in a conversation.
• PerplexityBot (Perplexity): shows pages and links in Perplexity results. According to the provider, it does not collect content for training models.
• Perplexity-User (Perplexity): fetches a page when a user asks a question.
• Google-Extended (Google): a separate robots.txt directive with which a site owner can limit the use of content in Google’s generative models. It is not a separate crawler that fetches pages.

The practical conclusion: blocking GPTBot or ClaudeBot does not block search in ChatGPT or Claude. Blocking a search crawler limits access and may reduce visibility, and its effects depend on the provider and how the page is fetched. Names and purposes of crawlers change, so before deploying rules we check each provider’s current documentation.

How to decide

Blocking a crawler that serves answer search lowers the chance of the site being cited. Blocking a crawler that collects training data is a separate decision, depending on the company’s policy on its own content. A publisher with paid access to its material will decide differently from a service company that wants to be found.

Robots.txt is not a security measure. Crawlers that follow the standard respect the rules, and confidential content is protected by a login, not by an entry in a file. A server firewall, bot protection or CDN settings can block AI crawlers even when robots.txt allows them, so in an audit we check both levels.

We checked how this looks in practice in a study of 296 Polish online shops from Google’s top 10. 48 of the 282 shops whose settings we could assess blocked at least one AI bot, and in 12 of them the rules were added by a setting in the Cloudflare dashboard.

The llms.txt file: what it gives and what it does not

llms.txt is a proposed file in the root directory of a domain that briefly describes the site and points to the most important resources in text format. The idea is to help models find the core of the offer without crawling the whole site.

It is a convention. Assistant providers have not confirmed it as part of ranking or source selection. So we treat it as an extra. Maintaining such a file costs little and may help tools that read it. Adding it does not replace a technically accessible site, clear content, consistent data and reliable external sources.

If the file is created, it should contain a concise description of the company and a list of key pages with a one-sentence description of each, and it should be updated when the offer changes. An outdated file with URLs that no longer exist can send tools to the wrong resources, so it is worth keeping it current.

Common mistakes in work on AI visibility

• Believing in a single file or markup. Adding llms.txt or schema.org alone, without improving content and accessibility, does not guarantee better visibility. The effect of changes has to be checked in measurements.
• Blocking all AI bots. A rule put into robots.txt after reading one article can cut a site off from search in assistants, even though the aim was only protection from training.
• Judging by one answer. A single question asked by hand does not show a stable picture. You need a set of questions repeated over time.
• Content written for machines. Texts made up only of keywords and question headings, with no substantive answer, give no passages worth citing.
• Ignoring external sources. Expanding your own site is not enough when assistants describe the brand from third-party sites where it is missing or out of date.
• Inconsistent company data. Different names, addresses and service descriptions in different places make it hard for systems to recognise a single brand.
• No reference point. Without measuring before the changes, you cannot tell what worked.
• Neglecting classic SEO. If a page is missing from the index of the search engine an assistant uses, that can limit its discovery.