AI SEO and GEO Without the Myths: What Actually Works

A brand gets into the answers of ChatGPT, Gemini, Perplexity and AI Overviews mainly by two routes that overlap, and AI SEO is the work on both:

1. Through search. While answering, the model searches the web and cites some of the pages it found. The same things work here as in SEO: access for bots, the index, visibility for a keyword and its variants, and passages that answer the question without the rest of the page. As in SEO, the effect shows after weeks or months.
2. Through the model's memory. The model answers from what it learned before its training ended, and that memory also affects whom it names when it does search the web. What matters here is what others have written about the brand, and one consistent description of the brand across many sources. I have not measured how fast the memory changes, so I promise nothing here within a quarter.

Most “GEO tricks” have no evidence, and I tested some of them on my own data. There is no evidence that an llms.txt file or structured data produces citations. More headings did not lift pages into the top 3 in my test. The test of whether position for sub-queries predicts citation better than position for the main keyword settled nothing, and my first conclusion from it was overturned by a cross-check. The evidence for each of these sentences is below.

What is AI SEO, and what is GEO?

AI SEO is the work of making a brand appear in answers generated by AI: in AI Overviews and AI Mode in Google, and in ChatGPT, Gemini, Perplexity and Copilot. The same work has several names, and the differences between them are mostly marketing:

The name GEO has a scientific source, and the foundation of all this work is the same as in SEO. GEO comes from a research paper from November 2023, accepted at the KDD 2024 conference. Its authors, Pranjal Aggarwal and colleagues, showed on their own test set that changes in content can raise a source's visibility in generative answers by up to 40% (arXiv). Google writes in its documentation that optimizing for generative search is still SEO (Google Search Central). Cyrus Shepard of Zyppy answers the question of whether a company needs a separate GEO strategy in a similar way:

People do not have one name for this work either, and it shows in Google results. The question “What is AI SEO called now?” appears in the “People also ask” section in results from the USA, Sweden and Singapore, and its counterparts appear in results from Poland, Germany, Spain and Finland. The question about the difference between GEO (or LLMO) and SEO appears in 9 of the 27 markets I checked, from Brazil to Japan.

The name that dominates depends on the market. In Japan, 72% of about 16,730 searches a month on this topic are LLMO, and the keyword “llmo” alone has about 12,100 searches, more than “ai seo” in the USA (about 6,600). In Germany and France GEO dominates (64% and 66%), in South Korea it has 83%. In Poland, 58% of about 2,265 searches a month are AI SEO, together with the keyword “ai seo” and Polish descriptions (“pozycjonowanie ai”, “pozycjonowanie w ai”), and GEO has 32%. In the USA AI SEO and GEO are almost level: 42% and 39% of about 25,400 searches a month, and in Spain it is a tie (44% and 44%). In the Czech Republic there is no demand for any of the names in the data.

The keyword “ai seo” is ambiguous, so I treat the share of AI SEO as an upper bound. For “ai seo” Google shows, next to texts about optimizing for AI answers, AI tools and guides on how to do SEO with the help of AI: in the USA that is more than half of the top 10, in Denmark at least 6 of 8. The AI Overview for this keyword in the USA and in Denmark does describe AI SEO as optimization for AI answers, so Google treats it as a name for this work. Without this one keyword GEO leads in 25 of the 26 markets of my blog, and in Japan LLMO stays on top. In Poland the unambiguous Polish descriptions (“pozycjonowanie ai”, “pozycjonowanie w ai”) have about 580 searches a month together, and GEO about 725. I will devote a separate post to AI tools for SEO.

I do not count the bare keyword “aeo”, because in Poland it mostly means a customs certificate, and in the USA some of the results are stock quotes for the clothing chain American Eagle. For “aeo” in Polish Google, three of the first four results are pages of the Ministry of Finance. Keywords with the full name, for example “answer engine optimization”, I count in the “Other” series.

Each language version of this post uses the name people search for in that market. The Japanese one talks about LLMO, the German and French about GEO, and the Polish about “pozycjonowanie w AI” and AI SEO. This English version uses AI SEO next to GEO, because in the USA the two names are almost level in searches, with AI SEO slightly ahead.

How does AI choose the sources for its answers?

A model answers from search or from memory, and that determines what you can change at all. When it searches the web, it splits the question into several queries, pulls the results, picks passages from them and cites some. In its documentation Google calls this “query fan-out”: the system sends many related searches about subtopics and from various sources (Google Search Central). When the model does not search, it answers from what it learned during training, and then a page's position in Google changes nothing.

These routes overlap, because the model's memory also affects the answer when it searches the web. Dan Petrovic of Dejan writes that a brand mention in the text of an answer rests primarily on the model's prior knowledge, and a single page hardly moves it (Dejan, July 7, 2026). That is why I measure the two routes separately, but work on them in parallel.

How often a model searches the web depends on the model: in my test from 10 to 94 keywords out of 94. I checked it through the API on 94 keywords from projects I run, in English and Polish, overnight from October 3 to 4, 2026 and again in the morning. In the post on SEO strategy I described the same effect on five how-to questions, and here it is on a larger set.

• 93 of 94The number of times Gemini 2.5 Flash used search in the first run (94 of 94 in the second).My API test, October 3-4, 2026
• 71 of 94The number of times Gemini 3.5 Flash searched (69 of 94 in the second run).My API test, October 3-4, 2026
• 25 of 84The number of times Gemini 3.1 Pro searched (10 calls ended with a rate-limit error; 32 of 94 in the second run).My API test, October 3-4, 2026
• 10 of 94The number of times Gemini 3.8 Flash searched (15 of 94 in the second run). For the other keywords it answered from memory.My API test, October 3-4, 2026

In this setup newer versions of Gemini Flash searched less often, but Gemini 3.1 Pro does not fit that pattern. Pro has an older version number than 3.5 Flash, and it searched less often than that model. This is one model family and one API configuration, not the Gemini app that users see, so I treat the numbers as a signal, not a rule. They are enough, though, to make me check, in every measurement, on how many questions the model searches at all.

Models' memory of a new Polish brand turned out to be full of holes, and some models made up its description. In the same project, about 500 answers to questions about brands I work for were graded blind by a separate model. One of the new Polish brands in that set was unknown to all the models I checked, and in 9 of 12 answers about it the description was made up. The share of made-up answers depended on the model: from 5% to 77% of answers about all brands. The answers were graded by a single judge model, so I treat differences below about 10 percentage points as noise.

The same domain can be visible on one route and invisible on the other. One site was cited in Google's AI Mode for 10 of 13 keyword pairs, and in the answers of models without search it did not appear even once (0 of 13). It is one domain and 13 pairs, so it is an illustration, not a rule.

A related effect shows up in a larger study. Dan Petrovic reports a Google Research study on 13 models and 2,150 facts from Wikipedia: the leading models among them, such as Gemini 3 Pro and GPT-5, hold 95-98% of this knowledge in their weights, but cannot recall 26-34% of it (Dejan, August 17, 2026). These are encyclopedic facts, not knowledge about brands, but the conclusion for a brand is similar: having information in the training data does not guarantee that the model will say it.

How do you get into AI answers through search?

On the search route the same things work as in SEO, with one difference. Google writes that to appear in AI Overviews and AI Mode, a page has to be in the index and eligible to appear in results with a snippet, and that there are no additional technical requirements (Google Search Central). The difference is that the model asks several queries at once, so a page visible only for one main keyword has fewer chances of being cited than a page visible for its variants too. ChatGPT also has its own search bot (more on it below), so visibility in Google does not carry over to it directly.

Access for AI bots

AI search bots can be blocked without the site owner knowing. A firewall or a content delivery network (CDN) setting can block them, often only after a request threshold is passed, so one test will show nothing. I check a series of requests with the names of Google, OpenAI, Perplexity and Anthropic bots and compare the server's response with what an ordinary browser gets.

Blocking model training is a different decision than blocking search. OpenAI separates these roles: GPTBot collects content that may be used to train models, and OAI-SearchBot decides about presence in ChatGPT search results. OpenAI writes that sites that block OAI-SearchBot will not be shown in ChatGPT search answers, and blocking GPTBot does not change that (OpenAI). Similarly, Google-Extended controls the use of content by Gemini models, but according to Google it does not affect presence or position in search (Google Search Central). Whoever blocks all AI bots with one switch may lose presence in ChatGPT search that they never meant to give up.

Position for the keyword and sub-queries

I do not know whether a higher position increases the chance of citation, because my data does not settle it. In the post on SEO strategy, in the step on hypotheses, I described an article that was cited in an AI Overview on only one of five very close keywords, and on the one where it ranked lowest. Sub-queries are composed by the model and can differ with every answer, so there is no fixed list of “sub-queries for keyword X”. In my set of 94 keywords the model generated 493 different sub-queries in one run, on average more than five per keyword.

I tested whether position for sub-queries predicts citation better than position for the main keyword, and my first result did not survive a cross-check. On keyword-page pairs from projects I run, I wrote down the rule before I computed the result. The measure was the area under the ROC curve (AUC), that is, how well a signal separates cited pages from uncited ones (0.5 is the level of a coin toss, 1 is separation without errors). Sub-queries were to win if their AUC was higher by at least 0.10. On 29 pairs held out for the AI Overview test, the difference was 0.44, so the test “passed”.

The cross-check showed that I had compared two measurements of different accuracy. The reviewer who recomputed everything from scratch reproduced my numbers to the digit, but noticed that I took the position for the main keyword from a single pull of Google results, and the citation from five. After I computed the position from the same five pulls, the difference fell to 0.04, with an interval from -0.28 to +0.36. So I have not shown that sub-queries are better than position, nor that they are worse. Both signals tell me something, so I work on both.

Position from a single DataForSEO pull turned out to be a draw of one of several variants of the results. When I pulled the same 94 keywords three times about an hour apart, two pulls of the same keyword had few URLs in common in the top 10 at the median: the Jaccard index, a measure of agreement between two sets from 0 to 1, was 0.14-0.15. Some of the variants are probably results that a person in a browser does not see, because on 5 keywords the browser matched one of them. The AI Overview was stable: its presence matched between pulls in 97-99% of cases. That is why I now compute position from this tool as a share across about 10 pulls sent at once. I have not tested other tools this way.

Passages that answer on their own

A model usually uses not the whole page but the passages that answer its query. That is why I write so that every passage can be understood without the rest of the page: the answer in the first sentence under a heading, concrete numbers with a denominator and a date, names instead of “this” and “that”. Every paragraph in this post starts with a sentence that states the conclusion for exactly this reason. It is good writing practice that also helps people, not a separate technique for AI.

That does not mean you have to cut content into small pieces or multiply headings. Google says plainly that there is no requirement to split content into small chunks so that AI understands it better (Google Search Central). I also checked whether more headings help reach the top 3. On 632 top 10 pages for 119 English queries from the business software niche (USA, September 26, 2026), the odds ratio was 0.95 (95% confidence interval: 0.81-1.12), and 1.00 after accounting for text length. The share of pages in the top 3 was almost the same in three groups from the fewest to the most headings: 32%, 34% and 32%. It is the same test I describe in the post on SEO strategy: one niche and one day of measurement. It is enough, though, for me not to add headings to improve position: a heading should say what the passage under it is about.

Polish questions, English sub-queries

In my set, some of the sub-queries for Polish questions were in English. 81 of 277 sub-queries generated for Polish keywords showed no features of Polish. For example, a Polish question about a ranking of products in one category turned into an English query with the year in it. It is a rough heuristic on one set, but it is enough for me to check, for a Polish brand, what English-language sources say about it too.

How do you get into a model's memory?

You cannot change a model's memory with your own site alone: what others write about you is what counts. A model learns from large collections of text gathered before its training ended. If a brand is written about mainly by itself, and third-party sources are old or contradictory, the model will repeat an old or contradictory picture. I have no measurement of how fast models' memory refreshes. I assume it happens only in the next versions, so I promise nothing here within a quarter.

On my own brand I see both sources of error at once. Gemini described me with an employer from several years ago in the present tense, and described rogala.pro as a site “under construction”. The two descriptions that models repeated about me most often come word for word from my bio on an industry site that ranks high in Google for my name. At the same time, none of the 26 pages I checked in the results for my name gives that employer as current, so that error is probably sitting in the model's memory. I will check it with the same questions once the bio is updated.

What a model remembers about a brand, you can check yourself, without AI visibility tools. Ask questions about the brand with no access to search, most reliably through the API without a search tool. Ask what the brand does, for whom and since when it has existed, each question three times. Then compare the answers with the facts and write down separately what is current, what is outdated and what is made up.

Mentions without a consistent brand definition repeat the old description, so I start with identity. My order:

• one definition of the brand, the same on the site, in profiles and in bios on other people's sites;
• structured data with identifiers that I checked by hand;
• entries in knowledge bases such as Wikidata, only where there are sources that confirm them;
• only then mentions and publications in places edited by someone else.

An identifier has to be opened and read before it goes into the page code. On October 5, 2026 I noticed that the Knowledge Graph identifier that had been in the structured data of rogala.pro since the day before leads to the panel of an actor with the same first and last name. For about a day my site was declaring to search engines that I am an actor. Checking takes a minute: the address google.com/search?kgmid= with the identifier appended shows who it really belongs to.

Wikidata will not give you a place that has no sources. Wikidata requires sources for information about living people, so you cannot declare in it a specialization that nobody but you has written about. That is why I started with publications in places edited by others, not with a database entry.

What works in AI SEO, and what is a myth?

I divide claims about GEO by the strength of evidence. Documented means the search engine's documentation confirms it, and my measurement is a result from one setup that I repeat. Refuted and not shown are claims I checked on my own data that did not survive or settled nothing. No evidence means popular advice with no measurement behind it, and sometimes with evidence to the contrary.

There is no evidence that an llms.txt file and Markdown versions of pages help you get into AI answers. Google writes that to be present in its search you do not need to create new machine-readable files, text files for AI, markup or Markdown (Google Search Central). When I checked 10 domains from my portfolio in July 2026, none served a Markdown version to AI agents. On rogala.pro every page has had a Markdown version and an llms.txt file since October 8, and until November 8 I am measuring in server logs whether anyone downloads them. In Poland the keyword “llms txt” has about 720 searches a month, so I will cover that topic separately.

Structured data describes a brand to a search engine, but there is no evidence that it produces citations. Google writes that structured data is not required in generative search and that there is no special schema.org markup to add (Google Search Central). Matt Buxbaum of Seer Interactive analyzed 8,500 keywords with an AI Overview and 6,354 pages that won a first citation in them: schema and E-E-A-T signals were the weakest predictors of a win, and pages with the most FAQ and HowTo markup did about 10 times worse (Seer Interactive, May 28, 2026). That is a correlation, so I do not know whether the markup hurts, but I do not read this data as evidence of a benefit. I use structured data to describe entities (who, what, where, with which identifiers), because that is what it is for. I have not finished my own test of its effect on AI answers.

A mention and a citation are two different things, and many reports confuse them. Petrovic separates three events: a page that the system retrieved as material for the answer, a citation, that is a page shown as a source, and a mention, that is the brand name in the text of the answer that the user reads and remembers (Dejan, July 7, 2026). A brand can be cited without being mentioned in the text, and named without its page being cited. That is why I measure the two separately and do not add them into one “AI visibility score”.

How do you check whether AI SEO works?

Without a baseline you cannot attribute an effect to any change, so I record it before the first fix. The baseline includes the questions recorded verbatim, the names and versions of models, the date and the market. I build the questions from demand data and from Search Console, and most of them do not contain the brand name, because a name in the question prompts the model with the answer. I describe the question set from two position bands in the post on SEO strategy.

I measure five things separately, because each can go in a different direction:

1. A mention of the brand in the text of the answer.
2. A citation of a page as a source.
3. Whether the model searched the web at all, that is, search and memory separately.
4. Traffic from AI assistants in analytics, which is always a lower bound, because some visits arrive without source information.
5. Traffic from Google in the same period, so that a change for AI does not come at the expense of regular results.

Chat answers are random, so I ask one question three times at once. An AI Overview is stable and one pull is enough for it. I always read a number from an AI visibility tool together with its denominator: it is a share of my question set, not a “share of the AI market”, which nobody knows.

What am I doing on my own site, and what don't I know yet?

I have been talking about AI in search publicly since 2016, and I do not yet have my own brand in AI answers, which is where my experiment began. The baseline from July 26, 2026: rogala.pro had not a single citation on any of the 7 AI surfaces checked in Ahrefs for Poland, and in 6 checks on three platforms, in questions without my name, no model mentioned me. Asked about me by name, a model repeated a description from several years ago. It is one measurement and it lacks ChatGPT, because the tool I collected the answers with returned no results for it, so I treat it as a starting point, not as evidence.

The order of work follows from the data: identity first, then reach. I put in order one definition across profiles (on LinkedIn in 30 language versions), checked the identifiers in structured data (hence the actor slip), and built a site and a blog in 27 languages with Markdown versions. Reach, that is publications and mentions in other people's media, comes only after that.

I do not know yet whether it works. I planned the first repeat measurement of AI answers from November 1, 2026, and the result of the Markdown file measurement will be available after November 8. I will add the results to this post, even if they come out at zero.

Frequently asked questions about AI SEO and GEO

Where to start?

The first three steps cost nothing but time. Check whether AI search bots get your site. Ask three times at once a few questions that your customers ask, without your brand name, and write down who is named and who is cited. Read what Google results for your brand name alone say about your brand, because models use the same sources. If the result surprises you, it is a good topic for a consultation.