Generative engine optimization, usually shortened to GEO, is the practice of making a company’s content easy for AI engines to find, understand and cite. They include ChatGPT, Gemini, Perplexity and Google’s AI summaries. Classic SEO aims for a position in the list of results. GEO aims for the company’s name to appear inside the answer the model generates.
TL;DR. An AI model reads your website in two ways, once when it learns, during training, and again when it searches live, at a person’s question. Special files for AI, such as llms.txt, are read almost only in the first case.
The answer your customer receives depends on the second reading, meaning pages that answer questions clearly and outside sources that confirm what you say about yourself.
Much is written on this subject, usually as a list of technical tricks. I work on systems and execution, so I preferred to measure. Since August 2026 we have built a full layer of files for AI agents on our own sites, and then watched the Cloudflare logs to see who reads them.
This article shows what we found, why Google openly says special files are not required, and what a model reads when it decides whom to recommend.
Some of the data comes from Romania, where our clients are, with the EU averages alongside for comparison. At the end there is a fifteen-minute test you can run on your own company today.
What is generative engine optimization?
Generative engine optimization is the optimization of content for engines that generate answers instead of lists of links. Its goal is that an AI model understands correctly what the company does and names it when a customer asks about its category. The term often appears next to answer engine optimization (AEO), which describes almost the same thing.
The term comes from an academic paper. Pranjal Aggarwal and his colleagues published the study “GEO: Generative Engine Optimization” in 2023, presented at the KDD conference in 2024. They showed, in lab tests, that some content changes raise visibility by up to 40%.
The three methods that worked best were citing sources, adding quotations and adding statistics. The classic method of stuffing keywords scored 10% worse than the original text. Models prefer content that can be verified.
Where the search happens matters too. G2 published a survey of 1,076 software buyers in April 2026, and 51% of them start their research in an AI assistant instead of a search engine. On top of that, 69% chose a different vendor than the one they had planned on, after the assistant’s recommendation.
The G2 sample consists of software buyers, so the figure does not describe every B2B purchase. The direction is clear all the same. The shortlist of vendors is starting to form in a conversation with an AI model, before the person visits any company’s website.
How does AI read your website?
An AI model reads your website through bots, meaning programs that download pages automatically. Some bots collect text to train the model, and others search live, at the moment a person asks a question. The two readings run at different speeds, open different files and have different effects on the answer your customer receives.
Reading for training works slowly. Text collected for training enters the model’s memory only with its next version, which means months later. Live reading works in seconds, and the model uses it to complete what it knows with fresh pages found on the web for that specific question.
- The questionThe customer asks for a supplier for a concrete problem
- The live searchThe search bot looks for pages about that category
- Source selectionThe model keeps the pages that answer clearly and are confirmed by others
- The answerThe names that appear in the sources reach the recommendation
The red link is where the recommendation is won or lost. The model chooses among the pages it found, and a page that does not answer the question stays out, however well it ranks in Google.
On our own sites we saw how different the two families of bots are. In the last days of August 2026, the Cloudflare logs showed active, among others, the Meta bot, GPTBot and OAI-SearchBot from OpenAI, the Anthropic bots, Applebot, Amazonbot and PerplexityBot. Each has a different role, and some of them never send a visitor back.
This imbalance has been measured at scale. Cloudflare published a report in 2025 on the gap between crawling and clicks, and in July 2025 the Anthropic bots crawled about 38,000 pages for every visit they sent back. For OpenAI the ratio was about 1,091 to 1, and for Perplexity about 195 to 1.
Cloudflare notes that the ratio may be overstated, because apps do not always report where a visitor comes from.
The same report shows that 79% of the crawling done by AI bots in July 2025 served training, and only 17% served search. Most of the attention your site gets from AI does not produce a customer today. It builds the model’s long-term memory, which matters, but on a different time horizon.
Is SEO still worth it in 2026?
SEO is still worth it in 2026, because AI engines draw their sources largely from the same indexed pages as classic search. What changes is the measured outcome. Clicks fall while the mention in the answer grows in importance, so a well-indexed site is the GEO baseline. SEO is not dead, but it is no longer enough.
Google says so explicitly in its official guide to optimizing for AI features in Google Search. “While terms like Answer Engine Optimization (AEO) or Generative Engine Optimization (GEO) are common online, many suggested ‘hacks’ aren’t effective or supported by how Google Search actually works.”
Google’s recommendation remains original content, written by experts, that adds value beyond what is already known.
Clicks do fall visibly, though. Pew Research Center analyzed the browsing of 900 American adults in 2025, and those who saw an AI summary clicked a classic result in 8% of visits. Without an AI summary the share was 15%, almost double.
Ahrefs reached a similar conclusion on data from December 2025, in its update of the study on how AI Overviews affect clicks. On queries with an AI summary, the click-through rate of the first organic result fell by 58%.
- 51%of software buyers start their research in an AI assistantG2, 2026
- 8% vs 15%visits with a click on a result, with and without an AI summaryPew Research, 2025
- −58%click-through rate of the first result when the AI summary appearsAhrefs, 2025
Sources · G2, The Answer Economy, 2026 · Pew Research Center, 2025 · Ahrefs, December 2025
The red card describes the most uncomfortable position for a company that has invested years in SEO. You can stay first in Google and receive less than half the clicks of the past. The new measure is whether the model names you in the answer, and the position in the list becomes only a condition.
In Romania, where our clients are, the change arrives later than in other markets. According to Eurostat, in December 2025, only 17.8% of Romanians had used a generative AI tool in 2025, the lowest share in the European Union, where the average was 32.7%.
The gap is even wider among companies.
Eurostat’s statistics on the use of AI in enterprises show that, in 2025, only 5.21% of Romanian companies with at least ten employees used an AI technology, against a European average of 19.95%.
For a company in a market that is still catching up, the delay is a window of time. The company that moves ahead of its competitors takes the space in AI answers first.
Is GEO better than SEO?
GEO is built on top of SEO, because AI models choose their sources from pages that a search engine has already found and understood. The two aim at different outcomes and reward different qualities of content, and the difference shows best when they are set side by side.
| Criterion | Classic SEO | Generative engine optimization |
|---|---|---|
| What you win | a position in the results list | a mention in the generated answer |
| What you measure | position, clicks, traffic | how often and in what context you are named |
| What the engine rewards | page relevance and links to it | clear answers, figures with sources, outside confirmation |
| The unit of work | the page | the passage of a few sentences that can be cited on its own |
| Where it is decided | on your site and in the links to it | on your site and in what others say about you |
The last row explains why a company can rank well in Google and be absent from AI. The model looks for confirmation, and confirmation comes from sources you do not directly control, from press articles and reviews to associations, partners and industry directories.
In our methodology the mechanism has a name, the transfer of authority in search and AI engines. A new or little-known company associates itself, through articles, mentions and declared entities, with names the models already know, and the trust of those names gradually reaches it too.
Does llms.txt help?
The llms.txt file helps a little and indirectly, and on our sites it was read almost only by bots that collect text for training. It is a summary of the site written for AI models, proposed in 2024 as a voluntary standard. None of the major AI platforms has publicly committed to using it for live answers.
Our measurement had three stages. Between 20 and 27 August 2026, llms.txt received zero requests in eight days, on both domains, even though it was published. In the same period robots.txt was requested 153 times and the sitemap 97 times, which showed that the bots were visiting the site but not opening the file.
On 31 August we declared it in robots.txt, the file every bot opens. The next day, requests to llms.txt went from zero to 32. The practical lesson is that a file for agents gets read because it is declared, not because it exists. Between 11 and 15 September we also measured who was reading it.
Source · our own measurement, Cloudflare logs, unrivals.ro, unrivals.com and unrivals.ai
Of the 127 requests, 82% came from ClaudeBot, Anthropic’s training bot. None came from a bot that searches live for a person’s question. The file feeds the model’s slow memory and is missing from today’s answer.
Our conclusion matches the large-scale measurements. Ahrefs analyzed llms.txt files on 137,000 domains in May 2026, and 97% of the valid files received no request that month.
Google also says, in the guide cited above, “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.”
John Mueller of Google gave the shortest verdict in a discussion reported by Search Engine Journal. “To me, it’s comparable to the keywords meta tag – this is what a site-owner claims their site is about.” A model that can read the site directly has little reason to trust the summary.
We kept the file anyway, and we kept the markdown versions of the pages too. Their cost is almost zero, and in the August 2026 pilot the home page dropped in its markdown version from 159 KB to 16.2 KB, so a bot has almost ten times less text to download. They are a convenience for bots, and the real lever sits elsewhere.
Why doesn’t ChatGPT recommend you when Google ranks you first?
ChatGPT can ignore a company that Google puts in first place, because the model looks for pages that answer the customer’s exact question and use the words of the category. A company described with a general term loses to one described with the term the customer uses.
The clearest case from our audits is NordicaMoto, a distributor of enduro gear from Odorheiu Secuiesc. In the test of 29 September 2026, asked “Where do I buy enduro gear in Romania?”, ChatGPT recommended BBMoto, MxEnduro and Moto24, and NordicaMoto was missing, even though it ranked first in Google for “enduro helmet”.
The NordicaMoto audit found the cause in a single line. ChatGPT cited pages that had “Enduro” in the title, while the title of the NordicaMoto home page was “Moto store”. The model recommended the companies that described themselves with the customer’s word. NordicaMoto was missing from all three buying questions we tested.
AI answers vary from one run to the next, so each test of this kind describes that particular day.
This is where the principle we call the central attribute comes in, the rule in our methodology that a single word or short phrase must run through the name, slogan, offer and proof. For people, the central attribute makes the company easy to remember. For an AI model, it is the word that links the company to a category.
At Doctor SKiN the cause was technical. The dermatology clinic network has 2,970 Google reviews with a 4.85 rating, and its robots.txt blocked nine AI bots by name, among them GPTBot, ClaudeBot and Google-Extended. On Gemini, the Doctor SKiN audit found zero citations, against 26 for the category leader, according to Ahrefs, on 10 September 2026.
The same audit also found an llms.txt file of 12,516 bytes, published the day before, not declared in robots.txt and addressed to exactly the bots that the same robots.txt blocked. Someone did good work that could not reach anyone. Good content first needs the infrastructure that lets it be read.
- The category leader26
- Doctor SKiN, with AI bots blocked0
Source · Ahrefs, citations in AI answers, Romania, taken from the Doctor SKiN audit
The pattern repeats among producers. On 7 September 2026, asked about traditional smoked cheese, ChatGPT named 17 producers, including one with a single Google review, and did not mention Lactate Brădet, a company with a turnover of RON 34.6 million.
The size of the company did not matter to the model. The Lactate Brădet audit measured a gap of 127 to 1 between the first domain in the category and Brădet.
There is also the opposite case, which shows that the mechanism works in both directions. At Cablero, a steel cable manufacturer from Iași, ChatGPT put the company first among Romanian producers, while Google put it in tenth place. The Cablero audit found, among the three sources ChatGPT cited, an article published in 2021 in the Romanian trade magazine Ghidul Primăriilor, which supported the top position in AI.
What does an AI model actually read on your page?
An AI model looks on your page for short passages that answer a question on their own, figures with sources and clear wording about what the company does and for whom. A passage that can be cited without context wins, and text that requires reading the whole page to be understood gets lost.
On the page, this means four things. The first sentence of every section answers the question in its heading directly. The term the customer searches for appears in the title and in the first lines. Figures have their source next to them, and the real questions customers ask each have a written answer on the site.
For Therezia Prodcom, a dairy producer from Pănet, the audit framed the problem in exactly this way.
The company had a net margin of 11.51%, against 5.2% for the category average, and it was cited twice in AI answers, against 128 times for a processor in the same county. The move proposed in the Therezia audit was a series of pages that answer the questions asked before buying, so the machines have something to cite.
There is one more layer, the quality of the information the model receives.
In our methodology, AI works as a cognitive processor, an extension of thinking, and its quality depends on how precisely you tell it where to think. A model fed vague wording about a company produces vague answers about it. When information is missing, the model fills it in from what it finds about others.
This is also where it becomes clear why AI has become the new home of cognitive ownership.
The principle in our methodology says the final goal is to own the mental space of the category, so that when the need appears, the market comes to you automatically. When the need is phrased as a question to an AI model, cognitive ownership means the model names you first. The concept is explained in our article on cognitive ownership.
At which layers is AI visibility solved?
AI visibility is solved at the positioning layer, even though the first signs appear in technical files and on-site pages. That is why we read each problem across the four layers of our methodology, from L1, performance marketing, to L4, positioning and category, with the AI Brain on L3.
- L4 · Positioning and categoryThe company describes itself with a general term, so the model associates it with another category
- L3 · AI Brain, orchestrationThe information about the company is vague or scattered, so the model fills it in from other sources
- L2 · Revenue and commercial processThe questions asked before buying have no written answer on the site
- L1 · Performance marketingBots are blocked or pages are not indexed correctly
The figure reads from the bottom up, in the order in which a company discovers the problem, while the red floor at the top shows where it begins.
L1 · Performance marketing. The first thing to check is the infrastructure, meaning robots.txt, indexing and page speed. Our principle of infrastructure before marketing says that marketing built on a poor foundation burns money, and the Doctor SKiN case shows that even the best content stays invisible when the bots are blocked.
L2 · Revenue and commercial process. The questions a customer asks before buying are the same ones they ask an AI model. A page written for them works twice, for salespeople and for bots.
L3 · AI Brain, orchestration. The information about the company must exist in a precise and consistent form, from the name of the category to the key figures. How to build such a knowledge base I showed in the article on AI in marketing as a cognitive processor.
L4 · Positioning and category. The problems on the lower layers get fixed quickly, in days. The top layer decides whether the model has a reason to name you. Without a category word the company can own, even the best pages describe a company the model places next to all the others.
How do you test in 15 minutes how AI sees you?
The test takes about fifteen minutes and is done in ChatGPT, with web search turned on, and then on your own site. You ask three questions in order, note for each whether the company appears and in what position, then check two files and the title of the home page.
- The mission. Ask “What does [company name] do?”. Check whether the model knows who you are and whether its description is correct.
- The capability. Ask “Who supplies [your main product or service] in [your country]?”. Check whether you appear or whether the model names competitors in your place.
- Your own category. Ask “Who are the leaders in [your field] in [your country]?”. This is the question you should win, and being absent here is the most expensive.
After the three questions, open your site’s address followed by /robots.txt and look for the names GPTBot, ClaudeBot, PerplexityBot and Google-Extended. If they appear next to the word “Disallow”, those models are not allowed to read your site.
The decision can be a deliberate one, but it has to be made by you, knowingly, because sometimes the block comes from a plugin installed long ago and forgotten.
Finally, read the title of the home page, the one shown in the browser tab. If the title does not contain the word customers search with, the model has no way to learn it from you. That word is, most often, the name of the category you want to own.
If one of the three questions fails, the gap is not technical. It shows that the market and the models do not yet have a clear reason to associate you with your category, and our CEO’s guide to brand positioning covers how to build that reason.
If you want to see how your company looks from the outside, on its public figures, you can ask for a diagnostic.
Files for AI remain useful as a convenience, and their cost is small. The recommendation your customer receives is decided, however, in the pages that answer their questions and in the sources that confirm, from outside, what you say about the company.
Frequently asked questions
What does GEO mean in marketing?
GEO stands for generative engine optimization, the optimization of content for AI engines that generate answers, such as ChatGPT, Gemini, Perplexity or the AI summaries in Google. The goal is that the model understands the company correctly and names it when a customer asks about its category.
What is the difference between GEO and SEO?
SEO aims for a good position in a search engine’s list of results, while GEO aims for a mention in the answer generated by an AI model. GEO is built on top of SEO, because models choose their sources from indexed pages, but it rewards clear answers, figures with sources and outside confirmation more.
Does the llms.txt file help you appear in ChatGPT?
The llms.txt file helps a little and indirectly. On our sites, 91% of the requests to it came from training bots and none from live search bots. Ahrefs found that 97% of llms.txt files received no request in May 2026, and Google says it does not use them.
Why doesn’t my company appear in ChatGPT?
The most common causes are AI bots blocked in robots.txt, a description of the company in general terms instead of the word the customer uses, and a lack of pages that answer the questions asked before buying. Sometimes the model finds the company but prefers competitors confirmed by more outside sources.
How do I check whether ChatGPT knows my company?
Put three questions to ChatGPT, with web search turned on: what the company does, who supplies your main product, and who the leaders of your category are. Note whether you appear and in what position, then check in robots.txt whether GPTBot, ClaudeBot or PerplexityBot are blocked.
