Generative engine optimization: how AI reads your website, explained for a CEO

Generative engine optimization, usually shortened to GEO, is the practice of making a company’s content easier to find, understand and cite in AI answers, including those in ChatGPT, Gemini, Perplexity and Google. SEO aims for visibility in search, while GEO also considers how information appears in generated answers.

TL;DR. AI bots may collect pages for training or for search. In our measurements, special files such as llms.txt were accessed almost exclusively by bots associated with training. Those accesses do not prove that the texts were actually used in a model.

When an assistant searches the web for a customer’s question, the relevant pages it finds and outside sources supporting the information about the company matter.

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, what Google says about special files and how to check the information an assistant uses when answering questions about your company.

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. In their tests, some changes to content increased measured visibility by up to 40%; the result depends on the methods and test set used.

In those tests, citing sources, adding quotations and including statistics performed well, while forced keyword repetition scored worse than the original text. The results suggest ideas to test without guaranteeing the same improvement for every site.

Where research begins matters too. G2 published a survey of 1,076 software buyers in April 2026: 51% said they start their research with an AI assistant more often than with Google. Another 69% reported choosing a different vendor from the one they had planned on after an assistant’s recommendation. These are survey responses, not independently observed purchases.

The G2 sample consists of software buyers, so the figures do not describe every B2B purchase. For some of these buyers, conversations with AI assistants are already part of supplier research. The survey does not show that buyers have stopped visiting company websites.

The founders we talk to put the question more simply than the research does. We often hear the question “Do we show up in the market? Are we visible?” Someone who asks this wants an answer in numbers, and part of those answers now sits in conversations with AI assistants.

How does AI read your website?

AI assistants may receive information about your site through systems that collect or search pages. Bots are the programs making requests to the site. Some collect content that may be used for training, while others are associated with search. A request in a server log does not show which passage, if any, reached a user’s answer.

Collecting text for training is different from using it in a model and does not guarantee when it might influence an answer. Web search can bring information into the current response. OpenAI documents separately GPTBot, for content that may be used in training, and OAI-SearchBot, for ChatGPT search.

The figure shows what to check when an assistant uses web search. It may omit a relevant page, and including a page among its sources does not guarantee a recommendation. Google rankings and AI answers need to be observed separately.

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 AI-bot crawling in July 2025 served training and only 17% served search. These shares describe the purpose of the accesses without measuring the use of texts in a model or customers acquired later.

Is SEO still worth it in 2026?

SEO is still worth investing in during 2026. An accessible site and pages that answer questions clearly can be found through conventional search and by assistants that search the web. Track appearances in AI answers separately from Google rankings and clicks, because these systems do not always select the same sources.

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.

Browsing data show a difference worth tracking. Pew Research Center analyzed the browsing of 900 American adults in 2025: 8% of visits with an AI summary included a click on a conventional result, compared with 15% of visits without one. The comparison observes behavior across two kinds of searches; it does not isolate the summary’s effect for each query.

In its updated study of AI Overviews, Ahrefs analyzed about 300,000 desktop keywords using data from December 2025. For queries in its sample with an AI summary, it estimated a click-through rate for the top organic result 58% below the value expected in its comparison. This is an association in that sample, not a measure of how much traffic any particular company would lose.

These data suggest that ranking position alone does not tell you how many visits a site receives. For your own company, track rankings, clicks and relevant appearances in AI answers separately before attributing a traffic change to generated summaries.

In Romania, where our clients are, the use of generative AI was below the EU average. According to Eurostat, in December 2025, 17.8% of people aged 16 to 74 had used such a tool in 2025, the lowest share in the European Union, where the average was 32.7%.

Eurostat measured the use of AI technologies in businesses separately.

Eurostat’s statistics on the use of AI in enterprises show that, in 2025, 5.21% of Romanian companies with at least ten employees used an AI technology, against a European average of 19.95%. These figures describe business adoption of AI, not how often Romanian buyers ask assistants for recommendations or how likely a brand is to appear in their answers.

Is GEO better than SEO?

For Google’s AI search features and assistants that search the web, access to a site’s pages still matters. SEO work can therefore help with visibility in AI answers. Selection methods vary by product, and a strong position in conventional search does not guarantee a mention in a generated answer.

Outside sources can add to what a company says about itself. When checking an AI answer, look at whether it cites articles, reviews, associations or partners, and whether they support the claim. Absence from one answer has several possible explanations; the table does not establish a lack of external confirmation.

In our methodology, authority transfer describes the work of documenting a company’s activity in credible sources, including verifiable partnerships and mentions. Association with a known name does not automatically transfer trust or guarantee an appearance in an AI answer. What the sources actually say about the company matters.

Founders in markets where everyone knows everyone see recommendation differently. Some tell us that in their industry business is won through acquaintances and word of mouth: “With us it sells through recommendations, and a website or an ad changes little.” If recommendation is their channel, it is worth knowing what an AI assistant says when a new decision-maker, who does not know them yet, asks about the category.

Does llms.txt help?

Our measurements do not establish whether llms.txt helps a company appear in answers. On our sites, it was read almost exclusively by bots that collect text for training. It is a site summary proposed in 2024 as a voluntary standard. Accessing the file and using its information in answers are different things.

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 added a reference to llms.txt in robots.txt. The next day, requests to llms.txt went from zero to 32. This sequence is an observation from our measurement, without proving on its own that the reference caused every access. Between 11 and 15 September we also measured who was reading it.

Of the 127 requests, 82% came from ClaudeBot, Anthropic’s training bot. None came from a bot classified in this measurement as a live-search bot. The result describes accesses in the observed period without measuring the file’s influence on answers.

An Ahrefs study from May 2026 started with about 137,000 domains; around 38,000 had valid llms.txt files, and 97% of those received no request in the month studied.

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 compared llms.txt with the old keywords meta tag in a discussion reported by Search Engine Journal. Google Search does not use this summary as a special signal; how other assistants handle it needs to be checked separately.

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?

A company may appear in Google for one query and be absent from an assistant’s answer to a different question. Compare the wording of both questions, the sources the assistant used and the date of the test.

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 observed a wording difference. ChatGPT cited pages with “Enduro” in the title, while NordicaMoto’s home page title was “Moto store”. NordicaMoto was missing from all three buying questions tested. The title wording is a possible explanation to investigate, without being established as the sole cause.

AI answers vary from one run to the next, so each test of this kind describes that particular day.

In our methodology, the central attribute is the idea we want people to associate with the offer. We express it consistently in descriptions, offers and evidence, then check whether buyers understand it. The same clarity may help an assistant identify the company’s activity without there being a mandatory word the model searches for.

At Doctor SKiN, the audit also identified technical rules to check. The dermatology clinic network had 2,970 Google reviews with a 4.85 rating, and its robots.txt blocked nine AI bots by name, including 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. These observations do not establish the cause of the absence. Google-Extended does not control appearance in Google’s AI Search features; Googlebot rules must be examined separately.

The same audit found an llms.txt file of 12,516 bytes, published the day before, without a reference in robots.txt and intended for the bots it blocked. This calls for checking the rules applicable to each bot and access path.

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.

That test did not name Brădet despite its substantial revenue. Separately, the Lactate Brădet audit counted a 127-to-1 ratio between citations of the category leader and Brădet, aggregated across five AI engines. Neither observation shows how much company size affected the answer.

Another audit shows that Google results and an assistant’s answer can rank companies differently. In a test on 2 September 2026 for Cablero, a steel cable manufacturer from Iași, ChatGPT named the company first among Romanian producers, while Google displayed it in tenth place for the query tested. The Cablero audit found a 2021 article in Ghidul Primăriilor among the three sources ChatGPT cited. A citation alone does not explain the order of the answer.

What does an AI model actually read on your page?

Readers can use a page more easily when it clearly says what the company does, for whom and what supports its claims. Passages with enough context can also stand on their own, but there is no universal rule that an AI assistant will select a particular format.

In practice, check whether headings and explanations help people find an answer, whether figures have sources and whether real customer questions receive answers. Use buyers’ language where it fits naturally; assistants can understand synonyms without the same phrase being repeated in every section.

In the Therezia Prodcom audit, for a dairy producer from Pănet, we proposed pages answering buyers’ questions. On 10 September 2026, Ahrefs reported two citations of the Therezia domain and 128 for Mirdatod, using its subdomain setting for Romania. This snapshot does not show whether the proposed pages would change the citation count.

In our methodology, we use AI as a cognitive processor in internal work: we give it clear sources and check what it does with them. An external assistant uses public information it can access; when that information is missing, it may cite other sources, omit the company or give an incomplete answer.

AI visibility is another place to observe whether public information about the company is represented accurately.

In our methodology, cognitive ownership aims for people to associate the company with a relevant benefit when they need it. Appearance in an AI answer may tell us something about available public information, but it does not directly measure what buyers remember. We explain the concept in our article on cognitive ownership.

At which layers is AI visibility solved?

To investigate AI visibility, check access to pages and answers to buyer questions, as well as how the company describes its offer. Our methodology places these checks across four layers, from L1, performance marketing, to L4, positioning and category, with the AI Brain on L3 for organizing internal information.

The figure is a list of checks; it does not establish in advance why a company is missing from an AI answer.

L1 · Performance marketing. Check infrastructure, including robots.txt, indexing and page speed. Access rules need examining separately for search and training. At Doctor SKiN, the audit found a restriction to investigate without proving that it alone explained the lack of citations.

L2 · Revenue and commercial process. Record the questions your sales team hears and answer them clearly on the site where the information can be shared publicly. The pages help buyers and may be found by assistants that search the web.

L3 · AI Brain, orchestration. The team organizes information and sources for its own AI assistants. This internal knowledge base does not automatically reach external assistants; information that can be shared needs to appear on accessible public pages too. I wrote about internal sources in the article on AI in marketing as a cognitive processor.

L4 · Positioning and category. Check whether the offer description says who benefits, in which situation and with what evidence. Clear wording may help readers and systems using public pages without guaranteeing that an assistant will name the company.

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. Ask three questions in order and note the date, sources and whether the company appears. Then check /robots.txt and the home page title.

1. The mission. Ask “What does [company name] do?”. Check whether the model knows who you are and whether its description is correct.
2. 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.
3. The category. Ask “Which companies offer [the relevant category] in [your country]?”. Record whether you appear, which sources are cited and whether the offer is described accurately.

After the three questions, check /robots.txt rules for OAI-SearchBot and PerplexityBot separately from those for GPTBot, ClaudeBot and Google-Extended. Read the User-agent group and the path in Disallow, including any Allow rules. A training restriction does not automatically block search. For AI features in Google Search, Google points to Googlebot rules, separately from Google-Extended.

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 home page title shown in the browser tab and check whether it clearly names the products or services buyers search for. The title is an important reference, but an assistant can also find the information in other passages or sources.

If you are absent from the answers, check technical access, the available information and its relevance to the buyer’s question. One test does not establish the cause. Our CEO’s guide to brand positioning covers how to clarify the reason to choose the company.

To start a self-assessment, you can use our Romanian-language brand diagnostic. It gives a score from 0 to 100 in about two minutes based on your answers; it does not directly measure market perception.

Files for AI may be useful, and generating them costs little. Before attributing a commercial effect to them, check what information is public, which sources the assistant uses and how it answers buyers’ questions.

What founders tell us about visibility

When we talk with founders about visibility, the same three situations come back, in very different sectors.

The first is uncertainty about the market. Founders ask us directly where they stand against the others: “The market has filled up with brands and we do not know where we sit among them.” Behind the question is the fear of being overlooked without finding out in time.

The second is the wish to be more visible, said without jargon: “Anything that makes us more visible and grows our market share interests us.” The need behind it is the one the test below measures, whether the company’s brand reaches the person who asks.

The third comes from closed markets: “Here everyone knows everyone, and what matters travels by word of mouth.” The worry behind it is simple, what happens when the person who knows them leaves.

The fifteen-minute test below answers the first question, at least in part, because it shows whether an AI assistant names you when someone asks about your category.

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 visibility and traffic from search, while GEO also considers how information appears in AI answers. For Google Search and assistants that search the web, accessible, useful pages can help. A strong position in results does not guarantee an AI mention.

Does the llms.txt file help you appear in ChatGPT?

Our measurements do not prove an effect on appearance in ChatGPT. During the period observed on our sites, 91% of requests came from bots associated with training and none from the bots we classified as live-search bots. Ahrefs found that 97% of valid llms.txt files in its sample received no request in May 2026; Google says it does not use them in Search.

Why doesn’t my company appear in ChatGPT?

Check access for search bots, the clarity of the company’s description and pages answering buyers’ questions. Also examine the outside sources cited in the answer. These are checks to investigate, rather than a diagnosis based on a single test.

How do I check whether ChatGPT knows my company?

Ask ChatGPT, with web search turned on, what the company does, who supplies your main product and which companies offer the relevant category. Record the appearance, sources and test date, then check search rules for OAI-SearchBot separately from training rules for GPTBot.