For years, when a client asked me how to improve their website’s visibility, the starting point was fairly straightforward: understand what people were searching for on Google, look at which websites were ranking for those searches, and develop a strategy to compete.
That approach still makes sense, but it no longer tells the whole story.
Someone choosing hotel management software, looking for accommodation or comparing service providers might now begin their research with ChatGPT, Gemini or Perplexity. Rather than entering a few keywords, they describe their problem, specify their requirements, ask for comparisons and often continue the conversation with follow-up questions.
The result is not necessarily a list of websites. It may be an organized answer in which some brands are mentioned, certain sources are cited, and many others are left out altogether.
For anyone working in search marketing, this raises an important question: how can we make a website, a piece of content or a company relevant enough to become a source for these systems?
This is where GEO, or Generative Engine Optimization, comes in.
Unfortunately, the term is already attracting the kind of noise we have seen many times in SEO: supposedly essential tools, universal checklists and promises that a few quick changes will make a brand appear in AI-generated answers.
I would rather start with what we actually know, and distinguish it from what we can only speculate about.
What GEO is, and why it is not quite the same as SEO
Generative Engine Optimization refers to the practices aimed at improving the presence and use of content in answers generated by AI-powered search systems.
It sounds very similar to SEO, but at least part of the desired outcome is different.
In traditional search, we mainly work to ensure that a page can be discovered, indexed and ranked for relevant queries. With generative systems, we are also interested in whether the information on that page is retrieved, considered relevant and potentially cited in the final answer.
That distinction matters because being included in a search index, being used to generate an answer and receiving an explicit citation are three different things.
A website may be properly indexed yet never appear in AI-generated answers to a particular set of questions. Conversely, a brand may be mentioned without a link to its official website.
None of this means SEO has become less important. Quite the opposite: many of the requirements for making information accessible and understandable remain unchanged.
What changes is the scope of the analysis. We need to look beyond rankings and consider how information is retrieved, interpreted and reused.
How ChatGPT, Gemini and Perplexity generate answers
One common mistake is to treat ChatGPT, Gemini and Perplexity as conventional search engines with a different interface.
That is not always how they operate.
A language model can generate answers using knowledge acquired during training. Certain systems and modes can also search for information, retrieve documents and use external content to produce more up-to-date responses.
In these cases, the underlying process can often be understood, at a general level, through the concept of Retrieval-Augmented Generation, or RAG.
The basic principle is straightforward:
- A user submits a question.
- The system determines how to respond and, where supported, retrieves information from external sources.
- Retrieved documents are selected and used as context.
- The language model generates an answer using the available information.
- The interface may display links or citations to supporting sources.
This is a simplified explanation, not a technical description of every product’s internal architecture. Different platforms use different retrieval, selection and generation mechanisms. These can change over time and are not always publicly documented.
For GEO, however, it highlights something important: the system must not only find our content; it must also consider that content useful for answering the user’s question.
Consider a query such as:
“I’m looking for management software for a small independent hotel, with an integrated booking engine, rate management and connectivity to major booking platforms. Which solutions should I consider?”
A page titled simply “Hotel Management Software” might be relevant, but it may not contain enough information to support a useful answer.
A page that clearly explains features, available integrations, supported property types, limitations and selection criteria provides much more useful material.
There is no guarantee that the second page will be selected, but its value goes well beyond containing the right keyword.
Before optimizing content, make sure the website is accessible
When discussing GEO, many people start with content. I would start with the technical infrastructure.
A page may contain excellent information, but if the systems expected to retrieve it cannot access it, its editorial quality becomes a secondary issue.
My initial checks would be very similar to those in a technical SEO audit:
- Correct HTTP responses for important pages.
- No accidental restrictions in
robots.txt. - Proper use of
noindexand snippet directives. - Important content available in HTML accessible to crawlers.
- A clear internal linking structure.
- Consistent canonical URLs and unnecessary duplication kept under control.
- No unintended restrictions introduced by CDNs, firewalls or bot protection systems.
That final point deserves particular attention. Security configurations sometimes block legitimate crawlers, especially when broad anti-bot rules have been implemented without reviewing their impact.
OpenAI crawlers: OAI-SearchBot and GPTBot
OpenAI distinguishes between crawlers used for search and crawlers used to collect content that may be used for model training.
OAI-SearchBot is associated with ChatGPT’s search capabilities, while GPTBot is used for collecting content for potential model training.
These are separate purposes, and website owners can manage them independently.
For example, the following configuration allows the search crawler while disallowing GPTBot:
User-agent: OAI-SearchBot
Allow: /
User-agent: GPTBot
Disallow: /This is not a configuration that every website should copy. It simply demonstrates the distinction between allowing content to be accessed for search and allowing its use for training.
The appropriate settings depend on the publisher’s policies and the project’s objectives.
More importantly, allowing crawling does not guarantee a citation. It merely removes one potential technical obstacle to content discovery.
Google’s AI search features, meanwhile, rely on the established technical foundations of Google Search. Google has explicitly stated that appearing in AI Overviews or AI Mode does not require special GEO markup.
Content should answer questions, not merely target keywords
This is probably the most interesting editorial shift.
For years, SEO strategies have been built around primary keywords, secondary keywords, related searches and search volumes. These remain useful, but they are not sufficient when people begin expressing their information needs as detailed questions.
Consider an example from the hospitality industry.
A hotel website may have a page optimized for “hotel near the train station”. In traditional SEO, we would analyze search volume, competition and ranking opportunities for that query.
Someone using an AI assistant might ask:
“I’m staying in Naples for one night. My train arrives late, and I need to leave early the next morning. I’d like somewhere convenient, reasonably priced and preferably with flexible check-in.”
Several factors now matter: location, transportation, arrival times, pricing, services and operational details.
If the hotel’s website only describes itself as being in a “strategic location” or offering “the ideal solution for every need”, it provides very little verifiable information.
Actual distance from the station, practical directions, check-in hours, self-check-in availability and relevant travel information are far more useful.
This is a simple example, but the principle applies across industries.
GEO also means identifying the questions we want our website to help answer, and making sure it contains the information needed to do so.
How I would structure a page
I do not believe every article needs to become a long sequence of questions and answers.
I prefer a readable editorial structure, with each section addressing a specific problem.
For example, a guide to choosing hotel management software could include sections covering:
- Essential features for a small property.
- Differences between a PMS, a channel manager and a booking engine.
- Integrations to verify before purchasing.
- Costs beyond the monthly subscription.
- Questions to ask suppliers during a product demonstration.
FAQs can complement the main content, but they should not replace a properly developed explanation.
Content quality is not determined by the number of headings. It depends on how easily a reader, human or automated, can find a relevant and sufficiently precise answer.
Sources and evidence: why verifiable claims matter
Another important characteristic of generative systems that retrieve external information is their ability to combine material from multiple documents.
This makes a frequently overlooked aspect of SEO content particularly relevant: whether claims can be verified.
Consider two pieces of copy.
The first describes a software product as “the most innovative solution on the market, used by hundreds of companies and capable of dramatically improving productivity”.
The second explains its actual features, identifies supported integrations, documents usage conditions and, when reporting measurable results, provides the relevant context.
The second gives readers considerably more information to assess.
This is one reason I would avoid building a GEO strategy around heavily promotional copy. A page full of superlatives but short on useful facts does little for users or for systems attempting to answer their questions.
Academic research on GEO offers some interesting observations.
In GEO: Generative Engine Optimization, presented at KDD 2024, researchers tested different content modifications and observed improvements in visibility under specific experimental conditions. The methods examined included adding relevant statistics, citations and verifiable information.
The authors reported visibility improvements of up to 40% within their experimental framework.
That figure is frequently repeated as though adding statistics to a page could automatically increase its visibility on ChatGPT by 40%.
It does not mean that.
The findings relate to a particular benchmark and evaluation methodology. They do not establish that a real website will gain citations across all generative systems, or that increased visibility will necessarily produce more traffic or revenue.
My takeaway is more cautious: how we structure and support information can influence its use in generative systems, but the actual effect needs to be tested in each context.
Authority: what a business says about itself is not enough
There is another factor worth considering.
A company may explain its services extremely well on its own website, but that does not automatically make it an authoritative source when users ask for comparisons with competitors.
Online authority is not built exclusively through owned content.
Industry publications, editorial mentions, public documentation, reviews and consistent information about an organization can all contribute to its broader digital presence.
For this reason, I would combine website optimization with an assessment of the brand’s external visibility.
Is the information accurate and up to date? Are services described consistently? Are there independent sources that document the company’s activities and expertise?
This should not become a campaign to secure as many mentions as possible on low-quality websites. The old quantity-driven link-building approach was already questionable in traditional SEO, and there is little reason to reintroduce it under a new acronym.
The objective is to build a consistent body of reliable, identifiable information.
External mentions do not guarantee that a specific AI system will recommend a brand, but they are a reasonable component of a broader search visibility strategy.
Structured data and llms.txt: what is actually useful?
Whenever a new optimization discipline emerges, supposedly essential technical solutions tend to follow.
In GEO, two recurring topics are structured data and the llms.txt file.
Schema.org
Schema.org structured data provides a shared vocabulary for describing information in a machine-readable format.
It can help identify organizations, articles, products and relevant attributes.
I consider structured data good technical practice when it is appropriate, correctly implemented and consistent with what users can actually see on the page.
I would not treat it as a shortcut to obtaining citations from AI assistants.
Google has made it clear that its AI Search features do not require special structured data beyond standard Search requirements.
llms.txt
The llms.txt proposal aims to provide language models with a concise, easily accessible representation of a website’s content, typically including links to important resources.
The idea is interesting, particularly for websites with extensive technical documentation.
However, as of August 2026, there is insufficient evidence to treat it as a universal requirement for visibility in generative search.
It may be worth testing, provided we understand that implementation does not guarantee crawling, content use or citations.
Given the choice between creating a perfect llms.txt file and improving a website with weak content, I would prioritize the latter.
How to measure visibility in AI-generated answers
This is where, in my view, many commercial GEO offerings start to show their limitations.
If we promise greater visibility, we first need to define what we are actually measuring.
With Google Search Console, we are used to working with impressions, clicks, CTR and average rankings. Generative search does not always provide metrics that are equally transparent or comparable.
Google includes appearances in its AI Search features within the broader Web performance reporting in Search Console, rather than providing a separate report identifying every generative citation.
For other platforms, measuring visibility may require structured testing and analysis of the sources displayed in answers.
A practical testing approach
Suppose we want to assess the visibility of a company selling hotel management software.
I would begin by creating a representative set of questions covering different search intents.
| Intent | Example question |
|---|---|
| Informational | What is the difference between a PMS and a channel manager? |
| Exploratory | What features does a small hotel need in management software? |
| Comparative | What are the main software solutions for independent hotels? |
| Decision-making | Which PMS should I consider for a 20-room hotel? |
| Branded | Is software X suitable for a small independent hotel? |
I would then test those questions across the selected platforms, recording at least the following:
- Whether the brand is mentioned without being explicitly named in the question.
- Whether the official domain is cited as a source.
- Which competitors appear in the same answer.
- Which external websites are cited.
- Whether the information about the brand is accurate.
- How much the results vary when questions are repeated.
A single answer should never be treated as a stable result.
Responses can change depending on wording, timing, product mode and other factors that may not be fully observable.
For meaningful comparisons, I would retain the same question set over time and introduce controlled variations where useful. I would also record the date, language, country, available settings and full response text.
A small sample cannot tell us the universal probability that an AI system will recommend a brand. It can, however, identify potential problems, highlight changes and reveal which information different systems tend to use.
The metrics I would track
I would distinguish between three levels.
Presence. How frequently the brand appears in the tested answers, keeping unprompted mentions separate from questions that already include the brand name.
Citations. How often the official website receives a link, and in what context.
Business outcomes. How many visits and conversions can be reliably attributed to referrals from AI tools, where this information is available.
A brand mention does not necessarily generate a click. A click does not equal a qualified lead, and a lead does not equal a sale.
Anyone with a performance marketing background should be particularly careful about these distinctions.
What I would do in the first 30 days
If I were setting up a first GEO project for a company website today, I would not begin by rewriting every page.
I would start with a limited, measurable scope, designed to clarify which problems we are trying to solve.
Week one: technical audit and question research. I would review crawlability, indexability, content structure and crawler permissions, while developing a first set of questions representative of prospective customers.
Week two: visibility baseline. I would test the selected platforms, record brand mentions and citations, examine external sources and identify competitors appearing in relevant answers.
Week three: priority content improvements. I would focus on a limited number of pages, improving precision, completeness, internal links, supporting evidence and the clarity of commercial information.
Week four: validation and documentation. I would check that the changes had been implemented correctly, collect any early available signals and establish a schedule for repeated measurement.
I would not promise substantial results within 30 days.
Crawling, indexing and the potential use of updated content can happen on different timelines. Even an observed change in AI-generated answers does not automatically prove that our modifications caused it.
The first month should establish a sound technical and analytical baseline, not promise a particular position in an answer by a particular deadline.
Should GEO and SEO be separate strategies?
I would not build two completely independent strategies.
SEO remains essential for making a website accessible, developing quality content, targeting relevant searches and attracting qualified traffic.
GEO adds another perspective: understanding how that information may become part of AI-generated answers and how a brand is represented in those environments.
Many activities deliver value in both contexts.
A well-written service page with accurate information, clear structure, verifiable statements and answers to real customer needs is a more dependable investment than any tactic built around assumptions about how a particular language model works.
More specialized tools and practices may emerge over time. But based on what we can currently observe, the main risk is confusing a new area of analysis with the promise of another algorithm we can manipulate.
Conclusion: become a useful source before trying to get cited
GEO is worth studying, testing and incorporating into digital visibility projects. Ignoring it simply because measurement remains imperfect would be a mistake.
But presenting it as a collection of tricks to make ChatGPT or Gemini recommend a business would be equally misleading.
My approach is relatively simple: make sure information is accessible, understand the questions potential customers are asking, develop genuinely useful content and use a repeatable method to examine where and how the brand is mentioned.
Then compare those observations against outcomes that actually matter to the business.
Ultimately, I am not interested only in getting a website mentioned in an AI-generated answer. I want to understand whether that visibility makes a business more discoverable, more credible and, ideally, more likely to be chosen.
That is how I believe this work should be evaluated, regardless of which acronym we use.
References and further reading
- Aggarwal P. et al., GEO: Generative Engine Optimization, KDD 2024. https://arxiv.org/abs/2311.09735
- Google Search Central, Top ways to ensure your content performs well in Google’s AI experiences on Search, May 21, 2025. https://developers.google.com/search/blog/2025/05/succeeding-in-ai-search/
- Google Search Central, AI features and your website. https://developers.google.com/search/docs/appearance/ai-features
- Google Search Central, A new resource for optimizing for generative AI in Google Search, May 15, 2026. https://developers.google.com/search/blog/2026/05/a-new-resource-for-optimizing
- OpenAI, Overview of OpenAI Crawlers. https://platform.openai.com/docs/bots
- The llms.txt proposal. https://llmstxt.org/