By Rafaela Pietra

If you work in SEO or Digital Marketing, you've probably come across a list of "GEO best practices." This acronym stands for Generative Engine Optimization, promising to teach you how to make a company appear in ChatGPT, Gemini, AI Overviews, or any other search mechanism based on Artificial Intelligence. 

Create an llms.txt file. Turn all headings into questions. Answer everything in short blocks. Add FAQs. Use Schema. Repeat specific entities. Post on Reddit. Earn mentions. Write "the way an AI would answer."

Some of these recommendations make sense. Some have interesting studies behind them. Others are adaptations of best practices we already knew from SEO. And some have started to be treated as rules before there was sufficient evidence to support them.

This movement, in fact, isn't exactly new. Whenever a new channel, a new technology, or a relevant change in search behavior emerges, the market first goes through a period of experimentation. Soon after, the race begins for methodologies, tools, and formulas capable of turning uncertainty into a process.

With GEO, or optimization for generative engines, it would be no different.

However, this doesn't mean GEO is just another buzzword. Optimizing a company's presence in AI-generated responses has already become a relevant front for digital visibility and tends to gain even more importance as search habits shift.

The problem starts when a still-new discipline is presented as if we already understood all its inner workings.

Today, we can identify patterns, test hypotheses, and observe characteristics that appear more frequently among content retrieved, cited, or mentioned by AI systems. What we cannot do yet is turn these signals into a new definitive list of ranking factors.

Perhaps, then, the best way to discuss GEO right now is to start precisely with this distinction: what do we already know, what does the data suggest, and what are we still trying to figure out?

GEO (Generative Engine Optimization) is the practice of increasing a brand's, content's, or information's ability to be found, understood, cited, or recommended by generative Artificial Intelligence engines. 

GEO exists. A ready-made formula for it does not yet.

The term Generative Engine Optimization gained momentum following a study developed by researchers from Princeton, Georgia Tech, Allen Institute for AI, and IIT Delhi, later published at KDD 2024.

The work formalized the concept of generative engines and analyzed different ways to increase content visibility in the responses produced by these systems. In the experiments conducted, some strategies managed to generate visibility increases of up to 40%. 

This is an impressive figure and, naturally, it became one of the most repeated references when the market began talking about GEO.

However, there is an important detail in the study itself that often disappears when its results are turned into posts, presentations, or lists of "techniques that work": the effects varied depending on the topic and the type of query.

In other words, the researchers didn't find a universal formula capable of increasing the presence of any page in any generative response. They found strategies that showed relevant results within specific experimental conditions.

It seems like a small difference, but it isn't.

Subsequent research reinforced the need for this caution. An academic review published in July 2026 analyzed 45 studies on Generative Engine Optimization produced between 2023 and 2026 and reached a conclusion far less comfortable for those looking for a cookie-cutter recipe: so far, no analyzed technique has demonstrated a stable, longitudinal, and consistent causal effect across different platforms regarding organic discovery and subsequent business results. This doesn't mean GEO doesn't work.

It means we are trying to optimize for probabilistic, shifting, and partially opaque systems.

We can find patterns. We can experiment. We can identify correlations and improve the conditions for a piece of content to be retrieved or cited. What we shouldn't do is turn every positive result into a universal rule.

There is another complication. ChatGPT, Gemini, Copilot, Perplexity, AI Overviews, and other generative environments do not work exactly the same way. A recommendation observed in a particular model, type of search, or set of queries might simply not repeat in another.

The very idea of "ranking in ChatGPT," therefore, already warrants some caution.

GEO Best Practices: When Hypotheses Start Turning into Rules

In recent months, several practices have been presented as virtually indispensable for GEO.

The logic usually starts from a real technical characteristic. LLMs work with chunks, for example. From this, the conclusion emerges that all content should be fragmented into small, independent blocks.

Or a new file emerges designed to facilitate model access to content, and quickly, it is treated as the "new robots.txt."

The problem isn't testing these practices.

The problem is the leap from "this might help in a specific situation" to "this is necessary to appear in AI responses."

In July 2026, Google itself published specific guidance on optimization for its AI research features and used the document to clarify some of these points.

The company states it does not use llms.txt files in its generative systems, that there is no ideal content size for AI, and that it is not necessary to artificially fragment texts into small blocks. Google also reinforces that there is no special Schema markup required to appear in these results. 

This doesn't mean textual organization has stopped being important, much less that structured data has lost its function.

Text divided into clear sections remains better than a wall of content without hierarchy. A FAQ remains useful when it answers real questions. Schema remains important when it helps search engines correctly interpret information.

The difference lies in understanding why these practices work.

Good structure helps because it makes the content more understandable. This doesn't mean we need to turn every article into dozens of disconnected micro-blocks because "the AI likes chunks."

Similarly, inserting five artificial questions at the end of any page doesn't automatically transform that content into a good source.

This distinction between a best practice and a supposed AI visibility factor is, in my view, one of the most important ones we need to make today when discussing GEO.

Before being chosen by AI, content needs to be found

When we clear away the hacks, we begin to arrive at a much more interesting discussion.

Generative engines themselves still depend on information retrieval systems.

Google explains that its AI features utilize traditional Search infrastructure and different mechanisms to retrieve relevant and updated information before constructing a response. Among these is Retrieval-Augmented Generation, known as RAG, which allows for searching external information and using it as a basis for generating content.

Another mentioned feature is called query fan-out. In practice, a single question asked by a user can trigger several related searches, allowing the system to gather information on different aspects of the subject before crafting a more complete response. 

Imagine, for example, someone asking what infrastructure a company needs to implement Artificial Intelligence safely.

The question is one, but to answer it, the system may search for information on data architecture, governance, security, cloud infrastructure, compliance, costs, and implementation best practices.

The final response may use different pages to support different parts of the reasoning.

This operation helps explain why I don't see SEO and GEO as two opposing disciplines.

Before a piece of content is cited, summarized, or used as a source, it must be available for retrieval. It needs to be found, accessed, interpreted, and considered relevant to some part of that need.

In other words, before being chosen by AI, the content needs to enter the set of information the AI can choose from.

For many years, the main question in SEO was how to make a specific page appear among the top results for a query.

GEO adds another layer.

Now we also need to think about how to ensure the knowledge produced by a company participates in the responses built around a particular topic.

It’s a significant shift, but it doesn't mean abandoning everything we already knew about search.

Read also: Local SEO with Artificial Intelligence: The Complete Guide

Technical SEO is still infrastructure, even for AI

There is nothing particularly exciting about reviewing robots.txt, canonicals, internal links, or indexing.

Perhaps that is precisely why the discussion on GEO started much more with hacks than with infrastructure.

But a sophisticated strategy to increase a brand's presence in ChatGPT is of little use if the relevant content is blocked, poorly indexed, isolated within the site architecture, or available in a way that makes its retrieval difficult.

OpenAI itself advises publishers who want to allow their content to appear in ChatGPT search to grant access to OAI-SearchBot. The company also makes it clear that this access does not guarantee positioning or citation, as results are selected based on various criteria of relevance and reliability. 

It’s a good example of the difference between a condition and a guarantee.

Allowing the crawler to access your site won't make a company automatically appear in ChatGPT. Blocking it, however, certainly limits the possibilities for discovery.

The same logic applies to architecture, crawlability, indexability, internal links, canonicals, performance, and content availability.

These elements don't need to be reinvented as "GEO factors." They remain important because they help create the conditions for information to be found and understood.

Nothing very glamorous.

But fundamental.

The prompt shouldn't replace the keyword

Another trend that quickly gained traction was the attempt to turn prompts into the new unit of optimization.

If we used to track keywords, now we start monitoring full questions asked to ChatGPT and observing which brands appear in the answers.

This type of analysis is useful. It helps understand how a company appears in different contexts, which topics it is associated with, and how consistently its presence is maintained.

The problem arises when we treat a specific prompt as if it were the direct equivalent of a keyword.

Generative responses have much more variability.

The phrasing of the question can change the result. The context of the conversation can change the result. The model used can change the result. The way the system searches for information can also change the result.

A study published by Semrush in July 2026 analyzed over 50,000 brands, 1,094 categories, and hundreds of thousands of citations in ChatGPT. Only 15.2% of the analyzed topics had a clearly dominant brand. More importantly, appearing well for a single question didn't necessarily mean possessing authority over that subject. Consistency began to emerge when the same brand appeared in different prompts related to the same topic. 

This aligns GEO much more with the logic of topical authority than a simple battle for prompts.

If a company wants to be associated with data governance, for example, it is likely more useful to build a truly strong ecosystem on the subject than to create a single page trying to answer exactly the question "which is the best data governance company?"

This ecosystem can involve a solid service page, articles on data implementation, quality, and architecture, content written by experts, research, technical materials, case studies, comparisons, and answers to questions that actually arise during the decision journey.

The strategic question is no longer just "which prompt do we want to appear for?"

It becomes: on which subjects do we want our brand to be recognized as a relevant source?

This shift seems subtle, but it completely changes content planning.

If AI can already summarize the consensus, your company needs to add something to it

Perhaps one of the most interesting consequences of this transformation lies in the value of proprietary content.

For years, companies competed for organic traffic by producing better versions of subjects that had already been widely explored.

An article on "seven benefits of Artificial Intelligence for business," for example, could have better on-page optimization, more comprehensive content, and a superior structure than competitors. Often, that was enough to win top rankings.

Now there is a new problem.

Generative engines can gather and synthesize dozens of versions of this same content in seconds.

Google itself currently recommends producing unique content built from experience, real knowledge, analysis, data, and original perspectives. The company also warns against content that merely reproduces widely available information or that could be generated without any specific expertise on the subject. 

The difference becomes quite evident when we place two titles side by side.

“Seven benefits of Artificial Intelligence for business.”

“We analyzed 120 Artificial Intelligence projects in B2B companies and identified the five most frequent bottlenecks before implementation.”

The first content can be produced by virtually any company.

The second depends on experience, access to data, and proprietary knowledge.

It adds information that wasn't available before.

And this might be one of the most important shifts for content strategies in the coming years.

If AI can summarize the consensus, truly valuable content becomes that which adds something to the consensus.

Original research, case studies, benchmarks, methodologies, expert experiences, and internal data gain strength because they offer something that cannot simply be reproduced by reading the same twenty articles that all competitors also used as a reference.

Clarity helps. Turning it into a GEO ritual does not.

Among the practices associated with visibility in generative systems, some make total sense simply because they make information better.

A Semrush analysis published in January 2026 compared over 300,000 URLs cited by AI engines with pages present in traditional Google results.

Among the most frequent characteristics in the cited pages were clarity, synthesis capability, signs of experience and authority, question-and-answer formats, and good organization into sections. 

This is a relevant result.

But, again, we need to use the correct word: association.

The study does not demonstrate that adding a FAQ causes a citation. Nor does it prove that every H2 needs to be formulated as a question.

A FAQ can work very well in content about cloud migration if users actually tend to have questions about timelines, costs, security, or risks. In this case, the structure helps the reader, makes important answers explicit, and organizes the information better.

However, putting five generic questions at the end of a corporate page just to "do GEO" is unlikely to turn weak content into a relevant source.

My interpretation is simpler.

If a person opens a page and can quickly understand the subject, find the information they are looking for, and identify the evidence supporting a particular claim, we have well-constructed content.

If this same clarity also facilitates interpretation and retrieval by generative systems, great.

Not every GEO best practice needs to be born from a revolutionary discovery about LLMs.

Some work because they are, first and foremost, good communication practices.

Evidence is worth more than adjectives

There is a type of phrase that appears on virtually every corporate website:

“We are a leader in innovative and personalized solutions, developed by a highly qualified team.”

It seems to say a lot and, at the same time, could be on the site of a tech company, a financial consultancy, a manufacturer, an agency, or an architecture firm.

Now compare it with another piece of information:

“Following implementation, the time required to execute the process dropped from 25 to 6 minutes.”

Here there is context, transformation, and measurable results.

There is something that can effectively support an answer.

For an engine that needs to build content grounded in available information, the second sentence offers much more value than the first.

This is why I consider case studies to be one of the most interesting opportunities within a GEO strategy.

A good case study shouldn't just state that a solution "transformed the client's operation." It needs to explain what the problem was, in what context it existed, what decisions were made, how the solution was implemented, what obstacles arose, and, most importantly, what the observed result was afterward.

This produces specific, contextualized, and verifiable information.

Not because there is some secret markup that makes ChatGPT prefer case studies.

But because evidence is more usable than adjectives.

Discover Bull's success stories!

GEO also happens off-site

SEO has always had a dimension external to the domain itself, especially when we talked about backlinks, authority, and reputation.

With GEO, this discussion seems to gain even more importance.

A brand is not just what it claims about itself on its own pages.

It is also what different sources say about it.

Recent research is beginning to find interesting relationships between brand presence on the web and visibility in generative systems. An analysis by Ahrefs involving 75,000 brands found a relevant correlation between brand mentions on the internet and presence in ChatGPT responses. Mentions on YouTube also showed a strong relationship with visibility in these environments. 

Again, correlation does not mean causation.

It would be very easy to turn this result into another recipe: get as many mentions as possible and ChatGPT will start recommending your company.

That is not what the study shows.

But there is a very reasonable strategic interpretation.

Brands that have a real digital presence, are discussed in different channels, appear in relevant publications, and are repeatedly associated with certain topics end up producing many more signals about who they are, what they do, and in which contexts they are relevant.

This causes GEO to bring together disciplines that for a long time were worked on almost separately.

SEO, Digital PR, branding, content, press, reviews, experts, partners, events, and communities become part of a larger discussion about brand presence on the internet.

A company can say on its institutional page that it is an expert in information security.

It’s another thing to have experts speaking on the topic, published cases, interviews in sector media, studies cited by third parties, participation in events, and clients reporting their experiences with the solution.

These are very different situations.

The second creates a much broader set of references about that brand.

Being a source doesn't necessarily mean building a brand

At the same time, we need to be careful not to turn "AI citation" into the new Google #1 position.

Appearing as a source is, naturally, a great sign.

However, this doesn't necessarily mean the user saw, recognized, or remembered the company's name.

The research published by Semrush in June 2026 analyzed a phenomenon called ghost citations. In 62% of the citations observed in the study, the domain had been used as a source, but the brand did not appear directly in the text of the AI-produced response. 

This data is important because it helps separate things that often get mixed up when we talk about GEO measurement.

Being found is not the same as being cited.

Being cited doesn't necessarily mean being mentioned.

A mention doesn't automatically represent a recommendation.

And a recommendation still needs to generate some kind of action to produce a business result.

A company article can be used to support a technical explanation without the brand name ever appearing in the response.

In another case, ChatGPT might recommend a company by name, but the user might not click any link because they already received everything they needed within the interface itself.

It is also possible to receive traffic from an AI tool and discover that these visits generate no commercial opportunity.

Therefore, a serious SEO and GEO strategy needs to separate different stages:

being found, being cited, being mentioned, being recommended, and generating action.

We are, to some extent, repeating a discussion that SEO has already faced.

Positioning was never the full recipe.

Citation won't be either.

So, what would I prioritize in a GEO strategy today?

Given all this, I don't see much sense in starting a GEO strategy by implementing a long list of hacks.

I would start with the conditions that increase the chances of a brand being found, understood, used as a source, and recognized within the topics that truly matter to the business.

The first layer is findability. The site must allow access to relevant crawlers, have good architecture, indexable pages, coherent internal links, and clearly available content.

Next comes topical coverage. Instead of producing isolated pages for a collection of prompts, it makes more sense to identify the company's strategic themes and build depth around them.

The third layer is authority, and here content volume alone says little. Experts, research, cases, data, methodology, and practical experience help demonstrate that there is real knowledge behind what is being published.

I also consider it important to work on brand entity. It needs to be easy to understand who the company is, what it does, who it works for, what solutions it offers, and in which areas it has authority.

This construction doesn't end on your own domain. External presence also matters through press, partners, clients, experts, reviews, communities, events, and other legitimate environments where the brand can be cited and contextualized.

Finally, comes measurement.

Citations and mentions are new and interesting metrics, but they need to be tracked alongside presence by topic, traffic from AI engines, conversions, and business results.

This measurement itself is starting to gain specific tools.

In February 2026, Microsoft launched in public preview the AI Performance feature within Bing Webmaster Tools, allowing publishers to track when their content is used in responses produced by Copilot, Bing, and selected integrations, in addition to identifying which URLs appear as sources. 

It’s an interesting sign of where we are headed.

For years, much of organic performance analysis was concentrated on positions, impressions, clicks, and traffic.

Now we are starting to need to answer another question as well: 

how much of the knowledge produced by a brand is participating in the answers people receive?

GEO expands the battle for information

It is likely that in a few years, we will have much better answers to many of the questions still surrounding GEO.

We will have more extensive studies, more mature methodologies, better tools, and perhaps even a more consistent understanding of which characteristics increase the likelihood of a company appearing in generative responses.

We will also, inevitably, have new hacks promising to solve everything with one technical change and three content tweaks.

In the meantime, I believe a responsible strategy needs to coexist with two things at once: curiosity and skepticism.

It’s worth testing new formats. It’s worth monitoring prompts. It’s worth tracking citations. It’s worth analyzing how different platforms use content. It’s worth experimenting.

What isn't worth it is turning every discovered correlation into a new universal rule.

Interestingly, Google's current guidance for AI search ends up very close to what we have championed in SEO for a long time: solid technical structure, original content, useful information, real experience, reliability, and a focus on people. 

Perhaps this is one of the greatest ironies of this new race for GEO.

The more we try to figure out how to write for machines, the more we realize that the content with the greatest value is precisely that produced by people and companies that have something real to say.

This doesn't mean nothing has changed.

A lot has changed.

For many years, brands competed primarily for positions on a results page.

Now, they also compete for participation in the answer itself.

And participating in that response requires more than positioning a keyword, adding a FAQ, or adapting an article to the format we imagine an LLM prefers.

It requires a brand to be findable, understandable, relevant, reliable, and recognized within the topics it aims to speak on.

Above all, it requires producing information useful enough to deserve being retrieved.

For me, that is where a serious GEO strategy begins.

Not in the attempt to discover a magic formula, but in building a brand that has real reasons to be used as a source.

Want to understand how your brand is positioned in traditional searches and AI responses? Learn more about Bull Digital's SEO and GEO strategy.