What is GEO (Generative Engine Optimization)?

GEO starts from a different premise than SEO: there is no results page to win, there is an answer a model generates on the spot by combining several sources. Understanding how a brand reaches that answer means separating two independent mechanisms.

The first is the knowledge the model already carries from its training. An LLM trains on a massive text corpus, articles, forums, documentation, books, web pages, and adjusts millions of internal parameters to predict which word most likely follows another. A brand that appears frequently in sources the model treated as reliable during training, reference media, Wikipedia, technical documentation cited by others, ends up better represented in those parameters. That does not mean the model «remembers» a fact the way a person would. It means that, faced with a related question, it is statistically more likely to mention that brand without consulting any source at the moment it answers.

The second mechanism is the one used by Perplexity, Bing Chat, or Copilot, and to a lesser extent Google with AI Overviews: real-time information retrieval, or RAG (Retrieval-Augmented Generation). Before generating the text, the system runs a search, retrieves relevant fragments, and uses them as context. This explains why some AI answers include links and specific citations: the model is showing where it got each piece of data for that particular query.

In both cases there is a decisive step in between: source selection. The system does not use everything it finds, it picks a subset based on perceived authority, freshness, semantic relevance to the question, and, increasingly, whether an identifiable author or organization stands behind the content. Only after that selection comes generation: the model synthesizes that information into new text, it does not copy or link literally the way a traditional search engine would. For a brand, this means appearing in the answer depends on surviving two filters in a row, not just ranking well on Google.

The term GEO was not born as a marketing trend, it started as the title of an academic paper. On November 16, 2023, Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande, with ties to Princeton, Georgia Tech, and the Allen Institute for AI, published «GEO: Generative Engine Optimization» (arXiv:2311.09735), later presented at the KDD 2024 conference. The paper defined a benchmark of thousands of queries to measure brand visibility in generated answers, and showed, in a controlled experiment, that this visibility can be shifted deliberately.

The paper landed at a specific moment. Google had unveiled the Search Generative Experience (SGE) at Google I/O on May 10, 2023, an experimental layer available only through Search Labs. That changed on May 14, 2024, also at Google I/O, when Google dropped the beta label, rebranded it as AI Overviews, and started showing it by default in the United States.

Since then, each platform has taken its own path. Google integrates AI Overviews directly into search and, with AI Mode, offers a full conversational experience. Perplexity was built from the start as an answer engine with visible citations in every result. ChatGPT gained web browsing on top of its trained knowledge, and Microsoft carried the same approach into Copilot, built on Bing. GEO stopped being a technique for a single search engine: today it is a discipline that has to be worked separately for each engine, because each one decides differently which sources it retrieves and how it cites them.

The practices that work best in GEO do not replace SEO, they extend it in a different direction. Citable content is the foundation: AI systems cite specific fragments, not entire paragraphs, so a text broken into short sections, with verifiable data and a question-and-answer format, has better odds of being the fragment chosen.

Topical authority weighs more in GEO than in classic SEO, because a generative model does not evaluate an isolated URL, it evaluates whether a brand covers a topic with enough depth to count as a reliable source. A single well-optimized article delivers less than a body of content that covers the topic from several related angles.

Schema markup (FAQPage, HowTo, Article) helps a crawler or a retrieval system understand the content's structure without having to infer it from the visible HTML, which lowers the error margin when extracting a fragment. Digital PR, mentions in reference media and on Wikipedia, remains relevant because those sources tend to carry more weight in training and in source selection than a mention on a brand's own blog.

There is also an emerging standard, llms.txt, proposed by Jeremy Howard (Answer.AI) on September 3, 2024: a file at the domain root that summarizes the site's most important content. Treat it with caution: no major AI provider, not OpenAI, not Google, not Anthropic, has officially confirmed using it to crawl content, so today it is a supporting signal, not a guarantee of visibility.

Finally, monitoring closes the loop: without measuring which models mention a brand, at what position, and in what context, any content change is a blind bet. Tools like the AI Visibility Tracker turn that question into a repeatable data point. Underneath all of the above, E-E-A-T still carries weight, experience, expertise, authority, and trust, the same criteria Google uses to evaluate content.

• Structure each section as a complete, self-contained idea, retrievable as an independent fragment without depending on the paragraph before it.
• Add concrete, verifiable data, your own figures, dates, primary sources, instead of generic claims.
• Mark up content with Schema.org (FAQPage, HowTo, Article) so the structure is explicit beyond the visual design.
• Keep pages you want cited accessible to AI crawlers (GPTBot, PerplexityBot, Google-Extended).
• Pursue mentions in media and on Wikipedia through Digital PR: those sources outweigh a brand's own content.
• Publish content with a visible creation and update date, because a RAG system favors information it can verify as current.
• Measure brand visibility per model (ChatGPT, Gemini, Perplexity, Copilot) on a recurring basis, not as a one-off audit.

• Treating GEO as SEO with a different label, when it actually competes to be cited inside generated text, not to hold a position in a list.
• Blocking AI crawlers in robots.txt while expecting to appear in the answers of those same models.
• Relying solely on llms.txt as a visibility strategy, without building topical authority or Digital PR, when no major provider has confirmed using it.
• Publishing content without a visible date or identifiable author, which makes it harder for a model to treat it as a current, reliable source.
• Measuring success only by AI referral traffic, ignoring unlinked mentions, the most common form a brand's appearance takes in these answers.
• Expecting results within weeks: models refresh trained knowledge and retrieved sources on cycles the brand does not control, so the effect of a content change is not immediate.