For twenty years SEO had one clear goal: reach the top of the ten blue links. Today, for a growing share of searches, nobody sees those ten blue links anymore. The user asks ChatGPT or Perplexity a question, or reads the synthesized answer from Google AI Overviews, and stops there. Traffic doesn't disappear: it moves. And the optimization game moves with it. This is where GEO comes in, Generative Engine Optimization: the set of techniques for getting found and, above all, cited by AI-based answer engines. In this article we explain what GEO really is, how it differs from the AI search optimization you already know, and which concrete tactics we apply today on aicircus.it to show up on Perplexity, get cited by ChatGPT and land in Google AI Overviews answers.
What GEO is and why it isn't the SEO you know
GEO is the optimization of content so that it gets selected, synthesized and cited by generative engines: ChatGPT with its search function, Perplexity, Google AI Overviews, Copilot. The difference from classic SEO isn't cosmetic. In SEO the unit of success is ranking: you're tenth, third, first. In GEO the unit of success is the citation: the model used your content to build the answer and named you as a source, with a clickable link or even just your brand name.
The physics of the game change. A traditional search engine returns a list and leaves the user to choose. An answer engine does something different: it reads many pages, extracts fragments, recomposes them into a single answer and decides which sources deserve to be cited. You no longer win "a position": you win, or lose, the chance to be one of the three or four references the model places at the bottom of its answer. That's why GEO rewards those who write content a machine can understand, extract and recompose without misunderstandings.
One misconception to clear up right away: GEO doesn't replace SEO, it extends it. Pages still need to be indexable, fast, well linked. If Google can't see your page, Google AI Overviews won't use it either. GEO adds a layer on top of the SEO foundations, it doesn't tear them down.
How answer engines choose and cite sources
To optimize, it helps to understand the mechanism. Answer engines work, simplifying, in two ways. Some — like Perplexity or ChatGPT with active search — run a real-time query, retrieve a set of pages, read them and synthesize their content while citing the sources. Others draw partly on the model's internal knowledge, built during training on enormous amounts of text. In the first case, being retrievable and clear right now matters; in the second, what matters is being present, consistent and repeated across the web over time.
When the model decides which sources to cite, certain signals recur. Direct relevance carries weight: a paragraph that answers the question exactly is worth more than a page that circles the topic. Structural clarity carries weight: explicit headings, paragraphs that stand on their own, answers placed up front. Perceived authority carries weight: who you are, whether the information about you is consistent across sources, whether others cite you. And, plainly, extractability carries weight: a figure inside a clean sentence gets picked up readily, the same figure buried in a six-line clause often does not.
The practical consequence is sharp. Writing "for answer engines" means writing in self-contained blocks, putting the answer before the explanation, and making every piece of information retrievable out of context. It's the exact opposite of text that spins in place just to fill the page.
Tactic 1: citable, self-contained content with direct answers
The first tactic is also the most underrated: write paragraphs that work on their own. A model recomposing an answer doesn't read the article like a human from start to finish: it extracts the piece it needs. If that piece requires three preceding paragraphs to make sense, it's useless.
So: every section opens by stating what it's about. Definitions are given explicitly ("GEO is..."). Answers to predictable questions sit up front, not at the bottom. Lists and numbered steps help, because they're formats models extract easily. On our site we structure articles like this: first the answer in one or two sentences, then the reasoning, then the examples. It isn't a style meant only for machines, it's also more honest toward the human reader, who finds what they're looking for right away.
A detail that matters: titles and headings should say what the section contains, not try to be clever. "How answer engines choose sources" is a good heading because it's also a plausible query. "Behind the scenes of the magic" is not.
Tactic 2: original data, numbers and statistics
LLMs love to cite numbers. A concrete figure, a percentage, a measurement, a date: these are high-density units of information, easy to drop into an answer and to attribute. If your page is the origin of a number the model wants to use, you become the natural source to cite.
This is especially true for original data. A statistic you produced yourself — a benchmark, the result of an internal test, a measurement taken on your own projects — exists nowhere else, so if the model uses it, it has to cite you. That's very different from repeating a figure already published by twenty other sites, where the citation, if it comes, goes to the primary source. When we can, we include field measurements in our technical content (times, iteration counts, test results) precisely because they're citable material and they're ours.
Two cautions. Numbers need context (what they measure, on which sample, when) or they aren't credible and don't get picked up. And they need to stay current: a dated figure loses value fast.
Tactic 3: structured data, schema.org and JSON-LD
Structured data is how you tell a machine, unambiguously, what's on the page: this is an article, this is the author, this is the date, this is a FAQ, this is an organization. It's typically implemented with JSON-LD following the schema.org vocabulary, a block of code that describes the content in a machine-readable form without touching what the user sees.
Why it matters for GEO: it reduces ambiguity. An answer engine that finds an Article markup with author, date and organization spelled out has an easier time understanding who says what and when. A FAQ markup turns your questions and answers into pairs ready to be extracted. An Organization markup with the right fields helps the model build a consistent entity around your brand.
On aicircus.it we use schema.org JSON-LD to mark up articles (Article type with author and date), the organization (who we are, with the official links) and, where it makes sense, FAQs. It isn't a trick: it's simply telling machines, formally, what we tell humans in words.
Tactic 4: the llms.txt and ai.txt files
These are two files, still uncommon, that live at the root of the site and speak directly to AI-based tools. They're worth explaining because we use both.
The llms.txt file is a map designed for LLMs. In practice it's a markdown text file that summarizes what the site is and gathers the most important links, in a clean form free of noise (no menus, no banners, no JavaScript). The idea is to offer a model a "distilled" and reliable version of the key content, easier to read than the markup-heavy HTML of a normal page. On aicircus.it we publish an llms.txt that describes who we are, what we do, and points to the articles and pages we want to be easy to find and understand.
The ai.txt file, by contrast, is closer in spirit to robots.txt: it declares your rules toward AI crawlers and agents — what they can use, what you'd rather not have collected. It's a way to make your preferences explicit instead of leaving them implied.
An honest caveat: neither file is a universally respected standard, and on their own they guarantee nothing. But they're cheap to maintain, they signal care, and as the ecosystem settles, having your house already in order is an advantage. For us they're part of the baseline kit of a site built well today.
Tactic 5: authority, entities and consistency across sources
Answer engines reason by entities, not just by keywords. "ShadApps", "aicircus", a person, a product like CORA or Smartlet CRM: to the model these are entities for which it builds a profile by aggregating what it finds around the web. The more consistent that profile, the more you're a reliable source to cite.
From this come two concrete jobs. The first is consistency: the way you describe yourself must be the same everywhere — site, profiles, registries, third-party mentions. Name, description, location, scope: if the versions disagree with each other, the model struggles to build a solid entity. The second is the explicit link between your online presences, which in markup is done with the "sameAs" field of the Organization or Person schema: a list of URLs that say "this profile, this page, this registry are the same entity." It's a way to give the model the coordinates instead of leaving it to guess.
Real authority, though, isn't declared: it's earned over time, with content others find useful enough to link and cite. Schema helps make it legible, not manufacture it.
Tactic 6: freshness, clear headings and multilingual versions
Three more operational factors, but ones that move the needle. Freshness: answer engines, especially those that search in real time, favor up-to-date content. A publication and update date that's visible and explicit in the markup helps. It doesn't mean retouching everything every week, it means keeping the content that matters alive and declaring when you last revised it.
Clear headings, as we said, are at once structure for the machine and a map for the user. An article whose subheadings already form a sensible outline is an article a model navigates effortlessly.
Multilingual, finally: if you serve an audience in more than one language, hreflang tells engines which version to show to whom, prevents the versions from competing, and ensures the Italian answer draws from the Italian page and the English answer from the English page. On a site like ours, which publishes in Italian and English, it's how we avoid scattering the signal between the two versions of the same content — exactly what you're reading right now.
How to measure GEO
A fair question: how do we tell whether it's working, if ranking is no longer the only thing to watch? It can be measured, but with different and less convenient metrics.
The first is direct citations: do you show up in the answers of ChatGPT, Perplexity, Google AI Overviews when you ask the queries you care about? It's partly a manual check — you ask the questions and look — and partly automatable with tools that monitor brand presence in answer engines. The second is referral traffic coming from AI engines: visits arriving from domains like those of ChatGPT or Perplexity start appearing in your logs and analytics. They grow slowly, but they're the signal that the citation turns into clicks. The third is brand queries: if people discover your name inside an AI answer, they then tend to search for you by name; a rise in brand searches is a measurable side effect of good GEO.
None of these metrics is as precise as the old ranking. GEO is measured more by trend than by exact position, and it should be looked at over time windows, not on a single day.
Myths to debunk
Given the hype around the topic, a few sharp clarifications.
No, rewriting meta tags isn't enough. GEO isn't a handful of optimized titles and descriptions. Answer engines read the body of the page, its structure, its data, the entity behind it: meta tags are the least of their concerns. Anyone selling "GEO optimization" as a metadata touch-up is selling smoke.
No, SEO isn't dead. It's the opposite: GEO rests on SEO. A page that isn't indexable or is slow is invisible to AI engines too. Whoever announces the death of SEO every six months, for fifteen years now, is wrong this time too.
No, you can't buy a citation. There's no auction to land among the sources of an AI answer the way there is for paid ads. Citations are earned with useful content, legible structure and real authority. Anyone promising to "guarantee" you a spot in ChatGPT's answers is either lying or describing something they don't control.
And one last: no, it isn't a shortcut. GEO doesn't make the work of writing real content unnecessary. If anything, it makes it more important.
In practice: GEO is SEO done well, plus machine-readable structure
Strip away the hype and a simple summary remains. GEO is SEO done well plus an obsessive care for structure, so that a machine can read, extract and cite your content without misreading it. No magic, no shortcuts: citable, self-contained content, original and contextualized data, structured JSON-LD markup, files like llms.txt and ai.txt in their place, a consistent entity across sources, freshness and clear headings, hreflang where needed.
It's exactly what we do on our own sites, starting with aicircus.it: articles are structured to be citable, we mark up content with structured data, we maintain llms.txt and ai.txt, we keep our organization's identity consistent. Not because it chases a trend, but because it's how a site reads well today — for humans and for machines. If you have to start from one place, start here: write a page that genuinely answers a question, so that anyone, person or model, can take the right piece of it and say where it came from.