What is GEO, Generative Engine Optimization, and why SEO rules alone are no longer enough
GEO is the work you do on your content so that AI systems pick it and cite it. The term comes from a 2023 Princeton study that measured what works: sources, numbers, and quotations raise visibility by up to 40%, repeated keywords don’t. Google confirms there are no extra requirements, but in an AI answer a page is either there or it isn’t.
In brief
- GEO, Generative Engine Optimization, is the work you do on your content so that AI systems pick it and cite it in their answers. The term comes from a November 2023 study by researchers at Princeton and IIT Delhi.
- On Google, a page holds a ranking; in an AI answer, it is either there or it isn’t. When an AI summary sits at the top of the results, searchers click through to a site 8% of the time, versus 15% without one.
- The study tested nine techniques on 10,000 queries: citing sources, adding statistics, and adding quotations raise visibility by up to 40%. Repeating keywords does nothing.
When someone asks ChatGPT which vendor to choose, they get a finished answer with three or four names in it. A brand that isn’t among those names doesn’t get rejected: it never gets read. GEO, Generative Engine Optimization, is the work it takes to get into that answer. This guide covers where the term comes from, how an AI answer is built, what works according to the only study that has measured it, what Google says, and where to start.
Index
GEO in one sentence
GEO is the set of choices, about your content and your site, that raise the odds a generative engine will use your pages to build its answer and cite them as a source. Generative engine means ChatGPT, Gemini, Perplexity, Copilot, and Google’s AI Overviews and AI Mode: systems that read several pages, summarize them with a language model, and hand back a single piece of text.
The term has a birth date. On November 16, 2023, six researchers from Princeton, the Indian Institute of Technology Delhi, and two independents published the study that coined the name Generative Engine Optimization, later accepted at the KDD 2024 conference. The authors start from one fact: publishers no longer control when and how their work shows up, because a model now sits between the page and the reader and decides on its own what to summarize.
How an AI answer is built
An AI answer goes through three stages: search, select, summarize. Google describes them in its documentation for site owners. AI Overviews and AI Mode use a technique called query fan-out: the person’s question is split into several searches, issued at once across subtopics and data sources; the models then pick the supporting pages and write the text. That is why, Google says, an AI answer shows a wider and more diverse set of links than a classic search.
This is where GEO parts ways with SEO. In classic search, a page holds a position: third, sixth, tenth, still visible to anyone who scrolls. In a generative answer, a page is either used or not, cited or not. The outcome is binary, and a page left out has no second page of results to be found on.
There is a second change. Classic search starts from typed keywords; AI starts from full questions, often with a comparison built in: which of these, what makes sense for a company like mine, who can be trusted with this problem. The model answers by drawing on pages that contain verifiable facts about that comparison, and it tends to credit whoever wrote those facts.
What changes for searchers: fewer clicks
How people behave has already been measured. When an AI summary appears in the results, searchers click through to a site 8% of the time, versus 15% when there is no summary: the figure comes from the Pew Research Center, based on searches by 900 US adults in March 2025. In the US, according to SparkToro and Similarweb, 68% of Google searches end without a single click.
The traffic doesn’t vanish: it takes a different road. Target saw its AI-driven traffic grow 2,000% in a year, versus 393% for US retail sites overall. A visitor who arrives from a conversation has already compared, ruled out, and chosen: the page that gets that visit is one of the few still in the running.
What actually works: the study’s results
The 2023 study is the only one so far to measure, under controlled conditions, which changes to a page increase its presence in AI answers. The authors built GEO-bench, a benchmark of 10,000 queries drawn from nine sources, and applied nine techniques to real pages, one at a time, measuring how much space and how prominent a position each page earned in the answer.
- The three best techniques are citing sources, adding quotations from credible sources, and adding statistics in place of qualitative statements: up to 40% more visibility than the original page.
- Writing in a simpler, more fluent way: 15% to 30% more.
- Giving the text a more authoritative, persuasive tone: no measurable improvement.
- Repeating the query’s keywords, the staple of old-school SEO: little to none.
Two details matter more than the average. The techniques help lower-ranked pages the most: for a page ranked fifth in the search engine, citing sources more than doubled its visibility in the answer, while for the page ranked first the result was negative. And effectiveness varies by field: statistics pay off most in content about law and government, quotations in content that explains and tells a story.
These numbers have a limit, and it should be stated. The tests ran on Perplexity and on an engine the authors built, between 2023 and 2024, with gains of up to 37% on the real-world engine. Today’s ChatGPT, Gemini, and AI Overviews run on different models: the proportions may have shifted, and the figures should be read as lab results.
What Google says
Google has published two official pages on the subject, one on its blog for site owners in May 2025 and one in its documentation, updated at the end of 2025. The position is clear: to appear in AI Overviews and AI Mode, there are no additional requirements and no special optimizations. A page has to be indexed, eligible to be shown with a snippet, and in line with Search’s technical requirements. No dedicated files, AI-only markup, or special structured data are needed.
The practical guidance comes down to seven points: unique, useful content; a page that reads well on every device; Googlebot free to crawl and a 200 response from the server; structured data that matches the visible text; high-quality images and video alongside the text; preview controls to decide what gets shown; and measuring the value of visits, not just clicks. Google adds that clicks from results pages with AI Overviews lead to longer visits.
The two readings, Google’s and the study’s, fit together. Google says what it takes to make the shortlist: a healthy, readable, reachable page. The study says what it takes to get picked from that shortlist: facts, numbers, sources, clear sentences. The first part is the technical SEO you already know; the second is the new work.
Which bots need access to your site
AI systems read websites with bots that have different names and different jobs, and the difference matters. OpenAI runs three: OAI-SearchBot powers ChatGPT search, GPTBot collects text to train models, and ChatGPT-User opens a URL when a user asks for it. The two settings are independent: you can allow OAI-SearchBot and disallow GPTBot in your robots.txt file, and the choice takes effect in about 24 hours. A site that blocks OAI-SearchBot drops out of the answers, even when its content is the best available.
Google follows the same logic: Search, AI features included, runs through Googlebot; to limit the use of your content for training other systems, there is the Google-Extended control. Sites behind Cloudflare have had three separate switches since July 2026, for search, agents, and training, and since September 15 an instruction that keeps a site in search while opting out of training, which Apple, Google, and Microsoft have committed to honor. The check is simple: open your robots.txt and your Cloudflare dashboard and make sure search bots and agents aren’t locked out by mistake.
Where to start: measure first
Before changing a page, you need to know what AI says about the brand today, in what words, and from which sources. Sources don’t carry equal weight: Reddit, Wikipedia, YouTube, LinkedIn, and Forbes are the domains AI cites most often, according to the 2026 AI Platform Citation Source Index by Everything-PR, and a brand described well on its own site but poorly on those sources gets an answer that sounds like those sources. GEO work therefore starts with three questions: does AI name the brand when someone asks for a recommendation in its category, what does it say when it does, and where does it get what it says.
An AI Brand Audit answers those three questions: 80 questions across 6 AI platforms, with the sources behind each answer and a comparison with competitors. The full path, from measurement to fixes to tracking over time, always starts with measurement and repeats the check after the changes, because AI answers shift with every model update. The closing fact is the one this guide started with: in an AI answer a page is either there or it isn’t, and the difference comes down to the facts it contains.
Frequently asked questions
No. Google says there are no additional requirements to appear in its AI features: the page has to be indexed and meet the requirements of classic Search. GEO comes on top: once a page is among those selected, what gets it picked and cited is facts, numbers, sources, and clear sentences.
Not for Google: its documentation says you don’t need machine-readable files, text files, or AI-specific markup. What you do need is for the right bots to get in: OAI-SearchBot for ChatGPT search, Googlebot for Google, and, if you use Cloudflare, search and agents left open.
A robots.txt change takes effect for ChatGPT in about 24 hours. Content changes get recrawled by Google anywhere from days to months later, depending on how often the page is crawled. That is why the measurement has to be repeated after the changes, with the same questions, instead of being run once.
