Generative Engine Optimization (GEO): the complete guide

Generative Engine Optimization (GEO) is the set of techniques that raise the odds of a generative engine (ChatGPT, Perplexity, Claude, Gemini, or Google's AI Overviews) citing your website in its answer. The term was coined by a team from Princeton and Georgia Tech in November 2023, tested on 10,000 queries, and published at KDD 2024. This guide covers what GEO is according to that paper, what Google says in writing, which methods work with numbers behind them, and what we found when we ran the checks on the ten pages that rank for this search in Spain today.

What is GEO and who defined it?

GEO is optimizing content so that generative engines select it and cite it in their answers. The name comes from the paper "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, and four co-authors, posted on arXiv on November 16, 2023, and presented at the KDD 2024 conference in Barcelona.

The paper starts from one observation: a generative engine doesn't return a list of links. It writes an answer from several sources, and content creators lose control over when and how their work shows up. GEO is the response to that problem, and it brings two things the "SEO for AI" conversation lacked: a visibility metric and a 10,000-query benchmark, GEO-bench, on which nine different methods were measured.

If you want the difference from classic SEO, we lay it out with a table in GEO vs SEO: what sets them apart. If you want what to do with your site today, SEO for AI gives you seven concrete actions.

How is GEO different from SEO and AEO?

All three chase visibility in search, but at three different layers. SEO targets positions in the results list. AEO (Answer Engine Optimization) targets the direct answer a search engine shows at the top, like a featured snippet. GEO targets being cited by a language model inside an answer it wrote itself.

In one line: SEO is measured in positions, AEO in snippets won, and GEO in citation frequency.

The three layers share a foundation: without crawling, indexing, and authority, there is nothing to cite. Google says so in its official guide: its generative AI features rely on the same core ranking and quality systems as classic search. GEO doesn't replace SEO. It adds a layer on top.

How does a generative engine choose which sources to cite?

A generative engine answers in four steps: it takes the question, breaks it into several related searches, retrieves pages from its index or from a live crawl, and selects the passages that answer best to write with them. The citations are the pages those passages came from.

Google names the first two mechanisms in its guide for site owners. Retrieval-augmented generation, which Google also calls grounding, pulls relevant, up-to-date pages from its index and reviews their information to build the answer. Query fan-out is the set of simultaneous, related queries the model generates to gather more information. A page doesn't compete for one query. It competes for the five or ten queries the model turned it into.

Those four steps yield the three conditions a page has to meet. First, the engine's crawler has to be able to get in. Second, the page has to answer one of the fanned-out queries in an extractable way, in a short, self-contained passage. Third, that passage has to contain something the model prefers to cite over other sources, and that is exactly what the paper measured.

Which GEO methods work, according to the paper?

The paper tested nine methods on GEO-bench with two engines: a custom one built on GPT-3.5 and Perplexity in production. The main metric is called Position-Adjusted Word Count: how many words of the answer come from your source, weighted by how early they appear. The three methods that improved visibility most have one thing in common: they add evidence.

The top three, Quotation Addition (27.8), Statistics Addition (25.9), and Cite Sources (24.9), are the only ones that clearly beat the baseline; the other six, keyword stuffing included, fall below.

Source for the figures: Aggarwal et al., "GEO: Generative Engine Optimization," results table on GEO-bench, KDD 2024. Links to all sources are at the end of the article.

The number that sums up the paper: the best methods beat the baseline by 41% in word-based visibility and by 28% in the subjective impression of the answer. And the method that gets there is the cheapest to apply: quoting a named source verbatim. Repeating the keyword, the reflex of SEO for years, landed among the methods that add the least.

What does Google recommend?

In 2026 Google published an official guide titled "Optimizing your website for Google Search's generative AI features." Its thesis fits in one sentence from the guide itself: SEO still matters, because AI Overviews and AI Mode are built on the core ranking and quality systems of Search.

What Google asks for is familiar: content with a unique perspective based on experience, not generic; clear paragraphs and headings; quality images and video; a crawlable technical structure with semantic HTML; and less duplicate content. What Google advises against is less familiar, and it contradicts a good part of what gets sold as GEO:

1. You don't need to generate machine-readable files, text files for AI, special tags, or Markdown. Google names the practice without naming the file, but it means llms.txt and its relatives.
2. You don't need to chunk your content into small pieces for AI.
3. You don't need to rewrite your style for AI systems.
4. Chasing mentions that aren't genuine is less useful than it looks.
5. Structured data is not required for generative AI search.

Our position: we publish an llms.txt on auroraglobalgroup.com because it takes five minutes, does no harm, and stays available to any agent that looks for it. But it is not a GEO lever, and anyone selling it as one hasn't read the guide. Google also says how to measure: the Search Console performance report already includes traffic from generative AI features.

A worked example: the ten pages ranking for "generative engine optimization" in Spain

On September 3, 2026, we checked two things about this very search. Market data comes from OpenSEO, built on DataForSEO data for Google Spain.

First, demand. "Generative engine optimization" gets 320 searches a month in Spain, with a difficulty of 44 out of 100. The Spanish variants carry more weight and are easier: "geo seo" 590 searches, "geo posicionamiento" 480, both at difficulty 0. And across the seven queries on the topic we checked that day, Google Spain showed an AI Overview on all seven. Anyone searching for GEO in Spain already gets a generated answer before a link.

Second, crawler access. We requested the ten URLs in Google Spain's top 10 for this search with the User-Agent of three AI crawlers: GPTBot (OpenAI), ClaudeBot (Anthropic), and PerplexityBot. Nine of the ten returned 200 to all three. One returned 403 to any automated request, AI crawlers included. That page holds a top-10 spot on Google and, at the same time, is invisible to three of the most used generative engines. Its robots.txt doesn't say so: the block sits at the server.

Our own site passes that same test for GPTBot, ClaudeBot, CCBot, PerplexityBot, Applebot, and Meta's crawler, and a monthly monitor checks it again. It's the first condition of GEO and the least reviewed.

A third data point, from the same day, on who AI cites. We asked Perplexity about a pet food brand we work with. It cited the brand, but the source it used to do so was a press article about the brand, not its website. The model cites what it considers reliable and extractable, and sometimes that is what others say about you. That's where mentions and authority come in, and they remain half of GEO.

What to do with all this: an eight-point GEO checklist

1. Confirm that GPTBot, ClaudeBot, PerplexityBot, and OAI-SearchBot get a real 200 from your server, with their User-Agent, not just in robots.txt.
2. Decide what you allow: crawling to answer, yes; use for training, your call. They are separate agents in robots.txt.
3. In every section, answer in the first two or three sentences and explain afterward. The model extracts passages, not articles.
4. Add verbatim quotes with a name and a date, statistics with a source, and linked references. Those are the three methods that gained the most visibility in the paper.
5. Turn every H2 into a question someone would ask an assistant, and make the answer stand on its own.
6. Include one data point only you have: an experiment, your own measurement, a case with numbers.
7. Keep technical SEO in shape: crawling, indexing, canonicals, unique titles. Google's AI relies on them.
8. Measure monthly with the same ten questions across four engines, and in Search Console with the AI performance report. Without measurement, GEO is an opinion.

How is GEO measured?

The metric we use is citation frequency: the number of answers where your site is cited, divided by the number of questions times the number of engines. Ten fixed questions from your sector, four engines, once a month, same questions every month. Going from 0 to 3 out of 40 is a real change; a shift in average Google position says nothing about whether AI cites you.

Three readings complement it: traffic arriving from chatgpt.com and perplexity.ai in Google Analytics 4, AI crawler hits in your server logs, and the Search Console performance report for Google's AI features. We walk through it step by step in our GEO service, and it is the first deliverable of the audit.

Frequently asked questions