Generative Engine Optimization (GEO): The Research-Based Guide

Generative engine optimization (GEO) is the practice of shaping web content so that AI systems that write answers, such as ChatGPT, Perplexity, Microsoft Copilot, Gemini and Google AI Overviews, use it and cite it. The term comes from a 2023 research paper that found simple, evidence-rich rewrites raised a source's visibility in generated answers by up to 40%.

Key takeaways

• GEO was defined in a peer-reviewed paper by Aggarwal et al., published at KDD 2024.
• On a 10,000-query benchmark, adding citations, quotations and statistics lifted visibility by 30–40%.
• Keyword stuffing, a classic SEO trick, did not help and sometimes hurt.
• Lower-ranked pages gained the most, which makes GEO relevant for smaller sites.
• Google says optimizing for its AI features "is still SEO", so GEO builds on search basics rather than replacing them.

What is generative engine optimization?

Generative engine optimization is a set of methods for improving how visible your content is inside AI-generated responses. Visibility here means how much of the answer draws on your page and how prominently your source is cited.

The paper that introduced the term, GEO: Generative Engine Optimization, describes generative engines as systems that "combine conventional search engines with generative models". They retrieve relevant documents, then use a large language model to write a response grounded in those sources. At the time, the authors named Bing Chat, Google's SGE and Perplexity.ai as examples.

Today the same pattern runs in Google AI Overviews and AI Mode, ChatGPT search, Perplexity, Microsoft Copilot and Gemini. Wikipedia's article on generative engine optimization defines GEO as "the practice of structuring digital content and managing online presence to improve visibility in responses generated by generative artificial intelligence (AI) systems." For a shorter definition, see our glossary entry on GEO.

Why does GEO matter for brands?

GEO matters because more questions now end with a generated answer, and people click fewer links when one appears. If the answer leaves you out, many buyers never reach your site.

The usage numbers are large:

• Google said AI Overviews had "scaled to over 1.5 billion users" across 200 countries and territories (Google I/O 2025 recap, May 2025).
• A National Bureau of Economic Research working paper found that by July 2025 ChatGPT had been adopted by around 10% of the world's adult population (NBER, 2025).
• Pew Research Center found that 18% of Google searches in its March 2025 sample showed an AI summary (Pew, 2025).

The same Pew study shows why presence inside the answer matters. Users clicked a traditional result in 8% of visits when an AI summary appeared, versus 15% without one. Only 1% of visits included a click on a link inside the summary.

Where did GEO come from?

GEO came from academic research, not from marketing. The paper was first posted to arXiv on 16 November 2023 and published in the Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining in August 2024 (DOI 10.1145/3637528.3671900).

The authors are Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. Their abstract states the headline result directly:

"GEO can boost visibility by up to 40% in generative engine responses."

The paper also raised a fairness point. The authors wrote that content creators "have little to no control over when and how their content is displayed" in generative engines. GEO was proposed to give creators a way to improve their visibility.

How did the GEO study measure visibility?

The researchers built a benchmark, rewrote source pages with different methods and measured how much each rewritten source shaped the final answer. They used two metrics.

GEO-bench

GEO-bench is a set of 10,000 queries drawn from nine sources, including MS MARCO, Natural Questions, ELI5 and Perplexity.ai's Discover feed. The queries span 25 domains such as arts, health and games. Of the queries, 80% are informational and 10% each are transactional and navigational, matching real search patterns.

Two visibility metrics

1. Position-Adjusted Word Count measures how many words in the answer come from sentences citing your source, weighted so earlier citations count more.
2. Subjective Impression uses a language model to rate seven qualities, including relevance, influence, uniqueness and how likely a user is to click the citation.

These metrics are useful beyond the paper. They mirror what practitioners track today: whether you are cited, how early and how much of the answer you shape.

Which GEO methods worked best?

Adding citations, quotations and statistics worked best; keyword stuffing did not help. The table shows the Position-Adjusted Word Count scores from Table 1 of the paper (word sub-metric, higher is better).

In the authors' words, the top methods "achieved a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric." The table caption adds that the best methods beat the baseline by 41% and 28% on the two metrics.

Combining methods helped further. On a 200-example subset, pairing Fluency Optimization with Statistics Addition outperformed any single method by more than 5.5%.

Do GEO findings hold on real AI search engines?

Partly, yes. The authors repeated key tests on Perplexity.ai, a live engine, and saw the same direction of results. Quotation Addition gave "a 22% improvement over the baseline" on Position-Adjusted Word Count. Cite Sources and Statistics Addition improved by up to 9% and 37% on the two metrics. Keyword stuffing performed 10% worse than the baseline.

Two limits matter when you apply this today:

• Engines change. The tests ran in 2023–2024. Models, retrieval systems and citation layouts have changed since.
• Domains differ. The paper found that "the efficacy of these strategies varies across domains". What works for history questions may not work for software comparisons.

The practical lesson is to use the research as a hypothesis, then test it on your own prompts.

Why do smaller websites gain the most?

GEO helps smaller sites because generative engines judge the content they retrieve, not only the domain's ranking strength. In the paper's analysis, the Cite Sources method "led to a substantial 115.1% increase in visibility for websites ranked fifth in SERP". Over the same tests, the top-ranked website's visibility fell by 30.3% on average.

This matters for challenger brands. Classic rankings reward backlinks and history that take years to build. Inside a generated answer, a clear, well-sourced passage from a lower-ranked page can earn more space than a vague one from the leader.

How does GEO relate to SEO and Google's guidance?

GEO depends on SEO, because engines can only use what they can retrieve. Google's guide to optimizing for generative AI features says its AI features "are rooted in our core Search ranking and quality systems" and adds:

"From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

The same guide lists things you can ignore for Google Search, which is a useful reality check on GEO advice:

• LLMS.txt and other AI text files, which Google Search does not use.
• Breaking content into tiny chunks, which is not required.
• Special schema.org markup, since "Structured data isn't required for generative AI search".
• Chasing inauthentic mentions, which "isn't as helpful as it might seem".

Google Search Central discusses how AI is changing Search and SEO in this official video:

For how this compares with answer engine optimization, see our AEO guide.

Which generative engines does GEO apply to?

GEO applies to every engine that retrieves web pages and cites them in an answer. Each one has its own gate, and a page must pass it before any rewrite can help.

OpenAI says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features" (OpenAI docs). Perplexity says PerplexityBot is "designed to surface and link websites in search results on Perplexity" (Perplexity docs). Google says a page needs to be "indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link (Google Search Central).

How do you apply GEO research to your content?

Apply GEO by turning the paper's winning methods into editing habits, then checking the effect on the prompts that matter to you. Keep changes honest: add real sources and real numbers, never invented ones.

1. Add sources. Link official docs, research or government data next to each factual claim.
2. Add verbatim quotes. Quote experts or primary documents exactly, with a link.
3. Replace vague claims with numbers. "Most users" becomes a sourced percentage, or the claim goes.
4. Improve fluency. Short sentences, plain words, one idea per paragraph.
5. Define technical terms. Use the right terms, then explain them on first use.
6. Skip keyword stuffing. The paper found it did not help.

Our list of GEO strategies turns these into a fuller set of tactics, and the AEO content guide shows how to edit a page step by step.

How do you measure GEO results?

Measure GEO the way the paper did, adapted for live engines: run a fixed set of prompts, then record whether your brand is mentioned, whether your page is cited and how early it appears. Compare before and after each change.

When we designed the AnswerixAI visibility score, we included position alongside mentions and citations, because an early mention carries more weight for the reader. The score is 60% mention rate, 25% citation rate and 15% position in the answer, and the formula is shown in the dashboard. Answer Engine Insights tracks this across ChatGPT, Perplexity, Microsoft Copilot, Google AI Overviews, Google AI Mode, Gemini and Grok. Our team built the Site Audit to score a page on 47 AEO and GEO signals, and citation density, statistics and quotations are among them because of this research. For cross-team reporting, our guide to GEO for CMOs covers budget, ownership and KPIs.