Generative engine optimization (GEO) is the practice of structuring content so AI answer engines are more likely to use and cite it. The main study, presented at KDD 2024, found that adding quotations, statistics and cited sources raised visibility in a simulated engine, while keyword stuffing lowered it. The tests used a research setup, and Google says its AI features need no special optimization.

Generative engine optimization has plenty of confident advice and very little checked evidence. Most guides stack statistics without naming where they came from, then promise citations that nobody can guarantee. This post sticks to what a published study measured and what official documentation says.

I'm Janak Pokharel, an SEO and paid ads consultant. I have not run controlled GEO experiments, so I am not presenting test results of my own. I am reporting a peer-reviewed paper and documentation from Google, OpenAI and Anthropic, and I label my own analysis as analysis. You will see what the paper tested, which methods scored highest, where the study stops being useful, what Google tells site owners, and how a Next.js developer can apply the findings without overclaiming. The aim is to help you make better content decisions, not to chase a formula that the evidence does not support. Nothing here is a guarantee. Everything below was checked on October 16, 2026.

What did the GEO paper test?

The GEO paper, by Pranjal Aggarwal and five co-authors, tested nine ways of rewriting web content to increase its visibility in AI-generated answers. It used GEO-bench, a set of 10,000 queries, and a simulated generative engine. It was presented at the 30th ACM SIGKDD Conference in August 2024.

The paper, titled GEO: Generative Engine Optimization, lists Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande as the other authors. The version I read is v3, dated June 28, 2024.

The setup works in two steps. Each query was paired with the cleaned text of the top five Google results. A model, gpt-3.5-turbo at temperature 0.7, then wrote an answer from those sources. The authors measured two things: a position-adjusted word count, which scores how much of the answer is attributed to your source and how early it appears, and a subjective impression score.

Which methods scored highest and lowest?

In the paper's main table, Quotation Addition scored highest at 27.2 against a 19.3 baseline, followed by Statistics Addition at 25.2, Fluency Optimization at 24.7 and Cite Sources at 24.6. Keyword Stuffing scored 17.7, below the baseline. The scores measure visibility inside a simulated AI answer, not Google rankings.

I calculated the last column from the paper's printed scores. The scores themselves come from the paper's Table 1.

The pattern is useful for writers. The best methods add substance a reader can check: a quote, a number, a named source. The only method that made things worse is the one SEO has used for years.

What are the limits of the study?

The study used a simulated engine built on gpt-3.5-turbo, tested single-turn questions only, and found that effectiveness varies by domain. Its authors also note that methods may need to adapt as generative engines evolve. A Perplexity.ai test on a 200-query subset reported Quotation Addition improving position-adjusted word count by 22%, which is promising but still one test.

Here is how I read those limits. The paper does not measure how often ChatGPT, Gemini, Claude or Google's AI Overviews cite a page, because it did not test those products directly. The engine in the main experiment is a research setup, not a commercial one. The data also predates today's products. Treat the results as a reason to improve content quality, not as a ranking formula.

What does Google tell site owners to do?

Google's AI features documentation, updated December 10, 2025, says a page must be indexed and eligible to show in Search with a snippet to appear as a supporting link in AI Overviews or AI Mode. Google lists no extra technical requirements and says no special schema.org markup, AI text files or other optimization is needed.

Google's AI features guidance lists ordinary best practices instead:

• Allow crawling in robots.txt and in any CDN or host settings.

• Make content easy to find through internal links.

• Provide a good page experience.

• Keep important content in text form.

• Support text with relevant images and videos.

• Make sure structured data matches the visible page text.

• Keep Business Profile and Merchant Center information current.

• Verify your site in Search Console.

Google also says traffic from AI Overviews and AI Mode is counted in the Web search type in Search Console's performance report. There is no separate AI report.

How do you build citable pages in Next.js?

Build citable pages by putting the direct answer in server-rendered text near the top of each section, supporting it with a named source or a verified quotation, and keeping structured data consistent with the visible text. These steps follow Google's guidance and the paper's findings, but neither guarantees citations from any AI engine.

In Next.js, a component without 'use client' is a Server Component, so its text is in the initial HTML. A small wrapper keeps every section consistent:

// components/AnswerSection.tsx (a Server Component: no 'use client')
type Props = {
  id: string
  question: string
  answer: string
  children: React.ReactNode
}

export function AnswerSection({ id, question, answer, children }: Props) {
  return (
    <section id={id}>
      <h2>{question}</h2>
      <p><strong>{answer}</strong></p>
      {children}
    </section>
  )
}

Use it with a short checklist, which is my own analysis and not a proven method:

1. Write the answer in 30 to 60 words that still make sense if copied alone.

2. Name the subject instead of writing "it" or "this".

3. Add one statistic or one verified quotation only where it supports the claim, and name the source in the sentence.

4. Show a visible review date for topics that change.

5. If you add JSON-LD, make it match the text on the page. The Next.js JSON-LD guide shows how to render it safely.

Which crawlers does robots.txt control?

OpenAI and Anthropic each document separate crawlers for search, model training and user-initiated visits. Blocking a search crawler can keep a site out of that product's answers, while blocking a training crawler signals that content should not be used for training. Robots.txt changes can take about 24 hours to apply.

This table follows OpenAI's crawler documentation and Anthropic's crawler article. Bots and policies change, so re-check both pages before you edit your file. Whether to allow training crawlers is a business choice.

Conclusion

The research supports a modest conclusion: content with checkable quotations, statistics and named sources was more visible in a simulated AI answer, and keyword stuffing was not. Google says pages need to be indexed and snippet-eligible, with no special markup. So do the basics well, state answers clearly, and avoid promising citations. Pick one page this week, rewrite its main answer in plain, source-backed language, and confirm that the page is indexed in Search Console. Revisit the paper's findings as engines change, since its authors expect methods to need updating. If you want help applying this to a site, you can request an SEO proposal or read about advanced SEO services.

Frequently asked questions

What is generative engine optimization?

Generative engine optimization is the practice of structuring and improving content so AI answer engines can understand, retrieve and potentially cite it. The term comes from a 2023 research paper. It overlaps heavily with SEO, because Google requires pages to be indexed before they can appear as supporting links.

Does GEO replace SEO?

No. Google says a page must be indexed and eligible to show with a snippet to appear in AI Overviews and AI Mode, so crawlability and indexing still come first. GEO is better treated as an extra layer of clear, well-sourced writing on top of sound SEO.

Does keyword stuffing help in AI answers?

In the paper's simulated tests, no. Keyword Stuffing scored 17.7 against a 19.3 baseline, about 8.3% lower. That result comes from a research setup, so treat it as evidence against the tactic, not a measurement of any commercial engine.

Should I block GPTBot or ClaudeBot?

That is a business decision. GPTBot and ClaudeBot relate to model training, while OAI-SearchBot and Claude-SearchBot relate to search features. Blocking training bots signals that content should not be used for training. Blocking search bots can reduce your visibility in those products.

Can anyone guarantee citations in ChatGPT, Gemini or Claude?

No. The paper reports relative visibility gains in tests, and Google documents no special optimization. No study I reviewed shows a method that guarantees a citation. Be cautious of any provider that promises one.

Janak Pokharel is an SEO and paid ads consultant based in Kathmandu and a co-founder of Limi Creatives. He publishes guides at Janak Pokharel's website.