GEO: Generative Engine Optimization
Indian Institute of Technology Delhi and Princeton University · KDD 2024, the ACM SIGKDD conference, a top peer-reviewed venue in data mining · Tested on a 10,000-query benchmark across 25 domains and on a live generative engine (Perplexity.ai).
What the paper is. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande (IIT Delhi and Princeton), published at KDD 2024. They built GEO-bench, a benchmark of 10,000 real queries across 25 domains, and tested nine ways of rewriting a web page to see which ones made generative engines (including Perplexity.ai) surface and cite that page more. Three changes work best: adding credible citations, adding quotations and adding statistics. They raised visibility by 30 to 40% on the position-adjusted word count metric and 15 to 30% on subjective impression.
1. What we took from it
- Three changes work best: adding credible citations, adding quotations and adding statistics. They raised visibility by 30 to 40% on the position-adjusted word count metric and 15 to 30% on subjective impression.
- Making the text clearer and easier to read also helped, by 15 to 30%. Engines reward presentation, not only content.
- Keyword stuffing, the classic SEO trick, does not work for AI engines.
- Which method works best varies by domain, so pages need targeted changes, not one template.
2. Where we used it
The GEO page: the research card, the “How we apply it” line, and step 7 of the ten-step method (“GEO methods”). It also sits behind the Product discovery tile on the Journey page.
3. How we used it
Steps 1 to 6 of our method find what a product page is missing (audit, select pages, understand the page, competitors, search trends, find the gap). Step 7 fixes the gap using only the paper’s proven methods: relevant statistics, credible sources, useful quotations, clearer wording and a more authoritative tone, and never extra keywords. Step 8 applies those changes to the page and adds conversational FAQs, step 9 checks crawl access, sitemap, speed and AI access, and step 10 re-audits so the before-and-after is measured the way the paper measured it. Because the paper found results differ by domain, we audit each page instead of applying one fixed formula.
Engines reward presentation, not only content.
