What did the paper test?
Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, "GEO: Generative Engine Optimization" (arXiv 2311.09735, submitted 16 November 2023, latest revision 28 June 2024), built GEO-bench, a benchmark of 10,000 queries from nine sources. For each query the setup fetched the top 5 Google results, rewrote one source with a GEO method, had gpt-3.5-turbo write an answer with citations, and measured how visible the rewritten source was. Visibility was scored two ways: Position-Adjusted Word Count (how much of the answer cites you, weighted by position) and Subjective Impression (seven sub-metrics including relevance, influence, uniqueness and diversity).
What happened?
Position-Adjusted Word Count, Table 1 of the paper (baseline: no optimization = 19.5 on the word score):
| Method | What the page was changed to do | Word score |
|---|---|---|
| No optimization | Baseline | 19.5 |
| Keyword stuffing | Add more query keywords, as in classic SEO | 17.8 |
| Unique words | Add distinctive words | 20.7 |
| Cite sources | Add citations from credible sources | 24.9 |
| Statistics addition | Replace qualitative claims with statistics where possible | 25.9 |
| Quotation addition | Add quotations from credible sources | 27.8 |
The authors summarize that their best methods (Cite Sources, Quotation Addition, Statistics Addition) gave a relative improvement of 30 to 40% on Position-Adjusted Word Count and 15 to 30% on Subjective Impression, and that the best methods improved on baseline by 41% and 28%. They say keyword stuffing, the classic SEO method, did not perform well. On Perplexity.ai, using 200 sampled queries with source text uploaded as files, they report visibility improvements of up to 37%.
What are the limits?
- The main experiment is a simulated engine, not ChatGPT, Google or Perplexity's live retrieval; the Perplexity run used uploaded files, not live web retrieval.
- The authors write that methods "may need to adapt over time as GEs evolve" and that query types can change.
- They did not evaluate how the changes affect traditional search rankings.
- The effect of each method varied across domains, which the authors say calls for domain-specific optimization.
- It measures visibility inside an answer, not clicks, leads or sales.
- Bing's guidelines list keyword stuffing, unreviewed auto-generated content and attempts to manipulate language models among practices that can reduce visibility, so use these tactics honestly: real statistics, real quotations, real sources.
What to do
- Where a page makes a claim, add the number and its source in the same sentence, with a link.
- Quote the primary source, briefly and with attribution, instead of paraphrasing from memory.
- Do not invent statistics or quotations to chase visibility; they are checkable and can hurt trust.
- Test the change on your own prompts: run your fixed prompt list before and after (measurement loop, citation share).
- Do not drop the basics: Google says no special optimizations are needed beyond indexing and snippet eligibility (GEO vs AEO vs SEO).
Related
Sources
Read November 19, 2026:
- Aggarwal et al., GEO: Generative Engine Optimization (full text, PDF), Tables 1 and 7, section 9 (Limitations)
- Microsoft Bing, Bing Webmaster Guidelines
- Google Search Central, AI features and your website
Disclosure: Dominate The World is published by Neural Command LLC (NRLC.ai), which sells AI search visibility, schema and entity services.
Next step
Pick your most important page and add one sourced statistic or quotation to each section that makes a claim.