GEO
What the GEO Research Paper Found About AI Visibility
Key takeaways
- The 2023 GEO paper tested nine content tactics on a benchmark called GEO-bench.
- Quotations, statistics and cited sources performed best in the authors' tests; keyword stuffing performed below baseline.
- Only two engines were tested, and the authors list several limitations, so results are directional.
- Google says no special optimization is required for AI features, and Bing's guidance also stresses clarity, evidence and freshness.
- Measure with Search Console and Bing's AI Performance report rather than assuming a tactic worked.
What is Generative Engine Optimization?
Generative Engine Optimization (GEO) is both a term and a specific research paper. In the paper, authors Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande define "generative engines" as search systems that use language models to gather information from many sources and summarize it. Their core concern is that content creators have little control over whether or how their content shows up in those answers. The paper proposes GEO as a way to test what helps, and was first submitted to arXiv in November 2023 (latest version dated June 2024).
How did the researchers test it?
The authors built a benchmark called GEO-bench, a set of diverse user queries across domains paired with relevant web sources. They evaluated nine tactics against a no-optimization baseline: Authoritative, Statistics Addition, Keyword Stuffing, Cite Sources, Quotation Addition, Easy-to-Understand, Fluency Optimization, Unique Words and Technical Terms. Visibility was scored two ways: a position-adjusted word count, which weights how much of the answer cites your source and how early, and a subjective impression score judged by a language model.
They ran these tests on a custom generative engine built on GPT-3.5-turbo that pulled the top five Google results per query, and on Perplexity.ai using a 200-query subset.
Which tactics worked, and which did not?
According to the paper, Quotation Addition performed best overall, with roughly a 41% gain on the position-adjusted word count metric. Statistics Addition and Cite Sources also showed strong gains, at roughly 30% and 27% on that metric. Fluency Optimization and Easy-to-Understand improved results by smaller margins. The abstract summarizes the headline as visibility boosts of up to 40%.
Keyword stuffing, the classic SEO shortcut, did not help. The authors report it scored below baseline, about 8% worse on the main metric, and roughly 10% worse on Perplexity.ai. The paper also says effectiveness varies by domain, which suggests no single tactic works everywhere.
How far can you trust these numbers?
Not as far as the headline suggests. The authors themselves list limits: only two engines were tested, so generalizing is uncertain; some scores came from language-model judges that were poorly calibrated and needed normalization; the benchmark tags were produced by GPT-4 and may be noisy; and the combination analysis used only a 200-query subset. They also did not measure any effect on traditional search rankings. Generative engines change quickly, so a tactic that helped in 2023 testing may behave differently today. Read the findings as a hypothesis worth testing, not a recipe.
What do the search engines themselves say?
Google's Search Central documentation on AI features says standard SEO practices still apply to AI Overviews and AI Mode and that no additional technical requirements or special markup are needed. A page must be indexed and eligible to appear with a snippet to be a supporting link, and Google notes that indexing and serving are never guaranteed. It recommends crawlability, internal links, text-based content, and structured data that matches visible text.
Microsoft's Bing Webmaster Blog, in its February 2026 announcement of the AI Performance report, offers guidance that overlaps with the paper's findings: improve structure with headings, tables and FAQ sections, support claims with evidence such as data and sources, keep content current, and use IndexNow to notify participating engines of changes. Notably, that advice is not identical to the paper's tactic list, and Microsoft does not claim any tactic guarantees citation.
What should you do with this?
A reasonable reading is to treat the research and the engine guidance as pointing the same direction: make pages clear, specific and verifiable. Where it is honest to do so, include accurate statistics, attribute them to a named source, and quote primary sources briefly. Do not stuff keywords, and do not invent figures to look authoritative. Then measure. Google says AI feature traffic is counted in Search Console's Performance report under the Web search type, and Bing's AI Performance report shows total citations, cited pages and grounding queries (a sample). Both are citation counts, not rankings, so use them to spot which pages get referenced and build from there.
Frequently asked questions
What is the GEO paper?
It is a 2023 arXiv paper by Aggarwal and colleagues that defines generative engine optimization and tests nine content tactics on a benchmark called GEO-bench.
Does keyword stuffing help in AI answers?
In the paper's tests it did not. The authors report it scored below the no-optimization baseline on their main metric.
Does Google require special optimization for AI Overviews?
Google's documentation says no additional technical requirements or special markup are needed beyond standard Search eligibility and helpful, people-first content.
How can I see if AI engines cite my site?
Google counts AI feature traffic in the Search Console Performance report under Web, and Bing Webmaster Tools offers an AI Performance report showing citations and a sample of grounding queries.
Sources
- GEO: Generative Engine Optimization (arXiv:2311.09735) — arXiv
- AI features and your website — Google Search Central
- Introducing AI Performance in Bing Webmaster Tools Public Preview — Microsoft Bing Webmaster Blog
Keep reading
See how AIrecommend.ai helps businesses get found and recommended by AI assistants.