# Generative engine optimization (GEO): what it is and what works

URL: https://serpel.app/blog/generative-engine-optimization

Updated: 2026-10-10

Generative engine optimization (GEO) is the practice of making your content easy for AI answer engines to retrieve, understand and cite. The original GEO research found that adding statistics, quotations and source citations raised visibility in its test setup, while keyword stuffing lowered it. Google says optimising for its AI features is still SEO, so the sensible approach is to build on good SEO and measure the result.

## Key takeaways

- Generative engine optimization (GEO) means making pages that AI answer engines such as ChatGPT search, Google’s AI Overviews and Perplexity retrieve and cite. The term comes from a 2023 research paper accepted to KDD 2024.
- In that paper, adding quotations, statistics and source citations raised a page’s visibility in generated answers by up to 40% in a lab setup, and keyword stuffing did worse than changing nothing.
- Those results come from a simplified engine with the sources already in context. A 2026 survey of 45 studies finds no technique with a proven, lasting effect on organic discoverability, so read the numbers as an upper bound.
- Google says you don’t need special files, markup or rewriting for its AI features. Crawlable, indexed, snippet-eligible and unique content is the foundation.
- Measure GEO by repeating real prompts, reading Google’s Generative AI performance report and Bing’s AI Performance report, and checking your server logs for AI crawlers.

## What is generative engine optimization?

Generative engine optimization (GEO) is the work of increasing how often AI answer engines use your content when they write an answer. A generative engine retrieves pages, passes passages to a language model and returns a written answer, usually with citations. ChatGPT search, Google’s AI Overviews and AI Mode, Perplexity and Microsoft Copilot all work this way. GEO aims to make your pages the ones that get retrieved, quoted and linked.

Researchers at IIT Delhi, Princeton University and other institutions coined the term in the paper [GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735). Its first author is Pranjal Aggarwal. It was first posted in November 2023 and accepted to KDD 2024. You will also see AEO (answer engine optimisation), LLMO and “AI SEO” used for the same idea, and Google’s own guide treats GEO and AEO as common names.

One note on the phrase “geo seo”. It sometimes means geographic targeting for local search. This article is about generative engines, not locations.

## How do AI answers choose which sources to cite?

No provider publishes the full recipe, but the documentation of OpenAI and Google describes the same broad pipeline.

1. **Query rewriting.** ChatGPT search [rewrites your question into one or more targeted queries](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt) and may send more specific ones after reading the first results. Google describes the same idea as [query fan-out](https://developers.google.com/search/docs/appearance/ai-features) for AI Overviews and AI Mode.
2. **Retrieval.** The system fetches candidate pages. Google says a page must be indexed and eligible to appear in Search with a snippet to show up as a supporting link. OpenAI says sites that block OAI-SearchBot [will not be shown in ChatGPT search answers](https://developers.openai.com/api/docs/bots).
3. **Ranking.** OpenAI says ChatGPT ranks search results using multiple factors and that placement is not guaranteed. Its help page does not list the factors.
4. **Answer writing and citation.** The model writes the answer from the retrieved passages and attaches links. OpenAI warns that citations can be incomplete, outdated or incorrect.

That gives you three gates to pass: be crawlable, be retrievable for the sub-questions behind a prompt, and be worth quoting once retrieved. Results also vary. A July 2026 [survey of GEO research](https://arxiv.org/abs/2607.14035) reports that commercial audits found low source overlap and substantial run-to-run variability, so a single answer tells you very little.

## GEO vs SEO: what is different and what is the same?

**GEO compared with classic SEO**
| Aspect | Classic SEO | GEO |
| --- | --- | --- |
| Goal | Rank a URL for a keyword | Be retrieved, quoted or cited in a generated answer |
| Unit of competition | A page in a list of results | A passage or fact inside one written answer |
| Crawlers that matter | Googlebot and Bingbot | Those two plus OAI-SearchBot, Claude-SearchBot, PerplexityBot and others |
| Content levers | Relevance, links, page experience and unique content | The same, plus quotable facts, statistics and cited sources |
| Consistency | A position is stable enough to track daily | Answers vary by run, location and user, so you track rates over many checks |
| What you measure | Rankings, impressions and clicks | Citation rate, mention rate, cited sources, AI referrals and crawler hits |

The overlap is large. Google’s [AI optimisation guide](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) says optimising for generative AI search is still SEO, and that you don’t need new machine-readable files, AI text files, markup or special rewriting for Google’s features. The practical difference is that you now care about several crawlers and about being quotable, not only about being rankable.

## What did the original GEO research measure?

The authors built **GEO-bench**, a set of 10,000 queries from nine sources across 25 domains, with the text of the top five Google results attached to each query. Their test engine fetched those five sources and had gpt-3.5-turbo write a cited answer. For each query they rewrote one source with a GEO method and measured how much of the answer was credited to it, weighted by where the citation appears. That score is called position-adjusted word count.

They tested nine methods. The table shows each method’s score from the paper’s Table 1. Higher is better, and the unmodified source scored 19.3.

**Position-adjusted word count per GEO method, from Table 1 of the paper**
| Method | What was changed | Score |
| --- | --- | --- |
| Quotation addition | Added relevant quotations from credible sources | 27.2 |
| Statistics addition | Used quantitative statistics instead of qualitative discussion where possible | 25.2 |
| Fluency optimisation | Improved the fluency of the text | 24.7 |
| Cite sources | Added citations from credible sources | 24.6 |
| Technical terms | Added technical terms where possible | 22.7 |
| Easy-to-understand | Simplified the language | 22.0 |
| Authoritative | Made the style more persuasive and authoritative | 21.3 |
| Unique words | Added unique terms where possible | 20.5 |
| Keyword stuffing | Added more keywords from the query, as in classic SEO | 17.7 |

The best methods improved on the baseline by 41% on this metric, and keyword stuffing, the classic SEO tactic, scored below doing nothing. According to the authors, a more authoritative tone made no significant difference, which suggests the engines are fairly robust to persuasive phrasing.

Two further results matter for smaller sites. In an experiment where all sources were optimised at once, citing sources raised the visibility of the fifth-ranked result by 115.1% on average, while the top-ranked result lost 30.3%. And the effect depended on the topic: adding statistics worked best for law and government, debate and opinion queries, while citing sources worked best for statement and fact queries.

The authors also ran the methods against Perplexity.ai on a 200-query subset. Because Perplexity does not let you choose source URLs, they supplied the source text as uploaded files. Quotation addition was best on position-adjusted word count there, with a reported 22% improvement. Statistics addition reached 37% on the subjective impression metric, and keyword stuffing again fell below the baseline on position-adjusted word count.

## How far can you trust those results?

They are useful, but narrower than the headline “up to 40%” suggests. The test engine was simplified, the sources were already in the model’s context, and the score measures a share of an answer, not clicks or traffic. The July 2026 survey of 45 GEO studies (a single-author preprint) concludes that the paper’s gains hold when a source is already in context, but do not show better organic discoverability or lasting traffic effects. It also finds that generic optimisation heuristics transfer poorly and that citation-oriented rewrites can hurt retrieval.

> **Be careful with GEO promises:** Google advises being wary of third-party tools that promise ranking success or claim to use internal Google metrics, and OpenAI says placement in ChatGPT search is not guaranteed. Take the paper as evidence about what makes a passage quotable once it is retrieved, not as a guarantee of being retrieved.

## GEO checklist: what to do in practice

1. **Let the right crawlers in** Check robots.txt, your CDN and your firewall. Sites that block OAI-SearchBot are not shown in ChatGPT search answers, and Google’s AI features need Googlebot access. Our guide to [AI crawlers](https://serpel.app/blog/ai-crawlers) lists which ones to allow.

2. **Put the answer in the HTML** Serve key content as server-rendered text. In a [December 2024 analysis](https://vercel.com/blog/the-rise-of-the-ai-crawler) of Vercel’s network, GPTBot, ClaudeBot and PerplexityBot did not execute JavaScript. Google also asks that important content is available as text.

3. **Write what only you can write** Google says unique, useful, non-commodity content will likely influence your presence in AI search more than any other suggestion in its guide. Original data, tested procedures and expert detail also give a model something specific to quote.

4. **State facts with numbers and sources** The strongest methods in the GEO paper added statistics, quotations and citations to credible sources. Do it only where it is true: attribute every number, link the primary source and never invent a quote.

5. **Answer the sub-questions** Because ChatGPT and Google run several related searches per prompt, cover the follow-up questions a reader would ask, on one useful page or a few distinct ones. Google warns that producing a page for every query variation can breach its scaled content abuse policy.

6. **Earn genuine mentions** Google says its AI features can show what is being said about products and services across the web, and that seeking inauthentic mentions is less helpful than it seems. Documentation links, honest reviews and independent comparisons are the mentions worth earning.

7. **Skip what does not help** Keyword stuffing scored below the baseline in the paper. For Google’s AI features you don’t need an llms.txt file, special schema markup or content chunked for AI. Our [llms.txt examples](https://serpel.app/blog/llms-txt-examples) explain where the file does make sense.

## How do you measure GEO?

Use four sources, because none of them is complete on its own.

- **Prompt checks.** Pick 20 to 50 questions your customers really ask, run them on the AI products that matter and record whether your domain is cited (linked) or mentioned (named). Repeat the checks, because answers change between runs.
- **Google’s Generative AI performance report.** Search Console [reports impressions](https://support.google.com/webmasters/answer/16984139?hl=en) in AI Overviews and AI Mode by page, country and device. The report is built around impressions, and Google says it rolled out to all sites worldwide on 31 Aug 2026.
- **Bing’s AI Performance report.** Bing Webmaster Tools [shows how often your pages are cited](https://www.searchenginejournal.com/bing-webmaster-tools-adds-ai-citation-performance-data/566874/) in Copilot and AI summaries in Bing. It launched as a public preview in February 2026.
- **Server logs.** Count requests from OAI-SearchBot, Claude-SearchBot, PerplexityBot and user-triggered fetchers such as ChatGPT-User. They show which pages AI systems actually read.

Serpel automates the first one for ChatGPT with web search and Google AI Overviews, and also reports which AI crawlers your robots.txt allows. Add your questions, run the checks and read how often answers cite or mention your domain and which sources they cite most. [AI visibility in Serpel](https://serpel.app/features/ai-visibility) covers the details, and our comparison of the [best AI visibility tools](https://serpel.app/blog/best-ai-visibility-tools) shows the alternatives.

```bash
serpel ai add --project <id> --prompt "What is the best SEO tool for developers?" --prompt "How do I check if ChatGPT cites my website?"
serpel ai run --project <id> --wait
serpel ai status --project <id>
```

## Is generative engine optimization worth the effort?

At the level of fundamentals, yes. Crawl access, server-rendered text, unique content and attributed facts cost little and help classic SEO too. Paying for a service that promises guaranteed citations is a different matter: both OpenAI and Google say outcomes cannot be guaranteed.

Treat GEO as an experiment loop. Pick a set of prompts, record the baseline, change one thing, wait for the crawlers to return and measure again. That habit matters more than any single tactic, because the engines and their sources change often.

## Frequently asked questions

### What is generative engine optimization?

Generative engine optimization (GEO) is the practice of making content easy for AI answer engines to retrieve, understand and cite. The term was introduced in a 2023 research paper by Pranjal Aggarwal and co-authors, accepted to KDD 2024. AEO and “AI SEO” are common names for the same work.

### Is GEO different from SEO?

Mostly it builds on SEO. Google’s guidance says optimising for its generative AI features is still SEO, and the same foundations apply: crawlable, indexed, unique content. What changes is that you also care about AI crawlers such as OAI-SearchBot, about being quotable, and about measuring citations as well as rankings.

### Do generative engine optimization tactics really work?

In the original paper, adding quotations, statistics and source citations improved visibility by up to 40% in a test setup, and keyword stuffing did not help. A 2026 survey of the research finds those gains hold once a page is already retrieved, but no technique has a proven lasting effect on being discovered. Use them as good writing practice and check the results with your own measurements.

### What are the best generative engine optimization tools?

Tools fall into three groups: SEO suites with an AI add-on, dedicated AI visibility platforms and developer-first tools such as Serpel. They check whether AI answers cite or mention your brand for a list of prompts. Our guide to the best AI visibility tools compares what each one tracks and what it costs.

## Sources

- [Aggarwal et al., GEO: Generative Engine Optimization (arXiv 2311.09735, KDD 2024)](https://arxiv.org/abs/2311.09735), accessed 2026-10-10
- [Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (arXiv 2607.14035)](https://arxiv.org/abs/2607.14035), accessed 2026-10-10
- [Google Search Central: Optimizing your website for generative AI features on Google Search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide), accessed 2026-10-10
- [Google Search Central: AI features and your website](https://developers.google.com/search/docs/appearance/ai-features), accessed 2026-10-10
- [OpenAI Help Center: Searching the web with ChatGPT](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt), accessed 2026-10-10
- [OpenAI: Overview of OpenAI crawlers](https://developers.openai.com/api/docs/bots), accessed 2026-10-10
- [Vercel: The rise of the AI crawler](https://vercel.com/blog/the-rise-of-the-ai-crawler), accessed 2026-10-10
- [Search Console Help: Generative AI performance report](https://support.google.com/webmasters/answer/16984139?hl=en), accessed 2026-10-10
- [Search Engine Journal: Bing Webmaster Tools adds AI citation performance data](https://www.searchenginejournal.com/bing-webmaster-tools-adds-ai-citation-performance-data/566874/), accessed 2026-10-10