What Is Generative Engine Optimization (GEO)? What Actually Gets a Page Cited by AI

Generative engine optimization (GEO) is the practice of shaping web content so that AI search tools, such as ChatGPT search, Perplexity and Google’s AI Overviews, are more likely to use it and cite it in their answers. The evidence on what works is still thin: one research paper defined the field, and the most reliable advice is what Google and OpenAI publish themselves, which comes down to being crawlable and indexed and writing clear, sourced, people-first content.

Below: what the original study found and didn’t test, what Google and OpenAI say, and where llms.txt stands. No one can honestly promise AI visibility, and this guide doesn’t.

What is generative engine optimization (GEO)?

The term comes from GEO: Generative Engine Optimization, a paper by Pranjal Aggarwal and colleagues from Princeton University, IIT Delhi and independent researchers, first posted in November 2023 and published at the KDD 2024 conference.

The paper describes a “generative engine” as a system with two parts: a search engine that retrieves sources for a query, and a language model that writes an answer from those sources with inline citations. GEO is anything a site owner does to the content of a page to increase how visible that page is inside the generated answer.

Visibility here isn’t a ranking position. The paper measures it by how much of the answer is attributed to your source and how early in the answer that text appears, plus a second score where a language model rates the citation for things like relevance, influence and how likely a user is to click it.

Is GEO different from SEO?

Less than the new acronym suggests. Google’s documentation on AI features and your website says you can apply the same foundational SEO best practices to AI Overviews and AI Mode as to Google Search overall, and that “specific optimization isn’t required.”

The real difference is the output. A classic results page shows a list of links. An AI answer summarizes several sources and links to some of them. In our view, GEO is best treated as an extra check on top of SEO, asking “could a system quote this page accurately?”, not as a separate discipline with its own tactics.

What the GEO research paper actually found

How the experiment worked. The authors built GEO-bench, a set of 10,000 queries drawn from nine sources, including real search queries (MS MARCO, ORCAS, Natural Questions), questions from Perplexity’s Discover page, the ELI5 subreddit and queries generated by GPT-4. About 80% were informational. For each query they took the text of the top five Google results, and a generative engine built on gpt-3.5-turbo wrote an answer citing those five sources.

They then picked one of the five sources at random and had a language model rewrite it using one of nine methods, such as adding statistics, adding quotations, citing sources, simplifying the language or stuffing in keywords. They compared the rewritten source’s visibility with the original.

What they reported:

• Cite Sources, Quotation Addition and Statistics Addition performed best, with relative improvements of 30-40% on the position-adjusted word count metric and 15-30% on the subjective impression metric. The headline “up to 40%” figure comes from here.
• Fluency Optimization and Easy-to-Understand (clearer, simpler text) improved visibility by 15-30%.
• A more authoritative, persuasive tone showed no significant improvement overall, although it helped on debate and history questions.
• Keyword stuffing gave little to no improvement. In the test on Perplexity it did 10% worse than the unmodified text.
• Results varied by topic. Statistics helped most for law and government, debate and opinion questions; quotations helped most for people and society, explanations and history.
• Lower-ranked sources gained the most. With Cite Sources, sources ranked fifth in Google gained 115.1% visibility, while top-ranked sources lost 30.3% on average.
• On Perplexity, tested on 200 queries, the best methods improved position-adjusted word count by 22% (Quotation Addition) and subjective impression by up to 37%.

The limits you should know before acting on it

• It measures quotation, not retrieval. Every test started with the page already among the five sources. The paper tells you nothing about how to get your page picked in the first place. For Perplexity, the authors uploaded the sources as files rather than letting it search.
• “Visibility” is a lab metric. It is computed from the answer text and from ratings by GPT-3.5. It isn’t clicks, traffic, leads or sales.
• The engines were 2023 systems. The main engine was built on gpt-3.5-turbo. Google AI Overviews and ChatGPT search were not tested, and the authors say their methods “may need to adapt over time” as these systems change.
• Search rankings weren’t measured. The authors say they didn’t evaluate how GEO methods affect search engine rankings.
• Accuracy wasn’t checked. The rewrites were done by a language model, and the paper doesn’t report verifying that the statistics, quotations and sources the model added were real.

Our reading: the paper supports a modest conclusion, that specific, sourced, clearly written text is easier for an AI system to use, and nothing stronger. It is not evidence that any tactic gets a site cited by Google or ChatGPT.

How do you appear in Google AI Overviews?

Google’s answer is short. To be eligible as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet. Google says, “There are no additional technical requirements.”

So the first step is ordinary indexing work. If important pages show up in Search Console as crawled but not indexed, they can’t be cited at all, whatever you do to the text.

Google then lists SEO fundamentals that “continue to be worthwhile”:

• Allowing crawling in robots.txt and by your CDN or hosting setup
• Making content easy to find through internal links
• Providing a good page experience
• Keeping important content available as text
• Supporting text with good images and video where relevant
• Making sure structured data matches the visible text on the page
• Keeping Merchant Center and Business Profile information up to date

Our technical SEO audit checklist covers how to check most of these.

Two more points set expectations. Both AI Overviews and AI Mode may use “query fan-out,” running several related searches across subtopics, which is why they can link to a wider range of pages than a classic results page. And AI Overviews only appear when Google’s systems decide they add something beyond classic Search, so they “often don’t trigger.”

To limit what Google shows from your pages, Google points to nosnippet, data-nosnippet, max-snippet and noindex. Google-Extended is a separate control: it covers AI training and grounding in some of Google’s other systems, not Search.

How do you get cited by ChatGPT?

OpenAI’s crawler documentation is the primary source here, and it covers access, not selection:

• OAI-SearchBot is used to surface websites in ChatGPT’s search features. OpenAI says sites opted out of OAI-SearchBot “will not be shown in ChatGPT search answers,” though they can still appear as navigational links. It recommends allowing OAI-SearchBot in robots.txt and allowing requests from its published IP ranges.
• GPTBot crawls content that may be used to train OpenAI’s models. Blocking it is a training decision, separate from search.
• ChatGPT-User visits pages when a user asks ChatGPT something. OpenAI says robots.txt rules may not apply to these user-initiated visits, and that this agent isn’t used to decide whether content appears in search.

A site that wants to appear in ChatGPT search but opt out of training could use:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Check your CDN or firewall too: blocking OpenAI’s IP addresses has the same effect as a robots.txt block. Once access is open, ChatGPT still decides which sources to cite, and the OpenAI pages cited here don’t describe how it does that.

Answer-first writing: what it is and why it helps

Answer-first writing means stating the direct answer to the question in the first one or two sentences, then explaining. It is how this article starts, and our writing guidelines require it for every post on this site.

There are two reasons to write this way. The first is the reader: Google’s self-assessment questions for helpful, people-first content ask whether someone leaves “feeling they’ve learned enough about a topic to help achieve their goal,” and whether headings give “a descriptive, helpful summary of the content.”

The second is mechanical. The engine in the GEO study was told that every sentence of its answer must be followed by a citation to a source that fully supports it. Under that kind of instruction, a clear, self-contained, sourced statement is easier to use than a claim spread across three paragraphs. Our rule of thumb: if a sentence can’t be quoted on its own and still be true, rewrite it.

In practice:

1. Put the answer under the heading that asks the question. Phrase H2s and H3s the way people search.
2. Name the source of every fact and link it. This is the “Cite Sources” method, done honestly.
3. Use numbers only when you have a primary source for them. Statistics helped in the study; invented statistics are factual errors.
4. Quote primary sources directly where the exact wording matters, as this article does with Google and OpenAI.
5. Write plainly. Simpler, more fluent text helped in the study, and it helps readers.
6. Keep key content in the HTML text, not only in images, PDFs or scripts.
7. Show who wrote it and how. Google asks whether it is “self-evident to your visitors who authored your content.”

A content audit is a practical way to find existing pages that bury the answer or make claims without sources.

What doesn’t work, and what we avoid

Keyword stuffing. It did nothing or worse in the GEO study, and it is a named violation in Google’s spam policies.

Tone instead of substance. A more authoritative, persuasive style showed no significant overall gain in the study.

Page farms and mass-produced Q&A pages. A tempting shortcut is to generate hundreds of short question-and-answer pages so a chatbot has something to quote. Google’s spam policies now define spam to include “attempting to manipulate generative AI responses in Google Search.” Its scaled content abuse policy covers pages “generated for the primary purpose of manipulating search rankings and not helping users,” and gives as examples using generative AI to create many pages without adding value, and stitching together content from other pages. That’s why we don’t build pages whose only job is to be quoted by an AI system.

Made-up statistics and quotes. The study’s strongest methods are also the easiest to fake. Google’s guidance asks “Does the content have any easily-verified factual errors?” A fabricated number is one.

What is llms.txt, and does Google use it?

llms.txt is a proposal by Jeremy Howard, first published in September 2024, for a Markdown file at /llms.txt that gives AI agents a short guide to a website. It also proposes clean Markdown copies of pages (for example page.html.md). The format is:

# Site name

> One-paragraph summary of the site

Optional details

## Section name

- [Page title](https://example.com/page/): What this page covers

The proposal says the file is meant to be read on demand, when an agent needs information while helping a user, rather than for training, and that it is used most heavily for software documentation.

It is a proposal, not an adopted standard. Google’s AI features documentation doesn’t mention it and says there are no technical requirements beyond being indexed and snippet-eligible. We know of no confirmation that Google Search uses llms.txt for rankings, AI Overviews or AI Mode. Chrome’s Lighthouse tool has an llms.txt audit for AI agents, but it only flags server errors and marks a missing file as “Not Applicable,” calling the file optional.

This site publishes an llms.txt file because our build generates it automatically. We don’t count on it to get cited anywhere.

How do you measure whether GEO is working?

Google says sites appearing in AI Overviews and AI Mode are counted in the overall search traffic in Search Console’s Performance report, under the “Web” search type. It also suggests tracking conversions and time on site in a tool like Google Analytics.

We’d measure GEO with the same business numbers as any search work: organic clicks, referral visits from AI assistants, and the leads they bring. The study’s visibility scores are useful in a lab, but you can’t see them for your own site. Our guide to SEO KPIs covers which numbers to report.

Where to start

Make sure your important pages are indexed and open to Googlebot and OAI-SearchBot. Then rewrite your key pages so each one answers its question in the first two sentences, backs every fact with a named source, and says who wrote it. That is ordinary good content work, and it is what our content marketing service does: fewer pages, each one specific enough to be worth citing, with no promises about which AI system will cite it.