Generative Engine Optimization (GEO): How to Create Content AI Search Can Cite

Search is no longer a list of blue links. Increasingly, it is a synthesized answer—assembled by ChatGPT, Perplexity, Claude, and Google's AI Overviews from the handful of sources each engine judges most quotable. Getting your content into that answer is a different discipline from ranking a page, and it has a name: generative engine optimization (GEO).

This guide covers what GEO is, how AI citations actually work, what the founding GEO research measured, how retrieval differs across AI search surfaces, and how to audit whether the engines cite you. It pairs closely with our guide to writing SEO articles without slop: the same fact-dense, section-by-section discipline that satisfies Google's quality systems is what makes content citable by a generative engine.

What Is Generative Engine Optimization (GEO)?

GEO is the practice of shaping content so a generative engine cites it in the answer it synthesizes, rather than ranking it in a list of links. The term comes from GEO: Generative Engine Optimization, a paper by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande of IIT Delhi and Princeton University, first posted in November 2023 and published at ACM SIGKDD 2024. The authors define GEO as "a new paradigm where content creators aim to increase their visibility (or impression) in generative engine responses."

The difference from SEO is where the payoff lands. SEO competes for a position in a ranked list, and the reward is a click. GEO competes for inclusion in an answer assembled from several sources at once, and the reward is a citation: your claim, your numbers, your name inside the text the user reads.

Being more comprehensive is not by itself a strategy. The function deciding which sources get cited is, in the paper's words, "a black-box function to end-users," and an engine can use information from your page without attributing it to you. What the research shows is that how content is presented, not only how much of it there is, changes how often it gets cited.

The Mechanics of AI Citations: RAG Vs. Training Data

Two different things get confused here. Static training data is what a model absorbed before its knowledge cutoff; it sits in the weights and is not queried per question. Retrieval-augmented generation (RAG) is what search assistants do at answer time: fetch live documents and write from them. Citations come from the second process, not the first.

A RAG answer is assembled roughly like this:

1. The user enters a query.
2. The engine reformulates it into one or more search queries and retrieves candidate sources from an index.
3. It pulls the chunks or passages that look most relevant out of the top sources.
4. It feeds those passages into the model's context window.
5. The model writes the answer and attaches citations to the source URLs.

You cannot change model weights. You can change what happens at steps two through five, and that is the entire practical scope of GEO: make the page retrievable, make each section extractable on its own, and make the extracted chunk worth quoting.

The Five Tactics That Lifted Citations

The GEO paper tested nine content edits against GEO-bench, a benchmark of 10,000 queries with the top five Google results attached to each one. Visibility was scored on two metrics: position-adjusted word count (how much of your content appears in the answer, weighted toward the top) and subjective impression (a model-scored judgment covering relevance, influence, uniqueness, and diversity). Five of the nine edits moved the numbers.

Tactic	What it means	What you do	Measured effect
Quotation Addition	Adds relevant quotations from credible sources	Quote named experts, standards bodies, or primary documents, and attribute the quote in the sentence that carries it	Best method overall: +41% on position-adjusted word count, +28% on subjective impression
Statistics Addition	Replaces qualitative discussion with quantitative figures wherever possible	Turn "many teams struggle with this" into the actual number, with the source of the number named	Top three: 30–40% on position-adjusted word count, 15–30% on subjective impression
Cite Sources	Adds citations from credible sources	Link the study, filing, or dataset behind each factual claim, at the claim rather than in a footer list	Top three: 30–40% on position-adjusted word count, 15–30% on subjective impression
Fluency Optimization	Improves the fluency of the text	Rewrite awkward, tangled, or padded passages so each sentence reads cleanly on its own	15–30% visibility boost
Easy-to-Understand	Simplifies the language while preserving meaning	Cut jargon and unpack the ideas, without stripping the specifics that make a passage worth quoting	15–30% visibility boost

The other four tactics did not earn their place:

• Keyword Stuffing (adding more query keywords, the classic SEO move) scored below the unoptimized baseline. The authors conclude that such methods "offer little to no improvement on generative engine's responses."
• Unique Words (adding unusual terms wherever possible) produced a negligible change.
• Technical Terms (adding domain terminology) helped, but less than any of the five above.
• Authoritative (rewriting in a more persuasive, authoritative tone) produced "no significant improvement," which the authors read as evidence that generative engines "are already somewhat robust to such changes."

The most consequential result appears when everyone optimizes at once. The researchers re-ran the experiment with all five retrieved sources optimized simultaneously, and the benefit went to the pages that were losing:

Tactic	Visibility change, rank-1 source	Visibility change, rank-5 source
Cite Sources	−30.3%	+115.1%
Quotation Addition	−22.9%	+99.7%
Statistics Addition	−20.6%	+97.9%

Backlinks and domain age do not condition a generative model; the retrieved text does. If you are ranking fifth for a query that matters to your business, that is where your leverage is.

The authors also replicated the experiment on Perplexity.ai, a deployed engine rather than their own test rig. Quotation Addition led again at +22% on position-adjusted word count, Statistics Addition reached +37% on subjective impression, and Keyword Stuffing performed 10% worse than baseline.

Where the Engines Differ

Retrieval is not one system. The surfaces you care about pull from different indexes and respond to different controls.

Engine	Where it retrieves from	Controls that apply	What that means for you
Google AI Overviews and AI Mode	The Google Search index, crawled by Googlebot	robots.txt rules for Googlebot; nosnippet, data-nosnippet, max-snippet, noindex	Google states there are "no additional requirements to appear in AI Overviews or AI Mode." Ordinary Search quality work is the price of entry. Google-Extended governs Gemini training, not AI Overviews
ChatGPT search	Third-party search results plus OpenAI's own crawl	OAI-SearchBot, separate from GPTBot	Blocking OAI-SearchBot removes you from ChatGPT's search results even where you rank well elsewhere
Perplexity	Its own index built by PerplexityBot, plus Perplexity-User fetches at answer time	PerplexityBot in robots.txt; Perplexity-User generally ignores it, since a user initiated the request	Index access is a separate decision from training access
Claude	On-demand fetches at answer time	Claude-User and Claude-SearchBot, each controlled separately from the training crawler ClaudeBot	You can decline training and still be retrievable
Grok	Live web search plus real-time posts on X	Standard crawler directives	Your presence on X is part of the same retrieval surface as your pages

No single configuration covers all of them, and none of these engines publishes its selection logic, which is why coverage across surfaces is something to measure rather than assume.

How to Audit Your Own AI Visibility

An AI visibility audit is a repeatable manual test: you ask the engines the questions your buyers ask and record who gets cited. It takes an afternoon to set up and produces a content backlog grounded in observed gaps rather than guesses.

1. Pick your money queries. Twenty to fifty prompts you would want to win. Write them the way a buyer would type them into a chat box: full questions, comparisons, "best X for Y," rather than keyword strings.
2. Pick the engines that matter to you. Whichever ones your buyers actually use. Adding a fifth engine you have no evidence about costs more than it returns.
3. Run every query on every engine in a clean session. Logged out, no chat history, no memory, no personalization. Otherwise you are measuring your own past behavior.
4. Record one row per query per engine. Were you cited? Which of your pages? Who else was cited, and which of their pages? Which sentence in the answer did the citation support?
5. Turn the gaps into a backlog. Where a competitor is cited and you are not, open the page that got cited and compare it against the five tactics above: named sources, quoted experts, real numbers, and sections that answer the query on their own without the rest of the page.
6. Re-run on a fixed cadence and keep the old rows. Monthly is enough for most teams. A single run tells you where you stand; the series tells you whether anything you shipped moved.

Where Deft Fits

Everything above is your work. The audit, the sourcing, the robots.txt decisions, the section structure. No writing tool does that for you, and Deft does not claim to.

What Deft does is generate the prose. It is a writing model tuned on real human web text so its output distribution matches human writing rather than model-average prose, which matters here for a specific reason: the tactics that lift citations are all things you have to actually write. Sourced statistics, attributed quotes, and clean, extractable sections are sentences, and generic AI prose does not produce them by itself. If what you already have is a generic AI draft, Rewrite mode regenerates it in Deft's distribution instead of paraphrasing it.

The Deft SEO Agent is a separate application that runs the research-and-outline work around that model, section by section. Both follow the same separation of concerns described in our guide to building a zero-slop SEO factory: research and structure get engineered deliberately, rather than left to a single improvised prompt.

Frequently Asked Questions About Generative Engine Optimization (GEO)

Is GEO Just SEO?

They optimize for different outcomes. SEO gets your content crawled, indexed, and ranked so a user clicks through to your URL. GEO makes your content easy for a language model to extract, verify, and cite inside a synthesized answer built from several sources. SEO sends traffic to your site; GEO sends citations from it. In practice the two overlap heavily on Google surfaces, where AI Overviews and AI Mode retrieve from the same index as Search, and diverge on engines that run their own crawlers.

What Is Llms.Txt and Why Does It Matter?

llms.txt is a proposed Markdown file at the root of a site, in the same position as robots.txt, containing a short, structured summary of what the site covers and where its key documents live. It was proposed by Jeremy Howard in September 2024 and remains an informal community specification rather than an adopted standard. No major AI provider has committed to reading it in production, and measurements of AI crawler traffic have found almost no requests for the file. Publishing one is cheap; treating it as a ranking lever is not supported by anything public.

Can I Block AI Scrapers and Still Be Cited by AI Assistants?

Yes, because training crawlers and retrieval crawlers are separate user agents. In robots.txt you can disallow GPTBot and ClaudeBot, which collect content for model training, while allowing OAI-SearchBot, Claude-SearchBot, and PerplexityBot, which exist to surface and link your pages. Two caveats: user-initiated agents such as ChatGPT-User, Claude-User, and Perplexity-User are triggered by a person rather than a crawl, and the operators note that robots.txt rules may not apply to them. And on Google, AI Overviews and AI Mode are served from the Search index via Googlebot, so blocking Googlebot removes you from Search entirely; Google-Extended only governs Gemini training.

How Quickly Can I See Results from a GEO Campaign?

There is no published answer. Citation depends on your page being recrawled and reindexed by whichever engine you care about, and none of them publish a refresh schedule or guarantee. Rather than expecting a change on any particular timeline, measure it: re-run the query audit above on a fixed cadence and compare against your earlier rows.

Claiming Your Territory in AI Search

The blue-link hierarchy that rewarded domain age and backlink volume is giving way to synthesized answers assembled from whichever sources an engine judges most extractable and most verifiable in the moment. The research is encouraging for anyone who is not already first: a page ranked fifth can gain more visibility from citations, sourced statistics, and quoted experts than the incumbent it is competing against, because a generative model is conditioned on retrieved text rather than on backlink graphs.

None of that requires guesswork. Audit where the engines cite your competitors and not you, and write the pages that close those specific gaps.