What Is Generative Engine Optimization? Origin, Mechanism, and What Research Really Shows

By Pedro Sonda, dooop

Generative Engine Optimization (GEO) is the optimization of content, websites, and an entity’s presence so that AI systems generating answers, known as generative engines, can find, understand, use, and cite this information when responding to a query. The concept was formalized in 2023 by researchers from Princeton University and IIT Delhi, among others, and presented at the KDD 2024 conference.

The central point of this article is that GEO is not limited to "being cited." An AI-generated answer results from a sequence of steps, and a company can be lost at any of them. Understanding this mechanism is what separates a GEO strategy from a list of writing tricks.

Updated in October 2026. For the executive and summarized definition, see What Is GEO?.

Origin of the Term

The term appears in the article GEO: Generative Engine Optimization, by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, and Ameet Deshpande (arXiv). The authors start from the observation that generative engines combine search and language models to produce synthesized answers, which removes from content creators the control over how, and if, their information appears.

The study proposes three things:

1. A concept: GEO as a paradigm to help content creators improve visibility in generative engine answers.
2. A benchmark: the GEO-bench, with 10,000 questions from 25 different domains.
3. Visibility metrics adapted to answers, not link lists: Position-Adjusted Word Count (how much of the answer’s text is associated with a source, with greater weight for what appears earlier) and Subjective Impression (evaluation of relevance, influence, uniqueness, prominence, and click probability, among other dimensions).

Interpretation: the third contribution is the most important for companies: in a generated answer, visibility is not position, it is participation in the answer’s text.

What the Research Showed, and What It Did Not Show

The authors tested nine content alteration methods and measured the effect on source visibility.

Result	Detail	Correct Interpretation
The most effective methods were adding textual citations, statistics, and references to sources	In the position-adjusted word count metric, the baseline without optimization was 19.5; with textual citations, 27.8; with statistics, 25.9; with source references, 24.9	Content with verifiable evidence tends to be more used in synthesis
Keyword stuffing was below baseline	17.8 versus 19.5	The classic tactic of repeating terms does not help and can harm
Maximum reported gain reached about 40%	Effectiveness varied by domain	There is no universal technique
Test on a real platform (Perplexity)	Improvements up to 37%	Effects were not limited to the simulated environment

Source of all table data: Aggarwal et al., KDD 2024.

What the study did not show: it measures relative visibility of sources within answers in a controlled environment, with mechanisms available in 2023. It does not measure traffic, sales, brand recommendation, nor the current behavior of ChatGPT, Gemini, or AI Overviews, which have changed significantly since then. It is strong evidence that evidence and clarity increase content utilization, but not a manual guaranteeing results on any platform.

How a Generative Engine Builds an Answer

Details vary by platform, but the process usually has three steps.

Interpretation and expansion of the question. Google informs that AI Overviews and AI Mode may use query fan-out, launching multiple related searches on subtopics to build the answer (Google Search Central).
Source retrieval. The system searches its own index, the web, or search providers. OpenAI states it may obtain URLs from third-party search providers or by crawling other pages (OpenAI — FAQ for publishers).
Synthesis. The model selects excerpts, combines information from various sources, and writes the answer, sometimes with links.

Two practical consequences follow:

The excerpt matters more than the page. Microsoft describes that AI assistants break content into smaller parts and evaluate each by authority and relevance; in their words, ranking still exists but is less about ordering entire pages and more about which content pieces earn a place in the final answer (Microsoft Advertising).
Ranking for the exact phrase is not enough. Ahrefs analyzed 15,000 queries and found that, on average, only 12% of URLs cited by AI assistants were in Google’s top 10 for the same question (Ahrefs, 2025). The likely explanation is fan-out: the system searches variations of the question and combines results.

dooop Framework: The 8 Steps of AI Visibility

For dooop, AI visibility is a pipeline. Each step has a typical failure and a specific intervention.

Step	Question	Typical Failure	Where to Intervene
1. Activation	Does the question trigger a search for sources?	AI answers only with the model’s prior knowledge	Historical brand presence in broad sources
2. Discovery and retrieval	Is the content found and retrieved?	Page blocked, not indexed, or unrelated to subquestions	Technical access, indexing, coverage of subtopics
3. Exposure in context	Does the content enter the material the model actually reads?	Essential information hidden, diluted, or dependent on other parts	Self-contained and direct excerpts
4. Citation	Is the source referenced?	Content is used, but another source gets credit	Original and attributable statements
5. Highlight	Does the information appear in a relevant position?	Mention at the end, like a sidebar note	Answer the main question, not just details
6. Absorption	Does the content influence facts, structure, or language of the answer?	Source is cited, but the answer follows another logic	Clear thesis, frameworks, own definitions
7. Fidelity	Is what is attributed to the source correct?	Service, audience, or data distorted	Entity consistency across all sources
8. Behavior	Does the user click, consider, or act?	Visibility without business impact	CTA coherent with intent and origin measurement

Reading the table shows why "being cited" is an incomplete goal: content can be retrieved and not appear in the answer, be cited and misrepresented, or be excellent and never retrieved. A well-done GEO diagnosis identifies at which step the company is losing before proposing any action.

How to Make a Company Appear in AI Answers According to This Mechanism

The pipeline indicates the order of interventions:

Unlock access. Google requires the page to be indexed and eligible to appear with a snippet; there is no additional requirement or special schema for AI Overviews (Google Search Central). On ChatGPT, blocking the OAI-SearchBot prevents the site from appearing in search answers (OpenAI).
Cover subquestions. If the AI expands the question, the content must also answer neighboring questions.
Write self-sustaining excerpts. Definition, answer, evidence, and implication in the same block.
Add attributable evidence. Data, sources, and expert citations, which were precisely the most effective methods in the Princeton study.
Manage what third parties say. In 75,000 brands, web mentions had a correlation of 0.664 with presence in AI Overviews, versus 0.218 for backlinks (Ahrefs, 2025). Correlation is not causation but suggests that retrieval depends on more than the site itself.

The complete roadmap is at How to Make a Company Appear in AI Answers.

Preparing the Company: What Changes in Management

GEO shifts the management question from "how much traffic do we get?" to "in which decisions do we participate?" This requires three changes:

Plan by intent, not by keyword. An intent, such as "hire AI governance consulting," appears in dozens of formulations. Content must answer the intent, not a phrase.
Measure by sampling. Since answers vary with each execution, measurement must repeat the same questions multiple times and across platforms. See How to Measure AI Visibility.
Treat the entity as an asset. The company’s description, services, and proofs must be consistent on and off the site because the fidelity step depends on this.

What to Do Now

If your company does not appear in any answer, the problem is likely in steps 1 and 2: access, indexing, and presence in sources.
If it appears but is poorly described, the problem is in step 7: the entity is inconsistent.
If your content is used but another company is recommended, the problem is in steps 4 to 6: lack of own thesis and attributable evidence.
If it appears but nothing happens in the business, the problem is in step 8: lack of linking visibility to measurable action.

Frequently Asked Questions

Is Generative Engine Optimization an official term from Google or OpenAI? No. The term has academic origins. Google talks about "AI features" in search and states that SEO practices remain valid for them.

Does the GEO study apply to the current ChatGPT? The principles (evidence, clarity, sources) align with public platform guidance, but the study’s numbers were measured in a 2023 environment and should not be projected as expected results today.

Can you optimize for all AIs at once? The base is common (access, clarity, evidence, consistent entity), but each platform retrieves sources differently. Therefore, measurement should be done per platform.

About dooop

dooop is a consultancy based in Caxias do Sul (RS) that prepares leaders and organizations to transform artificial intelligence into business capability through strategy, governance, and training. This article is part of dooop’s series on GEO and AI Decision Infrastructure.