Generative Engine Optimization (GEO)

What is generative engine optimization?

Generative engine optimization is the practice of making content retrievable, quotable and citable for AI systems that write answers, such as Google AI Overviews, ChatGPT search and Perplexity. Its unit of success is a citation or mention inside the answer, not a ranked link.

A generative engine is a search system that answers a query in written prose instead of returning only a list of links. It combines a large language model (LLM), which writes the text, with a retrieval system, which fetches web pages for the model to read. Google AI Overviews and AI Mode, ChatGPT search, Perplexity, Gemini and Microsoft Copilot all work this way. When one of these engines answers a question, it reads a handful of sources and credits some of them with a link or a name.

Wikipedia describes GEO as structuring digital content and managing online presence to improve visibility in responses generated by generative AI systems. Wikidata records it as item Q134083964, an instance of digital marketing technique. Both definitions share one idea: the work targets the answer the AI writes, and every page that feeds that answer competes for a place in it.

GEO is a subclass of search engine optimization. It inherits SEO’s foundations — crawlable pages, indexable content, relevance to the query — and adds a concern SEO never had to handle: whether a model will lift a specific passage from a page and attribute it.

The practice goes by several names. Wikidata and Wikipedia list these aliases:

• AI search optimization — the most common commercial label
• LLM optimization (LLMO) — naming the model rather than the engine
• AI optimization (AIO) — a shorter form used in agency marketing
• AI SEO — a label that also refers to using AI tools for SEO work
• AI visibility optimization — naming the outcome being measured
• AI citation optimization and AI retrieval optimization — naming single stages of the pipeline

Answer engine optimization (AEO) is treated as a sibling practice, not an alias. AEO grew out of featured snippets and voice assistants, where the goal was a single extracted answer; GEO covers multi-source answers that a model writes itself.

What does GEO optimize for, if not rankings?

GEO optimizes for three outcomes inside an AI answer. A citation is a link to the page, shown inline, in a footnote or in a source panel. A mention is the brand or product named in the answer text, with or without a link. A recommendation is a mention framed as advice — the brand listed as a good option for the user’s need.

These outcomes do not map onto rank positions. An AI answer has no position 1 to 10; it has a sequence of claims, and each claim draws on one or more sources. A page ranked seventh in classic search can supply the passage an AI Overview quotes, and a page ranked first can go uncredited. The distinction matters for measurement too: a brand can be mentioned in hundreds of answers while its site earns few citations, because the engine learned the brand name from other sources.

Where did the term “generative engine optimization” come from?

The term comes from "GEO: Generative Engine Optimization", a paper first posted to arXiv on 16 November 2023 by Pranjal Aggarwal and five co-authors from IIT Delhi and Princeton University. It was published at the KDD 2024 conference in Barcelona.

KDD is one of the main academic conferences for data mining, so the paper went through peer review before publication. It did two things at once: it named the practice, and it built a benchmark to test it. That benchmark, GEO-bench, is why later studies and tools cite the paper as the field’s starting point.

The paper frames generative engines as a threat to the third stakeholder in search. Users get better answers and engines get more traffic, but website owners lose the click, and they have no control over whether the engine credits them. GEO was proposed as the creator’s side of that exchange.

The experimental setup explains what the paper’s numbers can and cannot tell you:

• Queries: GEO-bench holds 10,000 queries across 25 domains, such as arts, health and games, with 80% informational queries.
• Sources: each query was paired with the cleaned text of the top 5 Google search results.
• Model: the answer was written by gpt-3.5-turbo from those 5 sources.
• Sampling: 5 responses were sampled per query at temperature 0.7, and results were averaged over 5 random seeds.

Visibility was scored two ways. Position-Adjusted Word Count counts how many words of the answer come from a source, weighted towards sources cited earlier. Subjective Impression has a model rate each citation on 7 aspects, including relevance, influence on the answer, uniqueness of the material, subjective position and subjective count. In the test, one of the 5 sources was rewritten with a method and its score was compared with the unmodified version.

Which 9 methods did the original GEO paper test?

The paper tested 9 content rewrites. Each changes the text of one source without adding new facts about the topic:

Quotation Addition scored highest at 27.2 against a 19.3 baseline; Keyword Stuffing scored 17.7, below baseline. The paper summarises its top three methods — Cite Sources, Quotation Addition and Statistics Addition — as a 30–40% relative improvement on Position-Adjusted Word Count and 15–30% on Subjective Impression, and reports the best case as a visibility increase of over 40%.

The Keyword Stuffing result is the one SEO practitioners remember. A tactic built for term-matching ranking systems made a source less visible to a model that reads for meaning.

The authors repeated part of the test on Perplexity.ai, a deployed engine. Quotation Addition again performed best, with a 22% improvement in Position-Adjusted Word Count, and Keyword Stuffing scored 21.9 against a 24.1 baseline.

Why did citing sources hurt the top-ranked result?

The paper split its results by where the optimized source ranked in Google’s top 5. The effect reversed at the top:

Citing sources raised a rank-5 source’s visibility by 115.1% and lowered a rank-1 source’s by 30.3%. The authors read this as good news for smaller sites: a lower-ranked page that adds evidence can take share from the page above it.

The mechanism is share, not absolute quality. The answer has a fixed length, so words credited to one source come out of the others. A top-ranked source already holds a large share, and a rewrite shifts the model’s attention in ways that can reduce it. The finding holds inside a fixed set of 5 retrieved sources; it says nothing about how a page gets into that set.

How does a generative engine pick the sources it cites?

A generative engine picks sources in 4 stages: retrieval, selection, synthesis and citation. It expands the prompt into sub-queries, fetches candidate pages from an index, keeps the most relevant passages, writes an answer from them and links the sources behind the claims it used.

1. RetrievalExpands the prompt into sub-queries and fetches candidate pages from a search index.
2. SelectionScores passages from those pages and keeps the ones that answer a sub-query.
3. SynthesisWrites the answer from the kept passages, paraphrasing and combining them.
4. CitationAttaches links or names to the claims, crediting some sources and not others.

Every AI citation is the end of this chain. A page that fails at stage 1 cannot be cited at stage 4, however good its prose, and the fix for each stage is different.

Stage 1 — Retrieval: query fan-out and grounding

Google describes its AI features as using retrieval-augmented generation, “also known as grounding”: the model answers from documents fetched at query time rather than only from what it learned in training. Before fetching, the engine runs query fan-out — it breaks one prompt into several related sub-queries and searches for each at once. Google’s own example is a question like “how to fix a lawn that’s full of weeds”, which fans out into searches about weed types, treatments and lawn care. A page enters the candidate set by being indexed and relevant to one of those sub-queries. Retrieval fails when a crawler is blocked, a page is not indexed, or no sub-query matches what the page covers.

Stage 2 — Selection: passage scoring

The engine does not pass whole pages to the model. It splits candidates into passages — a paragraph, a list, a table row — and scores each against the sub-queries. Only the highest-scoring passages go into the model’s context window. Selection fails when the answer to a question is spread over several paragraphs, buried under an introduction, or phrased so that it does not match the question’s terms.

Stage 3 — Synthesis: the model writes the answer

The LLM reads the selected passages and writes a response. This is the stage the GEO paper tested: all 5 sources were already in context, and the rewrites changed how much of each the model used. Passages with specific numbers, named sources and quotable sentences give the model material to lift. Synthesis fails when a passage is vague, when it repeats what other sources say, or when it contradicts the consensus of the other passages.

Stage 4 — Citation: links and named sources

Last, the engine attaches attribution. AI Overviews show links beside and below the answer, ChatGPT search places inline source links, and Perplexity numbers its citations. A source that shaped the answer is not always the one credited: engines link the passage that best supports a specific claim, and a brand can be named without its site being linked. Citation fails when a page’s claim cannot be traced to a clear, self-contained sentence.

Each stage has its own fixes; the step-by-step playbook covers how to optimize for AI search. Each stage fails for a different reason, and each failure is testable.

What does Google say about GEO?

Google states that optimizing for generative AI search "is optimizing for the search experience, and thus still SEO." Its guide says pages need no new machine-readable files, AI text files, markup or Markdown to appear, and it adds a Search Console setting for generative AI features.

Google published “Optimizing your website for generative AI features on Google Search” in May 2026 and last updated it on 10 July 2026. It is the first Google document written for site owners asking about GEO, and three statements in it set the frame for any work on AI Overviews and AI Mode.

"Optimizing for generative AI search is optimizing for the search experience, and thus still SEO."

"You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search."

The first statement places GEO inside SEO for Google’s engines. The second says Google Search does not require new file formats to show a page in its AI features, so any case for such files has to rest on other engines. The third turns Search Console into a gate: a site excluded there cannot appear, whatever its content.

These statements cover Google only. ChatGPT search, Perplexity and Copilot use their own crawlers, indexes and citation rules, which is why GEO treats each engine separately after the shared foundations. For the full side-by-side, read GEO compared with SEO and AEO. Knowing the mechanism and Google’s position, the next question is what the evidence actually supports.

How strong is the evidence behind GEO techniques?

The evidence is narrow: controlled studies show that content already retrieved changes how often it is cited, but no technique has shown a stable cross-platform effect on being discovered. A July 2026 survey of 45 GEO studies reached this conclusion.

The survey (arXiv 2607.14035) reviewed 45 studies published after the original paper. Three findings from it shape how any GEO claim should be read:

1. Gains are conditional. The original improvements were measured with the source already inside a fixed set of retrieved documents. They show what happens at synthesis, not whether a page gets retrieved.
2. Results vary between runs. The survey reports substantial run-to-run variability, so a single check of an AI answer is not a measurement.
3. Relevance and position lead. Topical relevance and context position are the most reproducible levers across the studies reviewed.

This is why the paper’s statistics and quotation tactics are being retested here on deployed engines rather than taken on trust: a live GEO experiment, currently planned, will apply them to 20 live pages, keep 20 matched pages unchanged as controls, and track citations in ChatGPT and Perplexity for 8 weeks with a public log.

How is GEO success measured?

GEO success is measured by running a fixed set of prompts on each AI engine and recording mentions and citations per answer. The three standard outputs are mention rate, citation rate and share of model against named competitors.

The method behind those numbers — prompt selection, run frequency, and the five metrics that add answer position and sentiment — is set out in the guide to measuring AI visibility. Google adds one first-party source: the Search Console Generative AI performance report, which shows impressions in AI Overviews and AI Mode by page, country, device and date.

The measurement does not change with the label. Whether a team calls the work GEO, AEO or LLMO, the question is the same: how often do AI answers name the brand, and how often do they link to it.