Definition
Generative Engine Optimization (GEO) is the practice of editing web content so that AI systems which write answers — ChatGPT, Perplexity, Google's AI Overviews and AI Mode — quote it, cite it and lean on it while composing those answers. It is not simply SEO under a new name, and the sharpest evidence is that the oldest SEO tactic of all runs backwards here: in the research paper that named the field, stuffing a page with keywords lowered its share of the generated answer, from a baseline of 19.5% to 17.8%.
The two disciplines separate because they optimize different objects. SEO optimizes rank — a slot in an ordered list of ten links, where your page and the one below it occupy separate rows and a reader can scroll past one to reach the other. A generative engine returns a single block of prose with citations threaded through it. There is no row three. Your page is either woven into the sentences someone actually reads, or it is a superscript nobody clicks. Before you can optimize for that, you have to define what "visible" means when the result is a paragraph rather than a list — and that definitional work, not the tactics, is what GEO originally was.
The term has an unusually precise origin for a marketing acronym. It was introduced by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande in GEO: Generative Engine Optimization, posted as a preprint in November 2023 and published at ACM SIGKDD in August 2024. The same paper coined generative engine for the class of system being optimized for: a service that retrieves documents and then uses a large language model to synthesize an attributed answer from them, rather than returning the documents themselves. Every later use of the acronym traces back to that paper. None of the competing labels — AEO, LLMO, AIO — has an equivalent.
How It Works
Visibility has to be defined before it can be optimized
In a ranked list, visibility is trivially measurable: you are at position 4, and there is a well-studied relationship between position and clicks. In a generated answer, there is nothing that obvious. Five sources may be cited in one paragraph, interleaved, with one supplying a full sentence in the opening line and another supplying a clause in the last. Which of those is "more visible"? The paper's first contribution is that it had to answer this before it could measure anything, and it proposed three metrics:
Word Count is the share of the answer's words that live in sentences citing your source — if the response runs 200 words and 40 of them sit in sentences attributed to you, your Word Count impression is 20%.
Position-Adjusted Word Count (PAWC) takes that same share and weights each sentence by an exponential decay in its position, so a citation in the opening sentence counts for more than the same number of words at the end. The paper justifies the decay by pointing at click-through studies showing that attention falls off as a power law with rank in search engines — the same intuition, transplanted from a list of links to the sentences of a paragraph.
Subjective Impression gives up on counting words and asks a language model to score seven facets instead: relevance of the cited material to the query, how much the response actually leans on the citation, uniqueness of what it contributes, its perceived prominence, its perceived volume, how likely a user is to click it, and the diversity of the material presented. It is scored with G-Eval, an LLM-as-judge method.
Only the first two are objective. The third exists because word count is a bad proxy for attention, and the authors were candid that it needed a subjective complement.
One number worth doing in your head
The experimental engine fetched the top 5 results from Google for each query and asked a model to answer from them. Five sources dividing one answer means an even split gives each about 1/5 — 20%. The measured baseline with no optimization was 19.5% on PAWC. In other words, an unoptimized page in that setup gets almost exactly its equal share, and the shares sum to one.
That is the structural fact underneath the whole discipline, and it is what most GEO marketing omits. Ranking is not zero-sum in this way: your page moving from position 7 to position 4 does not delete anyone from the results page. A generated answer is a fixed budget of prose. Every percentage point of visibility you gain is a point taken from one of the other four sources cited alongside you.
What the paper actually measured
The evaluation ran on GEO-bench, a benchmark of 10,000 queries split 8,000 / 1,000 / 1,000 into train, validation and test. The queries were pulled from nine sources — MS MARCO, ORCAS-1 and Natural Questions for real anonymized search traffic; AllSouls essay questions, LIMA, a debate set, Perplexity's Discover feed and the ELI5 subreddit for the harder synthesis-style questions a generative engine actually receives; plus GPT-4-generated queries to fill gaps. Each query was tagged across 25 domains (Arts, Health, Games and so on), and the mix was held at 80% informational with 10% each transactional and navigational. Answers were generated with gpt-3.5-turbo, sampled 5 times at temperature 0.7 and averaged over five random seeds.
Nine content edits were then applied to a source and re-scored. The absolute impression scores, against the 19.5% and 19.3% no-optimization baselines:
| Edit applied to the source page | PAWC | Subjective Impression |
|---|---|---|
| No optimization (baseline) | 19.5% | 19.3% |
| Keyword stuffing | 17.8% | 20.2% |
| More unique words | 20.7% | 20.4% |
| Easier to understand | 22.2% | 20.5% |
| More authoritative tone | 21.8% | 22.9% |
| More technical terms | 23.1% | 21.4% |
| Fluency optimization | 25.1% | 21.9% |
| Cite sources | 24.9% | 21.9% |
| Add quotations | 27.8% | 24.7% |
| Add statistics | 25.9% | 23.7% |
Read the table as a ranking of what generative engines reward, and two things stand out. The winners are all forms of making a claim checkable and liftable — a quotation, a statistic, a citation — which is what a system that must attribute its sentences needs in order to attribute one to you. The losers are the tactics aimed at the retrieval layer rather than the reader: keyword stuffing is the only edit that finished below doing nothing, and adding unique words barely moved the needle at 20.7%.
Keyword stuffing failing is the single most useful line in the paper for anyone deciding whether GEO is a real distinction. The paper's own framing is that these methods "offer little to no improvement on generative engine's responses", and the reason is mechanical: keyword density is a signal about lexical matching, and the step it was designed to influence — retrieval — is no longer the step that decides what the reader sees. Something else already retrieved five documents. What happens next is a language model deciding which of the five is worth quoting.
The result that should temper the whole field
The headline numbers above assume you optimize and your competitors do not. The paper also ran the case where every source in the result set is optimized at once, and the outcome inverts by rank:
| Edit, applied to all five sources | Rank 1 | Rank 5 |
|---|---|---|
| Cite sources | -30.3% | +115.1% |
| Add quotations | -22.9% | +99.7% |
| Add statistics | -20.6% | +97.9% |
The paper's reading is that "lower-ranked websites, which typically struggle for visibility, benefit significantly more from GEO", because generative engines weigh presentation and credibility of the text itself rather than the backlink and domain signals a small publisher cannot accumulate. That is a genuinely optimistic finding for small sites. But note what it also says: when everyone adopts GEO, the rank-1 source loses a fifth to a third of its visibility. Universal adoption is a levelling, not a lift, which is exactly the dynamic that eventually neutralized most SEO tactics.
It transferred to a live engine
To check the findings were not an artifact of their own test harness, the authors re-ran a subset against Perplexity.ai, a deployed commercial generative engine. The pattern held: baseline PAWC 24.1%, keyword stuffing 21.9% (below baseline again), adding quotations 29.1%, adding statistics 26.2%. On Subjective Impression, statistics addition reached 33.9% against a 24.7% baseline. That transfer is what makes the paper more than a benchmark exercise — the same edits moved the same direction on a black-box system the authors did not build.
Types
The label is contested, and a reader who searched the acronym mostly needs to know which of the surrounding terms are synonyms. They are not a taxonomy so much as a naming collision:
- GEO (Generative Engine Optimization) — the term defined above, with a citable origin and a published benchmark. Targets systems that generate an answer and attribute it.
- AEO (Answer Engine Optimization) — circulated before generative engines existed, when the "answer engine" meant a featured snippet or a voice assistant reading out one result. It has been retrofitted to AI answers and is now used interchangeably with GEO by most practitioners.
- LLMO (Large Language Model Optimization) — a vendor coinage. Sometimes used to mean optimizing for what a model absorbed during training rather than what it retrieves at answer time, which would be a real distinction; usually used as a synonym for GEO, which makes it noise.
- AIO / GAIO (AI Optimization / Generative AI Optimization) — broader still, and often stretched to cover using AI to produce content, which is a different activity in the opposite direction.
- "AI SEO" — normally means using AI tools to do conventional SEO work, not optimizing for AI answers. The two senses are frequently swapped in the same article.
The practical rule: treat GEO, AEO and LLMO as the same discipline unless someone tells you specifically what they mean by the difference. A vendor drawing a sharp technical boundary between them is usually drawing a boundary around a product tier.
Real-World Applications
Perplexity as a measured target. The paper's Perplexity experiment is the concrete demonstration that these edits do something outside a lab: a real engine, unmodified, with a real user base, shifting its citation behaviour in response to changes in source text. It is also the reason to treat GEO as engine-specific rather than universal — the effect sizes differed from the authors' own harness even though the direction held.
Google telling site owners to do nothing special. Google's AI features and your website documentation, last updated 10 December 2025, states that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that "you don't need to create new machine readable files, AI text files, or markup to appear in these features." That is a decision-relevant primary source, and it is narrower than it first sounds: Google is saying there is no separate technical channel to configure, not that every page is equally quotable. The KDD results sit comfortably alongside it, because everything the paper found to work is a change to the prose a reader sees, not a file a crawler reads.
Measurement, as far as it goes. In June 2026 Google added a Generative AI performance report to Search Console, covering impressions in AI Overviews and AI Mode. It is thinner than the equivalent search reporting: the two surfaces are not broken out separately, two results from the same site count as a single impression, and there is no query-level view. A site owner therefore cannot reproduce anything resembling the paper's metrics from outside — which matters more than it sounds, and is picked up under Challenges.
For what this means for strategy — how AI citation has decoupled from organic ranking, and what carries over from SEO — see GEO vs SEO: what changes when AI answers the query. For the page-level tactics, see how to get cited by ChatGPT and Perplexity. This entry deliberately stops at the definition and the evidence.
Challenges
The evidence base is one moment, frozen. The measurements above were produced against gpt-3.5-turbo answering from the top 5 Google results, in work posted in late 2023. Current engines rerank, issue several sub-queries per question, and use far stronger models. The finding that presentation beats keyword density is likely durable because it follows from where in the pipeline the decision is made; the specific percentages are a photograph of one system and should not be quoted as current effect sizes.
Visibility is not traffic. PAWC measures the share of an answer's words attributable to your page. Nothing in the research connects a point of PAWC to a visitor, a signup or a sale, and the zero-click nature of AI answers means the two can move in opposite directions. A page can be quoted constantly and receive nothing.
You cannot see the query, so the claims cannot be audited. Classic SEO is falsifiable from your own logs: you know the query, the position, the click. In AI answers, you generally see neither the prompt nor the citation decision, and the platform reporting that does exist is aggregated. That combination — high commercial interest, no independent measurement — is the exact condition under which unfalsifiable vendor statistics multiply. Treat any GEO figure without a named methodology and a date as marketing.
Optimizing for a metric that rewards numbers does not reward true numbers. The paper measures visibility, not accuracy; nothing in its evaluation checks whether an added statistic or quotation is correct. "Add statistics" is therefore a finding about what generative engines currently reward, and simultaneously a description of how to make fabricated content more citable. This is the seam where GEO meets hallucinations and content provenance, and it is unresolved.
The gains partly cancel. As the rank table shows, once every source in a result set is optimized the top result loses 20-30% of its visibility and the bottom one roughly doubles. Effect sizes measured on an otherwise unoptimized web are an upper bound on what a late adopter can expect.
Future Trends
The retrieval assumption is already outdated. The paper models a generative engine as fetching a single top-5 list and summarizing it. Google describes AI Mode as issuing multiple related searches behind one question — a "query fan-out" — which means a page competes inside several different retrieval sets simultaneously, each with its own four rivals. Visibility metrics built for one answer with five sources will need to be redefined for an answer assembled from several retrievals.
Whether platform measurement becomes auditable. The June 2026 Search Console report is the first official visibility data of any kind, and it counts impressions without queries or per-surface splits. If query-level and surface-level breakdowns arrive, GEO becomes an empirical discipline in the way SEO is; if they do not, it stays a field where the effect of an intervention cannot be observed by the person paying for it.
The naming will resolve or the field will not. GEO, AEO, LLMO, AIO and GAIO currently compete for the same idea, which is a reliable sign of a young commercial field where naming is a positioning move. The academic term has the advantage of a definition and a benchmark behind it; the marketing terms have the advantage of budgets. Watch which one platform documentation eventually adopts — as of now, Google's own guidance uses none of them.