Generative engine optimization

What GEO is, and what six AI models actually do

Most writing about GEO repeats the same 2 or 3 studies at second hand. This page reads them at the source and says what each can and cannot show, then adds our own live checks: real buyer questions, put to the assistants with the web open, every answer and every source kept. What follows comes from 5,190 recorded answers carrying 68,102 citations, and you can open any of them.

Updated 3 Nov 2026 with the research behind the term, read at the source, and our 18 Oct 2026 benchmark of 349 answers.

GEO, AEO, AI SEO, LLM SEO: one job, several names

The vocabulary is unsettled because the field is new. These all describe the same work, and each name emphasises a different part of it.

AI SEO

The broad one. Any work aimed at getting a brand into the answers assistants give.

AI search engine optimization

The long form of the same thing, used when people mean search specifically rather than chat.

Generative engine optimizationGEO

The term the industry settled on for optimising toward systems that generate an answer instead of returning links.

Answer engine optimizationAEO

Emphasises the output: there is one answer, and the question is whether you are in it.

LLM SEO

Names the thing doing the reading. Same work, framed around the model rather than the search box.

AI visibility

The measurable side: whether, and how often, a brand actually appears.

One distinction is worth keeping straight, because it splits the search results in two: “AI SEO” is also used for using AI to do your SEO work, which is a different subject. This page is about the other meaning, getting into the answers.

Where the term comes from: the 2023 GEO paper

The name comes from a research paper, "GEO: Generative Engine Optimization", by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, posted to arXiv on 16 Nov 2023 and accepted at KDD 2024. It defined generative engines as systems that gather several sources and have a language model summarize them into 1 answer, and GEO as a way for content creators to raise their visibility inside those answers.

A July 2026 survey of 45 GEO studies (Olivier Martinez, arXiv 2607.14035) calls those gains valid within the paper's setting but conditional on a source already being in a fixed context: they establish neither organic discoverability nor durable traffic effects. That is the gap live measurement has to cover: which brands real assistants name when a buyer asks.

How a generative engine builds an answer

Worth walking through slowly, because every GEO decision falls out of it. When someone asks a buying question, four things happen in about ten seconds:

1. 1. The question is expanded. The engine does not search your buyer’s words once; it fans out into several rewritten queries. “Best CRM for a small sales team” becomes searches about pricing, about alternatives, about complaints. You are not optimising for one query string; you are optimising for the cloud of queries around an intent.
2. 2. Pages are retrieved and read. This is where the spread between engines is enormous: in our corpus the average number of distinct sites an engine exposes per answer runs from under one to nearly forty. What gets retrieved is decided by ordinary search infrastructure, which is why classic SEO is not dead, it has been demoted to a qualifying round.
3. 3. Brands are lifted from those pages. The engine names companies the retrieved pages describe. It has no opinion of its own to consult beyond what it already believed, and your own site rarely gets to be the witness: the answer leans on third parties saying you exist and what you are for.
4. 4. A shortlist is written, and it is short. Twenty retrieved pages collapse into an answer naming a handful of companies. This compression step is where visibility is won or lost, and it favours brands that are described the same way in many places, because agreement across sources reads as confidence.

One more path matters: when nothing is retrieved at all, the model answers from memory. That is not rare in our corpus, and it is why the last section of this guide is about the memory layer, not just retrieval.

How we measured this

Stated plainly so you can judge it, because most numbers in this field arrive without a method attached.

1. The questions are not ours. They are built from terms Google Keyword Planner reports real monthly volume for in each category, so they are things buyers type rather than things we would like to be asked.
2. Asked live, with the web open, to 7 assistants: ChatGPT, Claude, Gemini, Perplexity, DeepSeek and Grok. Not a cached prompt database.
3. Every answer is kept whole, with the sources the model actually read attached, and published. Nothing here is a summary you have to take on trust.
4. The numbers recompute as more checks run, so this page cannot drift away from the transcripts behind it. Last recalculated 12 November 2026.

The finding that should change what you do

The industry talks about “optimising for ChatGPT” as though winning there carries you elsewhere. Across 718 questions that more than one model answered, it does not.

Ask six assistants the same buying question and they will name roughly 28.3 companies between them, of which under one is common to all of them. Being the answer in ChatGPT tells you almost nothing about Claude, and a tool that checks a single model is measuring a fraction of your exposure.

Watch the disagreement happen

Percentages are easy to nod along to, so here is the thing itself: a real question from our corpus, one full recorded answer, and what every other model named when asked the same thing.

The engines do not read alike either

How many sources each one consults before it answers, and how many companies it tends to name. Per-answer averages across our corpus.

And 19% of all answers cited nothing at all. In those the model is working from what it already believes about your category, which is the hardest state to influence and the strongest argument for being described widely rather than perfectly in one place.

What changes AI answers: our benchmark of 349 responses

On 18 Oct 2026 we sent 50 buying tasks from our own categories (media placements, press releases, Knowledge Panels, review and content removal, Wikipedia, account recovery, verification and AI visibility) to 7 assistants through their APIs, with web search on and no system prompt, 1 answer each: 349 answers, 1 call failed. It is a separate study from the live checks above, and it scores 3 things.

What it shows

Limits, stated plainly: 1 answer per task and assistant, and the 95% intervals treat answers as independent, which understates the uncertainty; cells under about 20 answers can move by 20 points between identical runs. API answers are not what each consumer app shows. A vendor missing from our list of 103 counts as "no vendor". The tasks cover our own categories, so the results describe those markets, not every industry.

GEO vs SEO, row by row

Most comparisons of these two are vibes. The differences that change what you should do are specific:

The practical reading: keep the SEO you have, because retrieval runs on it. Then point the new effort at the row that changed most, which is where the signal lives.

GEO vs AEO vs SEO: 3 names, 3 lenses

The terms overlap, as the list above says. What separates them is what each puts first, which is what people searching "what is GEO" or "GEO SEO" usually want to know.

What actually moves a brand into an answer

Three things this industry keeps repeating

What the research agrees on, and what it does not show

Each study below measured something different, and none measured your category. Read together, they still point the same way.

What to do about it, step by step, is in the LLM SEO playbook: the buyer-question method and how to measure Google and AI separately.