In short
- GEO
- Getting generative engines to cite and recommend you.
- AEO
- The same, with the emphasis on the standalone answer fragment.
- mentioned · cited · chosen
- The three outcomes, counted separately: your name appears, your source is cited, you are recommended.
- query fan-out
- The sub-questions an engine splits one question into before it answers.
What is this field actually called?
Six names for largely the same thing. There is no settled distinction between GEO, AEO, LLMO, GAIO, AIO and AI SEO; anyone claiming otherwise is usually selling the word they picked themselves. We use GEO, and we do not pretend that is a technical choice.
| Term | What it means |
|---|---|
| GEO — Generative Engine Optimization | Arranging your site, your entity and your technical setup so generative engines cite and recommend you. The most-used name, and ours. |
| AEO — Answer Engine Optimization | Writing your page so a direct answer can be lifted out of it. In practice the same as GEO, with slightly more emphasis on the standalone answer fragment. |
| LLMO — Large Language Model Optimization | Same idea, with the model in the name instead of the engine. Less common, no method of its own. |
| GAIO — Generative AI Optimization | An umbrella term that coincides with GEO. Shows up mostly in sales material, not in research. |
| AI SEO | The loose collective name for everything above, used mostly by people coming from classic SEO. |
| AI visibility | Not the discipline but the outcome: how far AI assistants mention, cite or recommend you. |
| SEO — Search Engine Optimization | The older field: ranking higher in a list of ten blue links. Overlaps GEO technically, differs in what gets measured. |
| answer engine | A system that gives one answer rather than a list of sources. Perplexity describes itself this way. |
| generative engine | An engine that writes its own answer instead of quoting one. ChatGPT, Gemini, Claude. |
What exactly gets measured?
These measures, and they are not interchangeable. Being mentioned, being cited and being chosen are three different things, and a tool that sums them into one number hides which of the three is broken.
| Term | What it means |
|---|---|
| mentioned | Your name appears in the engine's answer. The lowest of the three outcomes. |
| cited | A source of yours is cited by the engine underneath its answer. |
| chosen | You are recommended, not merely listed. The only one of the three that produces revenue. |
| share of voice | What share of all brand mentions in a set of answers is yours. Counts mentions, not recommendations. |
| share of model | The same measure, split per engine. Useful because you can win in ChatGPT and be invisible in Perplexity. |
| share of choice | How often you are the recommended answer, set against the rest. Predicts revenue; share of voice does not. |
| citation rate | The share of answers in which a source of yours is cited. |
| visibility score | One composite number across several measures. Without a published weighting it cannot be checked, and then it is an opinion with a decimal point. |
| sentiment | Not whether you are mentioned but in what terms: the cheap option, the reliable one or the awkward one. |
| prompt volume | How often a given question is put to an AI. No reliable figures exist for this; anyone showing them is estimating. |
| zero-click | The visitor gets their answer without clicking through. The reason measuring traffic no longer suffices. |
| retest | Asking the same prompts again, in the same market and language. The only proof that a fix changed anything. |
How does an AI arrive at its answer?
Like this, and it differs fundamentally from a search engine. An engine fetches pieces of text, weighs them and writes an answer from them — it does not show a list. That is why you measure on prompts and not on keywords.
| Term | What it means |
|---|---|
| prompt | The question as a person types it: a whole sentence, not a keyword. "Who do you recommend in my area", not "plumber Ghent". |
| query fan-out | The sub-questions an engine splits one question into before it answers. So you are measured on questions the user never asked. |
| RAG — retrieval-augmented generation | The model looks up sources first, then writes its answer. Without retrieval it answers from memory, and then you do not count. |
| grounding | Anchoring the answer in fetched sources rather than in what the model thinks it knows. |
| chunking | Cutting a page into pieces for storage. Badly structured text gets badly cut, and then no piece stands on its own. |
| embedding | A piece of text turned into numbers, so a machine can judge which pieces resemble each other. |
| reranking | Re-ordering the fetched pieces before the model uses them. This is where it is decided who makes the answer. |
| context window | How much text a model can hold at once. If your key sentence sits deep in the page, it can fall outside. |
| token | The unit a model counts text in, roughly a word fragment. Determines what still fits in the window. |
| hallucination | An answer that is right in form but wrong in fact. About your company this happens mostly when your data contradicts itself elsewhere. |
| live search | The engine actually fetches pages while answering. If it does not, it answers from memory and the measurement does not count. |
| AI Overviews | The summarised answer at the top of Google, above the normal results. |
| AI Mode | Google's conversational mode, where you follow up instead of searching again. |
| prompt injection | Text on a page that tries to give an AI instructions. A risk for anyone letting agents roam the open web. |
What does an AI crawler see of your site?
Text and structure, not design. Block it and you do not exist for that engine — and that is the first question to settle, before all the others.
| Term | What it means |
|---|---|
| crawler | The program an engine uses to fetch your page. There are over twenty, and they do not all obey the same rules. |
| GPTBot | OpenAI's crawler, fetching pages for training and for answering. |
| ClaudeBot | Anthropic's crawler. |
| PerplexityBot | Perplexity's crawler, which shows its sources next to the answer. |
| Google-Extended | Not a separate crawler but a switch in robots.txt: it decides whether Google may use your page for Gemini, without touching ordinary search. |
| robots.txt | The file where you say, per crawler, what it may read. Close it and you disappear from that engine. |
| llms.txt | A proposed file telling AI assistants what you do and when you fit. Young, and not yet a standard. |
| structured data | Facts in machine-readable markup instead of running text. The difference between being able to read something and knowing it. |
| schema.org | The vocabulary that markup is written in, shared by Google, Microsoft and the rest. |
| JSON-LD | The form that markup usually takes in a page: a block of JSON in the head. |
| FAQPage | The schema type for questions and answers. The questions must also be visible on the page; code-only breaks the rules. |
| DefinedTerm | The schema type for a term and its definition. The machine-readable form of a glossary like this one. |
| SSR — server-side rendering | The page arrives fully filled from the server. With client-side rendering, a crawler that runs no JavaScript sees an empty shell. |
| semantic HTML | Elements that say what something IS — article, nav, table, dl — rather than how it looks. That lets a model skip the noise. |
| IndexNow | A way to tell search engines immediately that a page changed, instead of waiting for them to come by. |
| hreflang | The pointer saying which page is the version in another language. Pointing at a page that does not exist is worse than having none. |
What does an engine base its trust on?
On signals that sit largely outside your own site. That is the uncomfortable part: your page can be perfect and you still will not be chosen if nobody else writes about you.
| Term | What it means |
|---|---|
| entity | Your company as a machine knows it: one name, one address, one set of facts that are the same everywhere. |
| entity confusion | When an engine mixes you up with a same-named company, or finds your data contradictory and names someone else instead. |
| knowledge graph | The network of facts and connections in which a search engine or model organises the world. |
| Wikidata | The open fact database many models draw entities from. Not being in it is a gap you can close yourself. |
| NAP-consistentie | Name, address and phone number written identically everywhere. Differences sow doubt, and doubt costs you the recommendation. |
| E-E-A-T | Experience, expertise, authoritativeness and trust: Google's framework for whether a source can be relied on. |
| topical authority | Authority on one subject, built by covering it fully rather than writing one page about it. |
| source authority | How heavily an engine weighs a source. A trade publication counts differently from your own product page. |
| freshness | How recent a source is. On questions where currency matters, a new piece beats a better old one. |
| digital PR | Getting others to write about you in places an engine trusts. Often the strongest lever, and the slowest to build. |
| earned media | Coverage you did not buy. Weighs more in AI answers than anything you publish yourself. |
| UGC — user generated content | What users write themselves: reviews, forum threads, Reddit. Models draw on this remarkably often. |
| Decision Drivers | The factors that recur for those who do get chosen: source quality, freshness, entity, citations, local context and human signals. |
| design trust | Whether your site looks trustworthy to a first-time visitor. Affects the human who clicks through, not the model. |
What changes when AI starts acting?
Then being cited is no longer enough. An agent booking or ordering on someone's behalf needs data it can act on, not just text it can read.
| Term | What it means |
|---|---|
| AI agent | A system that does not only answer but takes steps: look up, compare, book, order. |
| agentic workflow | A chain in which several agents hand work to each other, with conditions and feedback loops. |
| agent-ready | The state in which an AI agent can not only read your data but act on it. |
| MCP — Model Context Protocol | An open convention letting an AI model reach tools and data sources. |
| tool calling | A model calling a function or service itself instead of only talking about it. |
| agentic commerce | Buying where the agent makes the choice and the human only approves. Anyone absent there is not considered. |
| verified entity record | Your profile as a machine-readable record an agent can request. Not a page to read, but a fact to verify. |
Which words are ceeme's own?
These, and we set them apart because they are not industry vocabulary. Anyone meeting them elsewhere is meeting us — and that should be clear from where they sit.
| Term | What it means |
|---|---|
| agent | At ceeme: a part that performs one task and hands the result on. The plural is agents. |
| agents | The crew: all the agents together that do the work. Seven run today, six are planned — that comes from the code, not from a list. |
| approval gate | The gate between a proposed change and a published one. Only you open it, and there is always a way back. |
| REWORK | The loop back from Judge to Scribe. If Judge rejects, Scribe rewrites, until it is right or a human decides. |
| rollback | The way back. Any published change can return to the previous version. |
| need question | A prompt in which someone describes their problem without knowing which kind of provider solves it. This is where the volume is. |
| exclusion | A prompt that rules out a kind of provider — "without a photographer". It measures nothing: the competition is gone before the answer. |
| jargon lock | The ban on your own trade words in the need questions. A place name is exempt. |
| execution | Actually applying a fix instead of recommending it. The line between a tracking tool and ceeme. |
| tracking tool | A tool that establishes whether your brand appears in AI answers, and stops there. Measuring is a craft; it is not fixing. |
| engine | An AI assistant that answers a person's question. The term stays English in every language. |
| schema | With us, shorthand for the structured data on a page, following schema.org. |
What is allowed and what is not?
The line sits in one place: is it visible who is writing. Structuring your own site, writing citably, asking for real reviews and joining a forum openly as a brand is allowed. What is not allowed is passing yourself off as someone else — and that is on the EU blacklist, not in our terms and conditions.
That is why these words are here and not only in our code. Anyone wondering whether this craft can be honest deserves a rule they can check rather than a promise they have to believe.
| Term | What it means |
|---|---|
| astroturfing | Pretending a brand message comes from ordinary users: fake accounts, staged "spontaneous" enthusiasm, coordinated inauthentic amplification. Banned in the EU — it is on the blacklist of the unfair commercial practices directive, which names posing as a consumer explicitly. We do not do this, and our rule is tighter and more checkable than the word: we never post under any name but yours. |
| community seeding | The market term for building mentions in forums and communities so AI answers meet you there. The word says nothing about the method — and that is exactly where the difference sits: seeding with fake accounts is astroturfing, seeding under your own name is simply taking part. The latter is what we call a community-response workflow. |
| Reddit citations | Mentions in Reddit threads that an AI engine cites as a source. Reddit is the most-cited source in AI answers, which makes it the heaviest lever outside your own site — and the one place you cannot publish yourself. We measure them as the driver `community_signals` and work on them through a community-response workflow, always under your name. |
| forum outreach | Approaching a community or trade forum to be mentioned or included there. Related to earned media, but with a different gatekeeper: not an editor but a moderator and an audience that spots advertising instantly. That is why our test is not "is this allowed" but "is this a valuable contribution". |
| community-response workflow | Allowed, and for us an execution channel: answering a real question in a forum, community or Q&A openly as a brand, with real knowledge and an identifiable sender. Every contribution goes through the same approval gate as an email to a journalist, for the same reason: sent is sent. |
| valuable contribution | The test that decides whether a community answer may go out: would this text be useful to whoever reads it — even if it came from nobody? If the answer is no, it is advertising in a forum and it does not go out, however honestly the sender is named. |
| bought reviews | Paying for a positive review — or for a negative one about someone else. Banned in the EU and ruled out with us, even if a client asks. What we do instead: set up a rhythm in which real customers are asked for a real review. Freshness weighs heavily in AI answers, and it can be earned honestly. |
| prompt injection | Text on a page that tries to give an AI instructions. A risk for anyone running agents on the open web. |
| transparent sender | The hard condition under everything ceeme publishes outside your own site: it says who is writing. That is not modesty but self-preservation — a service that hides who posts today is tomorrow the service that does it against you. |
Frequently asked questions
What is the difference between GEO, AEO and LLMO?
In practice, almost none. There is no settled distinction between the three, and the differences you meet usually come from the marketing of whoever uses the word. GEO is the most common name; AEO leans a bit more on the standalone answer fragment; LLMO puts the model in the name. Anyone claiming these are three disciplines is selling three.
Why do the terms stay English?
Because a translated trade word matches nothing. Write "question fanning" where the world says "query fan-out" and you are using a word nobody types. The term stays English, the sentence around it is your own language.
What is the difference between share of voice and share of choice?
Share of voice counts how often you are mentioned, share of choice how often you are the recommended answer. The first is what most tools show; the second predicts revenue. You can double the first without the second moving.
Are these definitions yours or the field's?
Both, and it is marked. The last group is ours — approval gate, jargon lock, need question — and it is set apart because it is not industry vocabulary. The rest is the common meaning; where we use a word more strictly, the definition says so.
Is it allowed to optimise your own AI visibility?
Yes, and the line sits in one place: is it visible who is writing. Structuring your own site, writing citably, asking for real reviews and joining a forum openly as a brand are all allowed. What is not allowed is passing yourself off as someone else — fake accounts, bought reviews, hidden instructions in the page. The first is a craft, the second is on the EU blacklist of the unfair commercial practices directive. Our own rule is stricter and checkable: we never post under any name but yours.
Do you also post on forums and Reddit?
Yes, as a community-response workflow, and always under your name. Reddit is the most-cited source in AI answers and therefore one of the heaviest levers there is — and precisely the place where you cannot publish on your own site. Every contribution must pass two tests: identifiable sender, and would this text be useful to whoever reads it, even if it came from nobody. After that it goes through the same approval gate as an email: sent is sent.
Why is prompt volume listed if you do not measure it?
Because people ask about it and tools display it. A glossary that only contains the terms you are strong at is a brochure. The definition does say that reliable figures do not exist.
Read on
Free, no account and no card.