AEO vs SEO vs GEO vs LLMO: what each one means, and which one you actually need

Four names for one worry

Every one of these acronyms is an answer to the same question: when a buyer asks, do you come up?

The buyer’s question has not changed. Who should we consider for this, and why? What changed is who answers it. It used to be a search engine that answered with a list and let the buyer do the reading. Now it is often an assistant that does the reading for them and hands back a paragraph.

A paragraph has room for three or four names. A page of results had room for ten, plus ads, plus the second page for the patient. So the stakes of being left out went up at the same moment the reasons for being left out got harder to see.

The acronyms are the industry’s attempt to name the new work. They were coined at different times, by different people, for slightly different surfaces. That is why they overlap, and why two agencies can sell you the same work under different names. The useful thing is not to pick a favourite acronym. It is to understand the surface each one points at, because the surfaces really do behave differently.

SEO: be ranked

Search engine optimization is the work of ranking in a list of links.

The surface is the results page. The buyer types a query, the engine returns a ranked list, and your job is to be high on it for the queries your buyers use. The levers are old and well understood: pages that answer the query, a site that loads fast and can be crawled, and other sites that link to you because you are worth linking to.

SEO is measurable in the plain sense. You can see your position for a query, the impressions, the clicks, and the visits. The feedback loop is weeks. The main risk is ranking for things nobody who buys ever types.

The reason SEO is not dead, whatever the headline says, is simple. When an assistant searches the web to answer a question, it mostly reads pages the search index already ranks. A page nobody can find is a page no assistant reads. SEO is the floor the other three stand on.

AEO: be the answer that gets quoted

Answer engine optimization is the work of being the one answer lifted out of a page.

The term grew up with featured snippets and voice assistants: the boxed answer above the links, the one sentence a smart speaker reads aloud. There is only one slot. The engine picks a page, lifts a passage, and shows it with a source line under it.

The levers are about shape more than authority.

• Headings phrased the way a person asks the question.
• A direct answer in the first sentence under the heading, before the nuance.
• Lists, steps and tables where the answer really is a list, a set of steps or a comparison.
• Structured data, such as FAQ markup, that tells a machine which text is the question and which is the answer.

AEO is still close to SEO: the page usually has to rank before it can be lifted. What it adds is the discipline of writing for extraction. It is also the part of this list that transfers most directly to AI assistants, because a page that answers a question cleanly is a page an assistant can quote cleanly.

GEO: be cited inside a written answer

Generative engine optimization is the work of being one of the sources a generated answer draws on.

The surface here is an answer written by a model, often with numbered sources beside it: the AI overview on a results page, or an assistant that searched the web before replying. The term took hold after a 2023 research paper from researchers at Princeton and elsewhere, which tested ways a source could raise its visibility inside generated answers. Adding citations, quotations and concrete statistics to a page helped. Stuffing it with keywords did not.

That finding is the heart of GEO. A model writing an answer leans on text it can attribute and trust. Specific claims with sources beat vague claims without them. A page that says what you do, for whom, and with what evidence is easier to use than a page that says you are innovative and customer-centric.

GEO is harder to measure than SEO because there is no position to look up. The answer is written fresh each time, the sources change between runs, and two people asking the same question can get different answers. That is not a reason to give up measuring. It is a reason to measure with more than one answer, and to write down how many.

LLMO: be remembered

Large language model optimization is about what a model knows about you before it searches anything.

Every assistant can answer from memory: from what it absorbed in training, with no search at all. Ask a question with web search off and you are reading the model’s memory. That memory has a cutoff date, it was learned from whatever was written about you across the web up to then, and you cannot edit it directly.

LLMO is the slowest of the four. Changes show up when a model is retrained on newer text, which you do not control and cannot schedule. The levers are indirect: a clear, consistent description of what you do, repeated in the places models learn from. Your own site, yes, but also directories, reviews, press, talks, podcasts and the pages of people who write about your category.

The main risk is not absence. It is wrong facts said with confidence. A model that remembers your pricing from two years ago, or your old positioning, or a product you retired, will repeat it to a buyer in a calm, certain voice. Not being named costs you a shortlist. Being named wrongly can cost you the deal after it.

Not being named costs you a shortlist. Being named wrongly can cost you the deal after it.

Side by side

The four overlap more than the acronyms suggest. The table shows where they differ.

Read down any column and you will notice the levers keep pointing at the same thing: a site that says clearly what you do, for whom, with evidence, in the words your buyers use. That is not a coincidence. All four are ways of making the same truth easy for a machine to find and repeat.

The split that actually matters: memory or the live web

Forget the acronyms for a moment. Ask one question of every answer: did the assistant remember this, or did it read it just now?

When an assistant answers from memory, it is telling you what the web said about you up to its training cutoff. When it searches first, it is telling you what the pages it found today say. Those are two different problems with two different fixes, and they move at two different speeds.

If you are named from memory but not with the live web, your reputation is fine and your pages are the problem: the assistant searched, read what ranks for the question, and your pages were not among them or did not answer it. That is SEO and AEO work, and it can move in weeks.

If you are named with the live web but not from memory, your pages are doing their job and your footprint is thin: not enough of the web says clearly who you are. That is LLMO work, and it moves on the model’s schedule, not yours.

If you are named in neither, start with the pages, because they are the part you control and the part that moves first. If you are named in both but with wrong facts, fix the facts at the source, everywhere they appear, and check again next month.

So which one do you need?

Usually not a new discipline. Usually the next gap in the order the answers are built.

If buyers cannot find your pages in ordinary search for the problems you solve, start with SEO. Nothing downstream works without it, and an assistant that searches will not read a page the index does not surface.

If you rank but your pages bury the answer under positioning, do the AEO work: one page per question your buyers ask, the answer in the first sentence, the evidence under it, structure a machine can read. This is usually the cheapest improvement on the list and the one assistants reward soonest.

If you rank and answer cleanly but written answers cite other people, do the GEO work: make your claims specific and sourced, publish the numbers and methods behind them, and earn mentions on the pages assistants already cite for your category.

If you are well represented today but models name you wrongly, or not at all, from memory, that is LLMO: the slow, patient work of saying the same true thing about yourself in many places. Start it now, because it pays out on a schedule you do not set.

What you should not do is buy all four as separate services with separate reports. They are one job seen from four angles, and the only way to know which angle comes first is to look at the answers.

Measured, not claimed

An AI visibility score without its working is a number someone would like you to act on.

The market for this work is young, and young markets sell scores. A visibility score of 62. A share of voice of 14%. Up 9 points this month. Some of those numbers are careful. Some are one answer per question, run once, on one assistant, and rounded into a trend.

Before you trust any number about AI visibility, yours or a vendor’s, ask four things.

• Which questions? Buyer questions in buyer language, or questions that contain your name?
• Which assistants, and was each answer from memory or with the live web?
• How many answers sit behind the number? One answer per question is a sample of one.
• On what date? Answers change as models update and the web changes.

A question that includes your brand name will almost always come back with your brand in it. That tells you the assistant can read a question, not that buyers will find you. The honest test uses the questions a buyer asks before they know you exist.

And because answers vary from one run to the next, one answer is an anecdote. Ask each question more than once. Write the count next to every percentage. When the sample is small, say so, in the report and in your own head.

The one-hour check

You do not need a tool to see where you stand. You need ten questions, a sheet, and an hour.

Run it like this.

• Write ten questions your buyers ask before they have heard of you. Use their words, not yours. “Who are the best firms for this problem at our size?” rather than anything with your name in it.
• Open ChatGPT, Claude and Gemini in a clean session: logged out, or in a new chat with memory and personalisation off, so the answers are not tuned to you.
• Ask each question twice in each assistant: once with web search off, once with it on. Where you cannot switch search off, note that the answer searched.
• For every answer, write down four things: named or not, whether the facts about you are right, who was named instead, and which sources were shown.
• Count. What share of answers named you from memory? With the live web? Who was named most often in your place?

Sixty answers, give or take, in about an hour. It is small. It is also more honest than most dashboards, because you know exactly where every number came from.

The count tells you which fix comes first. Missing with the live web means your pages. Missing from memory means your footprint. Wrong facts mean a cleanup. And the names that keep appearing in your place are your real competitors in the answer, which is not always the same list as your competitors in the market.

Run it again in a month with the same ten questions. The change between the two sheets is the only trend worth reporting.