What is LLM SEO?
The mechanics rather than the vocabulary. How a model actually reaches a page, why the same site can be visible in one assistant and absent from another, and what of that you can influence.
LLM SEO is search engine optimization aimed at large language models rather than at a results page. The full form is large language model search engine optimization. It differs from ordinary SEO in one way that matters: each model reaches the web differently, so the same page can be visible in Claude and absent from ChatGPT.
- Full form
- Large language model search engine optimization. Also written LLMO.
- The two routes
- Training data, which is frozen, and live retrieval, which is not.
- Why it is separate
- It is per model. GEO and AEO describe outcomes; this describes plumbing.
- What you can move
- Retrieval. Training data is not something you can be added to on request.
A model reaches your page by one of two routes, and they behave nothing alike
Almost every confusion in this subject dissolves once you separate these. When an assistant says something about your company, the information got there one of two ways.
Route one is training data. The model was trained on a snapshot of the web taken before a cutoff date and the knowledge is baked into the weights. You cannot be added to it on request, you cannot correct it, and nothing you publish today changes what a model was trained on last year. Content published after the cutoff is invisible to this route entirely until the next training run, whenever that is.
Route two is live retrieval. When the assistant searches the web mid conversation, fetches pages and writes an answer from what it just read, this is retrieval augmented generation. It runs on today's web, it responds to changes within a crawl cycle, and it is the entire practical surface of LLM SEO.
The consequence is worth stating flatly, because a large amount of nonsense in this category depends on blurring it: anyone offering to get your brand into a model's training data is selling something that does not exist. What can be influenced is whether you are found, fetched and used at the moment the model goes looking, which is a different and entirely tractable problem.
Why four assistants disagree about the same brand
Ask four assistants the same question about your company and you can get four different answers, including some where you are absent. That is not randomness, or not only randomness. Each one reaches the web through different plumbing, and knowing which is which tells you where a fix will land.
ChatGPT searches through Bing
Which produces the most common surprise in a first report: a site that ranks respectably on Google and thinly on Bing can be strong in classic search and close to invisible here. If you have never checked your Bing coverage, that is the first place to look before concluding anything else.
Claude fetches live and cites explicitly
It goes and gets pages during the conversation and names what it used. That makes it the clearest read on whether your own pages are reachable and quotable right now, because there is the least indirection between your server and the answer.
Gemini draws on Google's index
So it tends to track your classic Google rankings more closely than the others do. If your Google position is healthy and Gemini still does not name you, the problem is usually the shape of the page rather than its discoverability.
Perplexity runs its own crawler and index
It cites heavily and leans towards research intent, which makes it the engine most likely to surface third party sources naming you: directories, forums, reviews and press. It is the best early read on your corroboration rather than your own site.
What you can influence, and what you cannot
This section exists because the honest list is short, and most of what is sold in this category quietly implies a longer one.
You can control whether the crawler is allowed in. GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google Extended and Bingbot each read their own directive in robots.txt, so this is six decisions rather than one. Sites routinely end up open to two of them and closed to the rest, and it stays the most common single reason a brand is missing.
You can control what arrives when it does. These crawlers overwhelmingly read the raw response rather than running your scripts, so anything assembled in the browser is invisible to them. Moving the templates that carry your words to server rendering or prerendering is normally the largest available gain on a modern stack, and it is a build change rather than a writing one.
You can control whether a passage can be used. Answers near the top, headings that match the question, self contained paragraphs, dates and named authors so the model can tell when it was written and by whom.
You cannot control training data, model weights, or which sources a model prefers. You cannot buy inclusion, and there is no ranking to hold. What third party corroboration does is raise the odds, not set them, and any vendor describing it as more deterministic than that is describing a product that has not been built.
LLM SEO compared with GEO and AEO
These three are the most commonly confused terms in the category, and they are frequently used interchangeably by people selling all three. They describe different things.
| LLM SEO | GEO | AEO | |
|---|---|---|---|
| What it describes | The mechanism: how a model reaches and reads your pages | The outcome: being named in generated answers | The outcome: being the one extracted answer |
| Scope | Per model, per crawler, per index | Across every generative engine at once | Any answer surface, including non AI ones |
| Central question | Can this model find, fetch and parse me? | Am I named when the answer is written? | Is the answer given mine? |
| Typical first fix | robots.txt and server side rendering | Corroboration and quotable structure | Heading as question, answer directly beneath |
| Where it fails quietly | One crawler blocked while the others are fine | Named on some questions, absent on others | Ranked well and never extracted |
| Measured by | Per model prompt sets, plus crawler and render checks | A rate across engines over a prompt set | Answer ownership for target questions |
In practice they are one workstream. The reason the vocabulary is worth keeping straight is diagnostic: "we are invisible in ChatGPT but fine in Claude" is an LLM SEO problem with a specific cause, and treating it as a general GEO problem sends you off writing content when the actual fix was one line in a file.
How to do LLM SEO, per model
The first three steps are mechanical, checkable and free. Do them before commissioning anything, because they are also the most common causes.
- 01
Check every AI crawler separately
Not "is my robots.txt fine" but "is each of these bots allowed": GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, Google Extended, Bingbot. One blocked bot is one assistant that cannot see you while the others can. The AI crawler checker lists them individually.
- 02
Confirm the page renders without JavaScript
Fetch your own page with scripting disabled and read what comes back. If it is a loading state, every crawler that does not render is receiving that. The render gap checker shows both versions together.
- 03
Check your Bing coverage, not only Google
ChatGPT searches through Bing, so Bing is not a legacy concern here, it is the retrieval layer for the most used assistant in the world. Being absent from it is the single most underrated cause of AI invisibility and almost nobody checks it.
- 04
Date and sign what you publish
A visible date, a machine readable modified date, and a named author with a profile that resolves. Models weight recency and provenance when choosing between sources that otherwise say the same thing. Check it with the freshness checker.
- 05
Track a named prompt set per model
The same questions, asked of each assistant separately, recorded over time. Per model matters: an aggregate score across four engines hides the case where you are strong in three and absent from one, which is the most actionable finding there is.
- 06
Build corroboration where each engine reads
Perplexity surfaces third party sources heavily, so directory, review and press coverage shows up there first and earliest. It is the cheapest place to see whether your corroboration work is landing before the slower engines reflect it.
How LLM SEO is measured
The unit is a prompt set per model, and the per model part is the whole point. A single blended AI visibility score across four assistants averages away the only finding that tells you what to do next, which is that one of them cannot see you.
Record four things for each question and each engine: whether you were named, whether you were linked, who was named instead, and which sources the answer drew on. Named and linked are separate outcomes with separate fixes, and the competitor named in your place is reliably the line that gets forwarded internally.
Then hold the deterministic half separately, because it is the half that is actually verifiable: crawler access per bot, render output, schema validity, freshness and attribution signals. Those do not vary between two asks, any third party can check them on your live pages, and they are what you point at when somebody asks whether the work was real.
On attribution, the honest position: assistants pass very little referrer data, so traffic that started in an AI answer often arrives looking like direct. Anyone quoting a precise revenue figure attributed to LLM visibility is modelling rather than measuring, and the useful version of that claim always names its assumptions.
Check the mechanical half yourself
Per bot access and render output are the two things that decide whether any of the rest matters. Both free, both about a minute.
Related reading: generative engine optimization, answer engine optimization, how to rank on ChatGPT, and why there is no rank, AI search optimization, surface by surface, what llms.txt is, and what v2 changed, what AI visibility means and how it is measured, the AI visibility tools, priced and scored.
Questions people ask about LLM SEO
What is the full form of LLM SEO?
Large language model search engine optimization. LLM stands for large language model, the class of system behind ChatGPT, Claude, Gemini and Perplexity. It is also written LLMO, for large language model optimization, which means the same thing.
What is the difference between LLM SEO and GEO?
LLM SEO describes the mechanism and GEO describes the outcome. LLM SEO asks whether a specific model can find, fetch and parse your pages, which is a per model, per crawler question. GEO asks whether you end up named in generated answers across engines. In practice they are one workstream, and the distinction earns its keep diagnostically: being invisible in ChatGPT while fine in Claude is an LLM SEO problem with a specific cause, not a general content problem.
How does LLM SEO actually work?
It works on live retrieval rather than on training. When an assistant answers a question it frequently searches the web, fetches pages and writes from what it just read. LLM SEO is the practice of making sure that at each of those steps you are findable, fetchable and quotable: the crawler is allowed in, the page renders without JavaScript, and the passage answers the question in a form that can be lifted.
Can I get my brand into a model's training data?
No, and anyone offering to is selling something that does not exist. Training data is a snapshot taken before a cutoff and frozen into the weights; you cannot be added on request and nothing you publish today changes what a model was trained on last year. What you can influence is live retrieval, which is where essentially all practical work in this field happens.
Why does ChatGPT not know about my company when Google ranks me well?
The most common cause is Bing. ChatGPT searches through the Bing index rather than Google's, so a site that ranks respectably on Google and thinly on Bing can be strong in classic search and close to invisible in the most used assistant there is. The second most common cause is that your pages render client side and the crawler received an empty shell.
Is LLM SEO different from ordinary SEO?
It shares the foundation and diverges above it. Both need pages that are reachable, indexed and well structured. Ordinary SEO then competes for a position in a list against one algorithm; LLM SEO deals with several models that each reach the web differently, where the same page can succeed in one and fail in another, and where there is no position to hold at all.
Does llms.txt do anything?
Not much yet, and it costs almost nothing to have. It is a proposed file for telling models what a site contains and how to use it, and adoption by the major assistants is still limited, so treat it as cheap insurance rather than a lever. Anyone presenting llms.txt as the fix for AI visibility is skipping past the two things that genuinely decide it, which are crawler access and rendering.
Find out which models can actually see you
The audit checks crawler access per bot and what each one receives, against your live pages. Free, no card, and the report is yours whether or not we ever speak.