GUIDE · PART OF GENERATIVE ENGINE OPTIMIZATION
LLM optimization:
what a model knows, and what it looks up.
A language model answers from two places: what it absorbed when it was trained, and what it retrieves at the moment you ask. LLM optimisation is about being present in both. This guide explains the difference, because the two change at very different speeds and need different work.
What is LLM optimization (LLMO)?
LLM optimisation, or LLMO, is the practice of making a brand and its content more likely to appear in answers from large language models, such as those behind ChatGPT, Gemini and Copilot. A model draws on two things: what it learned in training, and what it retrieves from the web when it answers. LLMO covers both. The term describes much the same work as generative engine optimisation, with more attention paid to the model itself.
Key facts
- A model answers from what it learned in training, from what it retrieves at the time, or from both
- Training data has a cut-off. What a model learned changes only when a new version is released
- Retrieval can reflect a change on your site within days
- There is no way to submit a site to a model. You can only be readable, and be described accurately in many places
- Blocking a training bot keeps your content out of future training, which is a choice with a cost
- SEORCE asks the engines themselves, repeatedly, and records what each one says and which sources it used
Last reviewed 15 October 2026 by the SEORCE team.
A model has two kinds of memory.
Treating a language model like a search index leads people astray. It has no index you can inspect, and half of what it says about you was settled before you ever looked.
You can only ask
There's no report of what a model knows about your company. The only way to find out is to ask it, more than once.
Two clocks
What the model learned in training won't change for months. What it retrieves can change this week. A fix may show in one and not the other.
Answers without sources
When a model answers from training, it shows no sources. You can't see where a wrong idea came from, only that it's there.
Training, retrieval, and telling them apart.
Most of the confusion about LLMO goes away once you know which of the two you're dealing with.
What the model learned about you.
During training a model reads a vast amount of text and forms a picture of what your company is. You can't edit that picture. You can influence the next one.
- How you're described across the webYour own pages, profiles, reviews and coverage all feed it.
- ConsistencyThe same facts in many places are easier to learn than conflicting ones.
- PatienceA correction made today reaches the model when a new version is trained.
- The training botsIf you block them, your pages aren't part of what the next model learns.
What the model looks up.
Many assistants now search the web before answering. This is the part you can change quickly, and the part that works most like SEO.
- The search bots need accessThey are separate from the training bots, with their own rules.
- The page has to answer the questionPlainly, near the top, in text.
- The sources it trusts matterReviews, directories and forums are read alongside your own site.
Look for the sources.
You can usually tell where an answer came from by whether it shows any sources. That tells you which kind of fix will work.
- Sources shownThe model searched. Work on access, content and the sources it used.
- No sources shownThe model answered from training. Work on how you're described, and expect it to take longer.
- Track bothRun the same questions over time, and note which answers come with sources.
LLM optimisation, in the order that pays off.
Start with what changes fastest. Identity work matters most over time and shows last.
Let the search bots in
Check that the crawlers behind AI search can read your pages.
AI crawler optimization guide →Answer questions plainly
So that a page retrieved for a question has something to quote.
Answer engine optimization →State who you are, everywhere
The same category, audience and facts on your site and on the profiles you own.
AI brand monitoring guide →Decide on training bots
Allowing them means your content can inform future models. Blocking them is a legitimate choice, with that cost.
Ask, and keep asking
Track what each engine says over months, because the training side moves slowly.
AI visibility tracking guide →Check what your pages tell a model about you.
A model can only learn what your pages state clearly, and only retrieve what its bots can reach. These cover both.
Training and retrieval, side by side.
| Training | Retrieval | |
|---|---|---|
| What it is | What the model absorbed before it was released. | Pages the model fetches when someone asks. |
| How fast it changes | With each new version of the model. | Within days or weeks. |
| What shapes it | How you are described across many sources. | Whether your page can be read, and answers the question. |
| The bots involved | Training crawlers, such as GPTBot. | Search crawlers, such as OAI-SearchBot. |
| How to spot it | Answers that show no sources. | Answers that list their sources. |
Ask the model, and keep the answers.
AI Beacon runs your questions on the engines themselves and records what each one says, with the sources where there are any. Entity mapping shows how each engine defines your company, and Technical Suite checks which bots can read your pages.
Frequently asked questions.
LLM optimisation, or LLMO, is the practice of making a brand and its content more likely to appear in answers from large language models. It covers what a model learned in training and what it retrieves when asked.
Only in emphasis. LLMO pays more attention to the model and its training. GEO covers the whole job. The GEO guide sets out the method for both.
Not by request. There's nothing to submit. Your pages can be part of future training if the training bots are allowed to read them, and nobody outside the model's maker decides what is used.
It depends which side answered. Where the model searches the web, a change can show within days or weeks. Where it answers from training, expect to wait for a new version of the model.
That is a business decision. Blocking them keeps your content out of future training, which protects it and also means the next model learns less about you from your own pages. The search bots are separate, and blocking those can keep you out of AI search.
It might, and it's optional. llms.txt is a proposed convention. It never replaces pages that bots can reach and read.
You can check the inputs for free. The entity SEO analyser shows whether the knowledge bases know your brand, and the AI crawler log analyser shows which AI bots visit your site. What a model says about you has to be asked, repeatedly.
Find out what the models say about you.
Add your site and a few questions. You'll see how each engine describes you, and whether it looked anything up.
Not ready for an account? Use the free browser extension or any of the free tools.