What is LLMO (Large Language Model Optimization)?

LLMO, optimization for an LLM (large language model), sounds like a new discipline, but it describes almost the same work as GEO (Generative Engine Optimization). So the useful question is not what LLMO is in the abstract but what it adds to GEO. The short answer is one thing: an assistant can talk about your products without having visited your site that day, using what it learned in training, and that is the part the name LLMO puts up front.

How a brand gets into an answer

A language model can answer by two routes, and it often uses both at once.

What it learned in training. To gather training data, the companies that build models crawl the web. OpenAI, for example, does it with GPTBot: it crawls content that may be used to train its foundation models, and blocking it in robots.txt signals that your site's content should not be used for that training.⁠^{[1]OpenAI. Overview of OpenAI Crawlers} At Google, the equivalent control is Google-Extended, which serves to limit training and grounding in some of Google's other systems.⁠^{[2]Google Search Central. AI features and your website}

What it searches for in the moment. ChatGPT can search the web automatically when a question benefits from current information, and the answer may then carry citations to its sources. To search, it rewrites your question into one or more more specific queries that it sends to its search providers.⁠^{[3]OpenAI Help Center. Searching the web with ChatGPT} The sites it shows are crawled by a different agent, OAI-SearchBot: if you block it, your site does not appear in ChatGPT's search answers, although it can still show up as a navigation link.⁠^{[1]OpenAI. Overview of OpenAI Crawlers}

The two routes mix in shopping. To decide which products to show, ChatGPT takes into account structured metadata from first- and third-party providers, such as price or description, and also the answers the model itself generates before looking at new search results.⁠^{[4]OpenAI Help Center. Shopping with ChatGPT Search} Put another way: what the model already "believes" about a product goes into the answer before it searches for anything.

What LLMO adds to GEO

GEO has a precise origin: a research paper published on arXiv on November 16, 2023 and accepted at KDD 2024, which defines generative engines as systems that answer by synthesizing information from several sources and summarizing it with an LLM.⁠^{[5]arXiv (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande; KDD 2024). GEO: Generative Engine Optimization} That is, GEO was born thinking about the second route, the sources the engine retrieves and cites.

LLMO has no equivalent document. Google's guide to optimizing for generative search names AEO (Answer Engine Optimization) and GEO as terms used for the work of improving visibility in AI search, and answers that, for Google Search, it is still SEO; LLMO does not appear.⁠^{[6]Google Search Central. Optimizing your website for generative AI features on Google Search} What LLMO has of its own is the first route, training, and there a store controls little, and late: it can decide whether to let the training agents crawl, but none of the OpenAI pages cited here says how often its models are trained or how a particular page gets into them.⁠^{[1]OpenAI. Overview of OpenAI Crawlers}

In a store

The most concrete setting that separates LLMO from GEO lives in robots.txt. OpenAI gives this case as its example: allow OAI-SearchBot to appear in ChatGPT search and block GPTBot so that what it crawls is not used for training, because each setting is independent of the other.⁠^{[1]OpenAI. Overview of OpenAI Crawlers} In a fictional example store, it would look like this:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

With this file, the store can get cited when ChatGPT searches, but it asks that its content stay out of training. It is an LLMO decision in its purest form: it changes nothing about what ChatGPT finds today when it searches, and it decides whether your product pages can become part of what its models know tomorrow. For a purchase question, the first block weighs more: ChatGPT searches the web when a question benefits from current information, and today's price and stock only arrive by that route.⁠^{[3]OpenAI Help Center. Searching the web with ChatGPT}

How to check it

1. Open your robots.txt and see what you do with GPTBot and with OAI-SearchBot, separately. A change can take about 24 hours to show in ChatGPT search.⁠^{[1]OpenAI. Overview of OpenAI Crawlers}
2. Ask ChatGPT about your product in a new chat and see whether the answer brings citations or sources. Answers that use search can include citations; an answer with none probably came from what the model already knew.⁠^{[3]OpenAI Help Center. Searching the web with ChatGPT}
3. Repeat the question choosing the search tool and compare the two answers with your product page.⁠^{[3]OpenAI Help Center. Searching the web with ChatGPT}
4. Check that your product structured data is complete: price and description are part of what ChatGPT takes into account when choosing products.⁠^{[4]OpenAI Help Center. Shopping with ChatGPT Search}
5. When you get an LLMO proposal, ask which route it works on and how it will measure it. Google warns that no third-party tool has access to its internal ranking or AI systems.⁠^{[6]Google Search Central. Optimizing your website for generative AI features on Google Search}

Related concept: GEO

GEO has an origin paper and Google names it in its guide;⁠^{[6]Google Search Central. Optimizing your website for generative AI features on Google Search} LLMO is its model-centered variant. In practice they share almost everything: that your page can be crawled, indexed and cited, and that it says something of its own. GEO is measured in citations inside answers composed from retrieved sources;⁠^{[5]arXiv (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande; KDD 2024). GEO: Generative Engine Optimization} LLMO adds the question of what the model knows about you without searching, and no report from the platforms cited here answers that: you answer it by asking, as in step 2. If a proposal uses both names, what matters is whether it includes that second question and how it plans to answer it.

To keep going, the comparison GEO, AEO and SEO: what changes for an online store puts the names side by side, and the AI search reading path orders the rest of the terms.