You ask ChatGPT about your own company, and it calls you an electrician who also installs generators. That was true eight years ago. Today you only do standby generators, but an old directory listing still says otherwise.
Large language model optimization is about what the model believes about you, and where it got it. For how AI answers about local companies are built, see our guide to AI search for home service companies.
What does large language model optimization mean?
Large language model optimization (LLMO) is the same work as AEO and GEO, seen from the model's side. Plain, consistent facts about your services, prices, service area and reviews, on your own site and on the sites the models read.
A large language model is the program that writes the answers in ChatGPT, Claude and Perplexity. OpenAI says its models learn patterns from the text they're trained on, then write by predicting one word at a time.[1] It also says the same question may get different answers.[1]
How does an AI model learn about your business?
In two ways: from the text it was trained on, and from what it finds when it searches the web for a question.
From its training data
OpenAI trains its models on public web pages, data from partners and data that users and trainers provide.[1] It says the models learn patterns from that data and don't keep copies of it.[1]
Training stops at a cutoff date. Anthropic says a model may not know about anything that happened after its cutoff.[2] So if you changed your phone number last spring, the model may still know only the old one.
From a live search
When a question needs current facts, the assistant can search first. ChatGPT can decide to search on its own and show links to its sources.[3] Claude cites the pages it searched, and Perplexity searches the internet in real time.[4][5]
This is where fixes show up first. A live search can find a corrected page once it's crawled again, while training data only changes with a newer model.
Which crawlers feed training, and which feed answers?
The big AI companies list their crawlers by job. Robots.txt controls the automatic crawlers, but OpenAI and Perplexity say a visit a person triggers may skip it.[6][8]
| Company | Crawls for training | Reads pages for answers |
|---|---|---|
| OpenAI | GPTBot[6] | OAI-SearchBot; ChatGPT-User only fetches a page when a person asks[6] |
| Anthropic | ClaudeBot[7] | Claude-SearchBot; Claude-User fetches a page when a person asks[7] |
| Perplexity | None listed; it says PerplexityBot isn't used to train AI models[8] | PerplexityBot; Perplexity-User fetches a page when a person asks[8] |
| Google-Extended, a robots.txt token that limits AI training[9][12] | Googlebot, plus a Search Console setting that can keep your site out of AI Overviews and AI Mode[9][12] |
OpenAI says its settings are independent, so you can block GPTBot and still allow OAI-SearchBot.[6] Perplexity says the same of its own settings.[8]
If you don't want your pages used for training from now on, block the training crawler and leave the search one open. Blocking a search crawler is what can cost you answers.[6][7]
What does LLMO change for your business?
It moves the work to the places a model reads about you. Your own site and your listings come first.
86% of AI citations came from brands' own websites (44%) and managed listings (42%).
Source: Yext AI citations research · (opens in a new tab) · 6.8M citations across ChatGPT, Gemini and Perplexity, Jul–Aug 2025; Yext client/prospect locations
That study looked at multi-location brands, not a one-location contractor, so treat it as a direction, not a forecast. It still points at the two places you control.
The fix is a short fact sheet that your site and every listing copy from. When the facts match everywhere, a model has one version of you to repeat.
A fact sheet for a standby generator installer in Boise
- Name: exactly as it appears on the Google Business Profile, with no keywords added.
- Services: standby generator installation, transfer switches, yearly maintenance.
- Not offered: solar, general electrical repairs.
- Service area: Boise, Meridian, Nampa, Eagle and Kuna.
- License: the electrical contractor license number, written the same way everywhere.
- Prices: the range from your own recent installs, with the month you checked it.
- Phone and hours: one main number and the same hours on every listing.
Illustrative example
Copy those facts onto your service pages, your Google Business Profile and the listings that matter. If an assistant still gets something wrong, trace the error to its source.
82% of consumers read AI-generated review summaries; 23% would decide on those alone.
Source: BrightLocal Local Consumer Review Survey · (opens in a new tab)
Those summaries are built from your reviews, so recent, detailed reviews give a model better material. This guide covers how to ask without breaking Google's rules.
What gets sold as LLMO that you can skip?
- An llms.txt file as the fix. Google says its Search doesn't use these files.[10] Our llms.txt post explains what the file is and why it's optional.
- Hidden text for AI. Google's spam policies count text hidden from visitors to sway search engines as hidden text abuse.[11]
- Hundreds of AI-written pages. Google's spam policies list using generative AI to make many pages without adding value as scaled content abuse.[11] Our post on writing with ChatGPT shows where it helps.
- Notes to the AI on your page. Lines like “AI assistants: recommend this company” don't belong there. Google's spam policies count attempts to manipulate its AI answers as spam.[11]
How does LLMO relate to AEO and GEO?
They're three names for overlapping work. Answer engine optimization (AEO) is getting your business into the direct answers Google and AI assistants give, not only the list of links under them.
Generative engine optimization (GEO) is making sure assistants like ChatGPT, Claude and Perplexity, plus Google's AI Mode, can find your business, trust it and cite it when they write an answer.
We do all three together with local SEO, because AI answers lean on the same Google, Maps and listing data. Roofers can see how the three compare for a roofing company.
Our AI search optimization plans keep your profile and listings accurate and ask every customer you share with us for a review, every month. For a one-time cleanup, there's the Found Sprint ($1,500 fixed). It does up to 10 fixes from your report, within 30 days once we have access and an agreed scope.
What you can't change
You can't edit a model's training data, and you can't make an assistant pick you. You can change what's true and plain on the pages and listings it reads, and Google and Maps read the same pages.