LLM SEO, explained for a local business owner.

LLM SEO is the practice of improving how large language models (LLMs), the systems behind assistants such as ChatGPT, Gemini and Claude, describe and recommend a business. Most people use the term to mean the same thing as generative engine optimization (GEO), and the goal is the same. Underneath it sits a more basic question: how a model comes to know anything about your company at all.

The two ways a model knows about your business

What a model learned in training and what it retrieves when asked behave very differently.

Why retrieval matters most for a local company

Training data says little about a small local business. A national brand is written about on a great many pages, so a model absorbs a lot about it. A 12-person termite company in La Mesa appears on far fewer. Asked about it from training alone, a model may know nothing, may confuse it with a similarly named business, or may repeat a detail that has since changed. When the assistant searches the web first, it can read your current pages and listings instead, and my ChatGPT visibility work applies that to one assistant in particular.

What follows from that in practice

Access comes first. AI companies fetch pages with their own crawlers, and a site can block them in its robots.txt file or at its firewall, sometimes without the owner having chosen to. I check this alongside ordinary indexing and crawling.

Then the facts themselves. Services, cities served, hours and phone number should be in the HTML of the page, in plain text. A fetcher may not run JavaScript, and a phone number that exists only inside an image is easy for a machine to miss. Those details also have to be stated the same way across your site and listings, which is the job of entity and schema work, so a machine can treat them as one business.

When an answer cites a source for a wrong fact, I correct it at that source.

What LLM SEO can’t do

No one outside the company that built a model can reach in and change what it learned, and what it says next is never guaranteed. An offer to “train ChatGPT on your business” describes something an outside agency has no way to do. The real work is slower: I correct the sources, then ask again.

How I tell which source I’m looking at

Where the product allows it, I ask the same question both ways, with web search on and off, and compare. Answers that list sources point to the pages worth working on. Answers without them show what the model absorbed in training, which is useful to know and not something I can change this month. Both go into the AI visibility audit.