A language model can know about your business in exactly two ways: it absorbed something during training, or it retrieved something at the moment the question was asked. Almost every practical decision in this area follows from telling those apart, because only the second is open to influence.
Most confusion about “LLM SEO” comes from treating them as one thing.
Training: fixed, and not your lever
Models are trained on large collections of text gathered before the model ships. Whatever was absorbed about your industry, your competitors and possibly your business is baked in at that point and does not update because you published something afterwards.
Two consequences follow, and both are freeing. You cannot submit your site to a training set, so nobody can sell you that. And an older model describing your business inaccurately is not a problem you can fix directly, only one that later versions and live retrieval can correct around.
There is a slower, indirect version of influence here: what the web says about you over years is what future training sets contain. That is a reason to care about consistent, accurate descriptions of your business everywhere, but it is not a campaign with a completion date.
Retrieval: where the work actually pays
Modern assistants do not answer purely from memory. When a question needs current or specific information, they search, fetch pages, and generate an answer grounded in what came back. This is the same shape as how AI Overviews choose their sources, and it is the part that responds to work.
Retrieval behaves much like search, which is the useful insight. The assistant needs to find your page, fetch it, parse it, and judge it worth using. Every one of those steps is familiar: discoverability, crawlability, clarity, credibility. A business absent from assistant answers usually has a retrieval problem rather than a training problem, and retrieval problems are fixable.
What actually changes the outcome
Be reachable by the right crawlers. Retrieval depends on fetchers that are separate from the classic search crawler. Google documents its range of crawlers and fetchers, and OpenAI documents its own user agents, including the distinction between the crawler that gathers training data and the one that fetches pages to answer a live question. If your robots.txt blocks the second, you have opted out of being cited.
Publish text a machine can read without effort. Content that appears only after heavy client-side rendering, or lives inside an image or a PDF, is harder to use. Plain HTML text remains the most reliable format.
Write self-contained passages. A generated answer stitches together fragments. A section that states a fact plainly, with its context intact, is easy to lift. A section that depends on three paragraphs above it is not.
Be described consistently. Since models synthesise across many mentions, inconsistent descriptions of what you do produce a blurry characterisation. This is the one place where the training and retrieval paths converge: consistency helps both.
The citation-without-a-link problem
Assistants routinely describe a business without linking to it. That is not a bug to be fixed; it is a property of synthesis. It has one important implication for measurement: your analytics will never show it. There is no referral, no session, no attribution.
Which means the honest way to know how assistants portray you is to ask them. Build a list of the questions a prospective customer would type, run them across the assistants your audience uses, and record what is said about you, whether you are cited, and who appears instead. That record is the only reliable measurement available, and the competitor list it produces is worth the exercise on its own.
What this means practically
If you already run a competent search programme, LLM SEO adds three things: check that AI fetchers are not blocked, structure passages so they can be lifted whole, and run a monthly prompt audit instead of expecting a dashboard.
If you do not have a search programme, there is no shortcut here. Retrieval draws on the open web the same way search does, and a site that is hard to find, slow to load or thin on substance fails at the same step in both systems. The wider view of that work is in our generative engine optimization guide, and if you would rather it were handled, it is our job.
