LLM Basics for Marketers

Mechanic One: It Predicts, It Does Not Look Up

A large language model breaks text into tokens, roughly word fragments, and repeatedly predicts the most plausible next one given everything before it. That is the entire operation. There is no database of facts inside it and no lookup step.

Two consequences matter to you, and they pull in opposite directions.

The first is that a model can produce a fluent, confident, completely fabricated statement about your business without anything having gone wrong internally. That is hallucination, and it is not a bug in the sense of a defect: a system predicting plausible text will sometimes predict plausible text that is false.

It happens most where information is thin, which is to say most often about smaller and less-written-about companies.

The second is more useful. Because the model predicts what plausibly comes next, phrasing that commonly precedes a factual claim makes the claim easier to reproduce. A sentence built like "According to X, Y is Z" sits in a pattern the model has seen a great many times in reliable text.

This is a large part of why the original GEO research found that adding citations and statistics improved visibility: it was making pages look like the kind of text that gets quoted.

Mechanic Two: Its Knowledge Is Frozen

Everything a model knows without looking anything up was fixed when it was trained. That date is its knowledge cutoff, and what sits behind it is parametric knowledge: patterns in the weights, not documents on a shelf.

This is why an assistant with no web access can still discuss a well-known company confidently and will describe a two-year-old company as though its last funding round were its current state. It is not out of date by accident. It is a snapshot, and snapshots do not update.

You cannot edit this layer, and neither can anyone selling you a service that claims to. It changes when a successor model is trained, on a web that will by then have been written. What you publish today is a deposit against a training run that has not happened yet.

Mechanic Three: Anything Current Must Be Retrieved

Because of the cutoff, systems that need to be current search the live web, read the results, and place that text into the model's input alongside the user's question. This is retrieval augmented generation, and it is the layer nearly all actionable GEO work targets.

It is worth being precise about what happens, because the details are where the leverage is:

1. The system rewrites the user's question into one or more search queries of its own. These are frequently not the user's words.
2. It retrieves candidate documents.
3. It selects a subset that fits its budget, discarding the rest.
4. It reads the survivors and composes an answer, citing some of what it used.

Every one of those steps is a filter you can fail independently. You can rank for the user's phrasing and lose at step one because the system searched for something else. You can be retrieved at step two and dropped at step three. You can survive to step four and be read but not cited.

Mechanic Four: The Context Window Is Finite

Everything the model considers at once, the question, the retrieved documents, the conversation so far, must fit inside a fixed budget of tokens called the context window. Modern windows are large. They are not unlimited, and retrieval routinely finds far more material than will fit.

So something is thrown away, before your page is ever weighed on merit. This is the least discussed step in GEO and one of the most decisive.

What survives selection tends to share properties, and they are properties you control:

• The relevant part is dense rather than diffuse. A page where the answer occupies one clear section beats a page where the same answer is spread across two thousand words.
• It reads correctly in isolation. A section that depends on the paragraph above it loses its meaning when extracted, and an extracted fragment that means nothing is a fragment that gets dropped.
• The structure is honest. Headings that describe what is actually underneath them help a system locate the relevant chunk. Clever headings actively hurt.
• It is not padded. Introductions restating the question, and conclusions restating the introduction, consume budget without carrying information.

There is a broader point here worth stating plainly. Long-form content earned its reputation in an era where length correlated with depth and depth correlated with ranking. Under retrieval, length is a cost paid at the selection step. Depth still wins. Padding now loses in a way it previously did not.

Why The Same Question Gets Different Answers

Ask an assistant the same thing twice and you will often get two different answers. This confuses a lot of measurement, so it is worth knowing where the variance comes from.

Part of it is sampling: the model chooses among plausible next tokens rather than always taking the most likely one, which is what makes its writing readable rather than mechanical. Part of it is retrieval: the searches it generates vary, so the documents in front of it vary. Part of it is context: memory of earlier conversations changes the input.

The measurement consequence is not optional. A single prompt result is an anecdote. Run each prompt several times, in fresh sessions, and record how often you appear rather than whether you appeared. Frequency across runs is a real signal. One reading is not.

See The Mechanics Directly

Twenty minutes, and it makes the four mechanics concrete in a way reading cannot:

1. Ask an assistant about your company with web search explicitly turned off. What comes back is parametric knowledge. Note what it gets wrong, and whether it invents rather than declines.
2. Ask the same question with search enabled. The difference between the two answers is exactly what retrieval is contributing for you today.
3. Ask it what search queries it would run to answer your most important buyer question. Compare those to the terms you actually optimise for. The gap is usually instructive.
4. Paste one of your key pages in and ask which section it would use to answer a specific question. If it cannot point to one, your answer is too diffuse to survive selection.
5. Run your most important prompt five times in five fresh sessions. Count appearances out of five. That fraction, not any single answer, is your actual position.

Key Takeaways

• Models predict text rather than retrieving facts, so thin information about you produces confident invention, not silence.
• Parametric knowledge is frozen at a cutoff date and cannot be edited. Publishing today is a deposit against a future training run.
• Retrieval is a four-step filter: query rewriting, retrieval, selection, composition. You can fail each one independently.
• The context window forces material to be discarded before merit is assessed. Dense, self-contained, honestly structured sections survive that cut.
• Answers vary between runs by design. Measure frequency across repeated fresh sessions, never a single result.