Skip to content

Module 4 min

LLMs for Marketers

Context windows and retrieval, not training data, decide what an engine says about you today.

LLMs for Mar…Trained vs retrieved knowledgeWhy context windows cap what gets usedWhy engines favour fresh retrieval

Why it mattersIf you think the model "already knows," you optimize the wrong thing.

01

Definition & Foundation

What it is, in plain words

A large language model is not a database of facts; it is a text-prediction system with two very different sources of knowledge. What it absorbed during training (parametric memory) is frozen, fuzzy, and months out of date. What it fetches at answer time (retrieval) is current, specific, and citable. Almost everything an answer engine says about you today comes from the second source.

This distinction is the single most useful thing a marketer can know about LLMs, because it tells you where the leverage is. You cannot meaningfully edit a model's training data; it is a snapshot of the web from months ago, compressed and unattributed. You can control what its retrieval step finds this week. That is why GEO is mostly a retrieval game, and why the context window (the limited working memory an engine fills with fetched passages) is the real estate you are actually competing for.

02

Five Ideas That Explain the Machine

Just enough internals to make good decisions

trained-in

Parametric memory

What the model absorbed during training: a compressed impression of the web, not a lookup table. It is why a model "sort of knows" your brand but gets details wrong. Slow to change, impossible to cite.

the freshness wall

Training cutoff

The date the training snapshot ends. Anything after it (your rebrand, your new pricing, last month's launch) simply does not exist in parametric memory. Retrieval is how engines see past the wall, and why they lean on it.

working memory

Context window

The limited space a model can "look at" while answering: the retrieved passages, the conversation, the instructions. Only a handful of passages make it in per answer. GEO is a competition for these few slots.

your lever

Live retrieval

The engine searches an index, pulls the best-matching passages, and grounds the answer in them. This is updatable this week, and it carries your citation. See The RAG Pipeline for the full mechanics.

the currency

Tokens

Models read and write in tokens (word fragments), and every window slot costs tokens. Practical consequence: concise, dense, self-contained passages fit where rambling ones get truncated.

03

Myths vs Reality

Common misreadings, corrected

Myth"The model already knows our brand; we're covered."
RealityIt has a stale, compressed impression from its training snapshot, with no source attached. Ask an engine your pricing with retrieval disabled and you'll often get a confident, outdated guess. Current, correct answers come from what the engine can retrieve about you today.
Myth"The goal is to get into the training data."
RealityYou can't meaningfully target a training snapshot, and parametric knowledge never cites you. Target retrieval instead: it is measurable, changeable within weeks, and produces attributed citations. (Long-horizon parametric familiarity is real, but it is a by-product of broad, consistent mentions; see Brand Mentions.)
Myth"Bigger models will memorize everything and retrieval will go away."
RealityThe trend runs the other way: engines keep adding retrieval (ChatGPT Search, AI Mode, grounded Gemini) precisely because fresh facts, niche detail, and attribution demand it. Every major engine ships citations; that is a retrieval feature, not a memory feature.
04

Putting It to Work

Aim at the window, not the weights

You are not training a model; you are competing for a few retrieval slots in a context window. Three moves follow from that.

The retrieval-first playbook

1

Test what each engine says about you today

Ask ChatGPT, Perplexity, Gemini, and Copilot your five money questions ("what is X?", "what does X cost?", "X vs Y"). Note what is stale (parametric memory) versus what is cited (retrieval). This 20-minute audit shows exactly where the model is answering from.

2

Publish for the window, not the archive

Write dense, self-contained passages that answer one question each; they fit retrieval slots. A 3,000-word essay whose answer is smeared across ten paragraphs loses to a 150-word block that answers cleanly. See Passage Optimization.

3

Keep the retrievable facts current

The engine sees past its training cutoff only through retrieval, so your canonical pages (pricing, features, comparisons) must be the freshest, clearest source about you. Update them the week things change, not the quarter after.

05

Verification Checks

How to know it's really done

0/3 verified

You understand the machine when your effort points at retrieval. Click a check to mark it verified:

06

Works Together With

The nodes this one leans on

Tools for this from How AI Engines Work

Go deeper

Always current

These links resolve live: what you get is generated or filtered the moment you click.