In our last article we described AI development as an iceberg, with the prompt sitting visibly above the surface and the work that actually determines quality sitting below it. We closed on a point worth starting with here. The companies building foundation models are experts in general intelligence, but general intelligence is not where good marketing comes from. That comes from people who understand a brand, its market and its audience in depth: often the client's own team, and often the marketers who have worked alongside them for years and hold a great deal of hard-won knowledge about what works. A model has never had that knowledge. This article is about capturing it, and what changes when you put it into a form an AI can use. We made the fuller case for working below the surface in the previous article, The Iceberg of Truth in AI Projects.
What the model cannot know
A frontier model arrives knowing a great deal about the world in general and nothing about you in particular. It has read more marketing theory than any person alive and has never seen your brand guidelines, your last campaign, your buyer personas, or the product positioning your team spent months getting right. It is, in that narrow sense, the most capable colleague you have ever hired on their first morning, before anyone has told them anything.
General intelligence can tell you what a good marketing email looks like. It cannot write the right email for your audience, in your voice, making the claims you are allowed to make. The second thing is the only thing marketing actually needs, and it depends almost entirely on context.
This is why we think context deserves the attention usually reserved for model and tool choice. Teams spend a great deal of energy deciding which model to adopt and which tools to buy, and far less deciding what those tools are allowed to know about them. In our testing, the gap between one model and another was modest next to the gap between the same model with and without well-organised context. Model and tool choice set the ceiling. Context decides how much of that ceiling you reach.
What unstructured context costs
In the early baseline testing for our research project, before we had organised any of this, the pattern was consistent and easy to recognise. The outputs were competent and generic in equal measure.
Image and video generation drifted off-brand the moment there was no clear brand context to hold it in place. Marketing copy read as plausible filler, because the model had no grasp of what made a particular product or service valuable. Email was the clearest case of all: without the persona, the campaign the message belonged to, and a real understanding of the target audience, the adjustments the model made were weak and interchangeable, the kind of writing that could belong to anyone.
None of this was a model failure. Each output was exactly as good as the context behind it, which is to say thin. Structured marketing context is simply the raw material that fixes this: brand guidelines, product and service value, buyer personas, campaign intent, audience understanding, prior approved content. Most organisations own all of it already. What they rarely have is that material in one place, in a form a model can draw on at the moment it produces something.
The craft is in the structure
The instructive part of our research was not that adding context helped. We expected that. It was how much further the results moved once the context was organised specifically around the way marketing works, rather than simply gathered together. Quantity of context was not the lever. The shape of it was.
This is the part we would call a craft. More context is not automatically better, and unfocused or redundant context can make outputs worse rather than merely more expensive. The skill is in deciding what a given piece of work genuinely needs to know and presenting it clearly, which is the same judgement a senior marketer uses when briefing a colleague. A good brief is not everything they know. It is the right things, in the right order.
It also means context cannot be a technical artefact that only engineers can touch. The people who know what on-brand means, who own the campaign and oversee the content, are marketers. So the context they curate has to be legible and maintainable by them, not buried in machinery that assumes a background in programming or retrieval systems. Done well, this does not remove their expertise from the process. It captures it, and lets it shape far more output than any individual could review by hand.
A shared brain, built by your experts
Follow that through and you arrive at a single source of context that the whole team, and every tool, draws from: the context architecture we described in the last article, made real. We have come to think of it as a shared organisational brain. Not a document library, which is where most of this knowledge sits today, scattered and quietly going out of date, but one maintained source of truth.
The reason to centralise is reliability. When every person and tool works from its own private copy of the brand, results drift and depend on who happens to be driving. When they work from the same source, output becomes consistent and repeatable, an improvement made once lifts everyone, and there is a clear place where ownership and sign-off live. Common context is just common understanding, written down.
This is also where the more ambitious value will come from. A model that can see across the whole of your marketing knowledge, rather than one corner of it, is positioned to do more than execute the task in front of it. The problems most worth catching tend to sit between silos, where no single person holds all the pieces. We are early in exploring this, but the direction is clear, and it only becomes possible once the context is unified in the first place.
Where this goes next
None of this is exotic. It is mostly the unglamorous work of taking knowledge your organisation already has and putting it somewhere a model can reliably use. The imbalance worth correcting is one of attention: enormous care spent choosing a model, almost none spent structuring what it knows about you. For reliable marketing output, that ratio is backwards.
In the next articles we look at how this works in practice: how a campaign brief becomes context an AI can act on, and why workflows, rather than prompts, are the unit enterprise teams should be building around.
For now, one observation. If your AI output feels generic, the model is rarely the problem. It simply does not know enough about you yet.
.jpg?width=800&height=500&name=Luxid-2026-02-WhatToWatch-Report-heroimage%20(5).jpg)