The problem with AI in marketing is not the AI. It is the data

More exactly: that the data is out of date, or that it simply doesn't exist.

These last two years the industry has been arguing enthusiastically about the wrong thing. Which model is smarter, who has the longer context, who writes better code, whose prompt engineering is more sacred. And the real bottleneck has been standing off to one side that whole time, smirking quietly. The bottleneck is not the model. It's the data it works with when you come to it with a real marketing problem.

And until you understand that, any "AI marketing" is no more than a nicely packaged guessing game.

The model doesn't know your market. And it hides that very confidently

Let's be honest, without the magic. A large language model is not an oracle and not an analyst. It's a very well-read gadget that learned the average version of the internet at the moment of training and learned to speak convincingly. The key words here are "average" and "at the moment of training".

When you ask the smartest model "how do I go to market with my product", it physically cannot know your market. It hasn't gone into the search results today, hasn't looked at what your channel costs this week, hasn't read your competitors' recent reviews and hasn't counted how many people really search for what you make. It takes the common place from the thousands of articles it absorbed and hands you the median advice at the level of "define your audience, make useful content, launch on Product Hunt".

It's not that the advice is incorrect. It's that it is nothing. The same for a fintech startup, for a coffee shop and for a B2B SaaS. Underneath there isn't a single number, not a single source, not a single reason to do it that way and right now.

And here's the treacherous part: the model is not going to tell you "I don't know". It's trained to sound confident. So at the point where it has no data it doesn't go quiet: it makes it up. Pretty, rounded off and with a clever face on. In a world where decisions cost money, a confident invention is more dangerous than an honest "no idea".

What "marketing data" actually is

When people say "data", many of them picture a dashboard with charts. But to make a decision several radically different layers matter, and almost all of them live outside the model's head:

• Demand. How many people really search for your topic, with which formulations, whether that is growing or deflating. Not "the market seems big", but volumes.
• Competitors. Who is already here, where their traffic comes from, what they are growing on, where they have holes. That isn't guessed, it's looked at.
• Voice of the customer. What people write in reviews, on Reddit, in forums and chats: their real words, their pains and their objections, not your fantasies about them.
• Channel benchmarks. What acquisition costs, which channel is alive in your niche and which one turned into a graveyard of budgets long ago.

All of that is live data. It changes every week and it exists outside, in the market. And the model, by default, only handles what got frozen inside it during the training phase. They are two different worlds, and swapping one for the other is the root of almost every failed "AI strategy". Your marketing analytics tells you what already happened inside your own funnel; none of these four layers is in there.

A good formula to keep in your head: real data, not training data.

Going out of date is the silent killer nobody sees

The absence of data at least shows. With expiry it's worse: it dresses up as knowledge.

Every model has a knowledge cut-off date. Everything that happened after it does not exist for the model. And marketing is a field where half a year is an eternity. The channels that worked two years ago have either doubled in price, or got saturated, or collapsed. CAC has drifted. The algorithms have been rewritten. The audience has moved house. New formats, new platforms and new rules of the game have appeared.

And then you get from the model a self-assured plan which actually describes the market of two years ago. It's logical, coherent, argued - and dead. It's like using the map of a city that a digger has already driven through: everything looks drawn, only those buildings aren't there any more and where your street used to be there's now a motorway junction.

The most disgusting thing is that in a plan like that there is no way to see where it's lying. The error doesn't stick out. It's spread evenly across the whole pretty document. This is a data quality problem, and it's the one kind you cannot spot by reading the output.

A decision without data is not a decision. It's a bet

Here comes an important thesis that touches a nerve for a lot of people. If you made a decision, it worked out, but you leaned not on data but on your nose, you're not a hotshot. You simply got lucky.

And luck is a terrible business strategy. First, it has the habit of running out at the worst possible moment. Second, it doesn't scale: you cannot put luck on an assembly line or teach it to the team. And third - and this is the most treacherous part - a blind shot that hits teaches you the wrong lesson. You decide you have divine intuition, you repeat the same move under different conditions and you break your face, as was to be expected, because the first time it wasn't the "decision" that worked, it was chance.

Multiply that by speed. Through a normal R&D department there go something like ten hypotheses a week. Validating each one with numbers is slow as hell and expensive, there simply aren't the hands. And discussing a hypothesis only as an idea, at the level of "well, it sounds logical", is the same fortune-telling, only now collective, in a meeting room and on the expensive salary of several people at once. It comes out a dead end: either slow and with data, or fast and blind.

How to live with this: data first, model afterwards

The good news is that this isn't cured by giving up AI, but by changing the order of the actions. The model is not a source of truth, it's an accelerator of the work with data. A few practical principles that really change the quality of decisions:

1. No hypothesis comes to the discussion naked. Underneath there has to be a minimal layer of material: whether there's demand, who is already in the market, where the audience sits, what the channel costs. Without material it isn't a hypothesis for a meeting, for now it's a corridor idea.
2. Separate "the model thinks so" from "the market shows it". They weigh differently. The first is a starting point for thinking. The second is a basis for deciding.
3. Demand a source under every claim. If behind a number there is no where-it-comes-from, it isn't data, it's somebody's opinion dressed as data. The source makes the conclusion reproducible, and therefore checkable and arguable.
4. Where there is no data, say so. An honest "there isn't enough data here, we're going at our own risk" is worth a hundred times more than invented confidence. Throw the section away, don't fill it with fiction.
5. Use AI for the speed of gathering and synthesis, not for the verdict. Let it quickly drag the material in from every corner and comb it. Thinking and deciding is still your job - with your head, not by delegating the choice to something that doesn't answer for the result.

That is also the honest test for any AI marketing tools you're evaluating: do they bring the outside market in, or do they only rephrase what the model already believed? Most generative AI marketing products are the second kind, and they are priced like the first.

The point is not to turn marketing into a bureaucracy that justifies every sneeze. The point is that the team's discussion starts not from zero but from the material. Arguing when there is a real cut of the market on the table is a radically different conversation from arguing with somebody's "well, it seems to me". The quality of the debate goes up an order of magnitude just because everybody has the same reality in front of them and not five different fantasies.

Conclusion

AI is not going to kill marketing and it isn't going to make everybody a genius. It does something else, and it's far more interesting: it amplifies what you already have. If there's data, it amplifies the data and gives you a real acceleration. If there's no data, it amplifies your guessing game very fast and very convincingly, and then you spend a long time wondering why that beautiful strategy didn't take off.

The efficiency bar only goes up because of this. And the ones it will shake hardest are those who all these years have been moving data from the left-hand pile to the right-hand one without asking what is written in it or where it came from.

So the question is really not "which model do I choose". The question is what data you are going to make decisions on. Any AI marketing strategy that skips that question is decoration. Everything else is secondary.