The real problem with AI in marketing isn't the AI. It's the data.

More precisely — that the data is stale, or barely there at all.

For the last couple of years the industry has argued about the wrong thing. Which model is smarter, whose context window is longer, who writes better code, whose prompt engineering is more sacred. Meanwhile the real bottleneck stood off to the side the whole time, quietly smirking. The bottleneck isn't the model. It's the data the model runs on the moment you bring it a real marketing problem.

And until you get that, any "AI marketing" stays a nicely packaged guessing game.

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

Let's be honest, no magic. A large language model isn't an oracle or an analyst. It's a very well-read thing that learned an averaged version of the internet as of its training date and got good at sounding convincing. The key words here are "averaged" and "as of training."

When you ask the smartest model "how do I go to market with my product," it physically cannot know your market. It didn't crawl today's search results, didn't check what your channel costs this week, didn't read your competitors' fresh reviews, didn't count how many people actually search for what you do. It pulls the common denominator from the thousands of articles it absorbed and hands you median advice: "define your audience, make useful content, launch on Product Hunt."

The advice isn't exactly wrong. It's just nothing. The same line for a fintech startup, a coffee shop, and a B2B SaaS. Underneath it there's not a single number, not a single source, not a single reason to do this and to do it now.

And here's the sneaky part: the model won't tell you "I don't know." It's trained to sound confident. So where it has no data, it doesn't go quiet — it makes things up. Beautifully, coherently, with a straight face. In a world where decisions cost money, a confident invention is more dangerous than an honest "no idea."

What "data" in marketing actually means

When people say "data," they often picture a dashboard with charts. But for a decision you need a few fundamentally different layers, and almost all of them live outside the model's head:

• Demand. How many people actually search for your topic, in what words, growing or fading. Not "the market feels big" — volumes.
• Competitors. Who's already here, where their traffic comes from, what they're growing on, where their gaps are. You don't guess this, you look at it.
• Voice of customer. What people write in reviews, on Reddit, in forums and chats — their real words, pains and objections, not your fantasies about them.
• Channel benchmarks. What acquisition costs, which channel in your niche is alive and which one became a budget graveyard long ago.

All of this is live data. It changes every week and lives outside, in the market. The model by default operates only on what was frozen into it at training. Two different worlds — and swapping one for the other is the root of almost every failed "AI strategy."

A useful formula to keep in mind: real data, not training data.

Staleness — the quiet killer no one notices

Missing data is at least visible. Staleness is worse — it masquerades as knowledge.

Every model has a knowledge cutoff. Everything after it doesn't exist for the model. And marketing is a field where six months is an eternity. Channels that worked two years ago either doubled in price, got saturated, or collapsed. CAC drifted. Algorithms got rewritten. The audience moved. New formats, new platforms, new rules of the game.

So you get a confident plan from the model that actually describes the market of the year before last. Logical, coherent, argued — and dead. It's like using a map of a city a bulldozer drove through long ago: everything's drawn, only those buildings are gone and there's an interchange where your street used to be.

The nastiest part: you can't tell where such a plan is lying. The error doesn't stick out. It's smeared evenly across the whole pretty document.

A decision without data isn't a decision. It's a bet.

Here's a point that stings some people. If you made a call, it worked, but you leaned on a gut feeling rather than data — you're not smart. You just got lucky.

And luck is a lousy business strategy. First, it has a habit of running out at the worst possible moment. Second, it doesn't scale: you can't put luck on a conveyor and teach it to the team. Third — the sneakiest — a lucky blind shot teaches you the wrong lesson. You decide you have god-tier intuition, repeat the same move under different conditions, and predictably break your face, because the first time it wasn't the "decision" that worked, it was chance.

Multiply that by speed. A normal R&D team runs around ten hypotheses a week. Validating each with numbers is brutally slow and expensive — there simply aren't enough hands. And discussing a hypothesis as just an idea, at the "well, sounds logical" level, is the same fortune-telling, only now collective, in a meeting room, on the expensive salaries of several people at once. Dead end: either slow and data-backed, or fast and blind.

How to live with this: data first, model second

The good news is that the cure isn't dropping AI, it's changing the order of operations. The model isn't a source of truth, it's an accelerator for working with data. A few practical principles that genuinely change decision quality:

1. No hypothesis shows up to the discussion naked. Under it there must be a minimal layer of facts: is there demand, who's already in the market, where the audience sits, what the channel costs. No facts — it's not a hypothesis for a meeting, it's still a thought in the smoking room.
2. Separate "the model thinks" from "the market shows." Different weights. The first is a starting point for thinking. The second is grounds for a decision.
3. Demand a source under every claim. If there's no "where from" behind a number, it's not data, it's someone's opinion in a data costume. A source makes the conclusion reproducible — which means it can be checked and challenged.
4. Where there's no data — say so. An honest "not enough data here, we go at our own risk" is a hundred times more valuable than invented confidence. Cut the section, don't fill it with fiction.
5. Use AI for the speed of gathering and synthesis, not for the verdict. Let it quickly pull and tidy the facts from every corner. You still think and decide — with your own head, not by delegating the choice to something that doesn't answer for the result.

The point isn't to turn marketing into bureaucracy with a justification for every sneeze. The point is that the team's argument starts not from zero but from facts. Arguing when a real slice of the market is on the table is a fundamentally different conversation than arguing with someone's "well, it feels that way to me." The quality of the discussion jumps an order of magnitude just because everyone is looking at the same reality instead of five different fantasies.

Bottom line

AI won't kill marketing or turn everyone into geniuses. It does something else, and it's far more interesting: it amplifies what you already have. Got data — it amplifies the data and gives you real acceleration. No data — it amplifies your guessing game, very fast and very convincingly, and then you'll spend a long time wondering why the pretty strategy didn't fly.

The efficiency bar only goes up. And the ones who'll get shaken hardest are those who spent all these years moving data from the left pile to the right one without ever asking what's written in it and where it came from.

So the real question isn't "which model to pick." The question is: on what data are you going to make your decisions. Everything else is secondary.