Where GPT Actually Earns Its Keep in Marketing

I co-founded Moonb, a creative studio, and I have watched GPT go from a novelty someone opened in a side tab to a tool that touches almost every task my team runs in a week. That happened fast. What did not happen fast is the part everyone assumed would follow. The results getting obviously, measurably better.

So I want to skip the article you have already read a hundred times. No list of ten magic prompts. Instead, the thing I actually care about as someone who ships marketing work for a living. Where does an LLM earn its place, and where does it cost you more than it saves? After two years of using it daily, my answer is narrower and more useful than the hype suggests.

The real state of GPT in marketing right now

Usage is basically universal. In HubSpot’s 2025 research, 91% of marketing leaders say their teams use AI to assist their work, and 65% plan to spend more on it this year. If you are still framing GPT as an edge, you are a year late. It is table stakes now. Your competitor down the street is prompting the same model with the same generic instructions and getting the same generic output.

Here is the uncomfortable part. Adoption and results are not the same curve. McKinsey’s 2025 State of AI found that marketing and sales saw the biggest generative-AI adoption jump of any function, more than doubling since 2023. In the same study, over 80% of organizations reported no material impact on enterprise-level profit from generative AI. Everyone is using it. Almost no one is winning with it at scale.

That gap is the whole story. It tells me the problem is not the model. The problem is where teams are pointing it.

What an LLM is good at, and what it only pretends to be good at

An LLM is a pattern machine. It has read a staggering amount of text and it predicts what plausibly comes next. That makes it excellent at anything where “plausible and fast” beats “original and considered.”

It is very strong at:

• Turning a messy pile of notes into a structured first draft.
• Rewriting one thing into five formats (a blog post into a thread, an email, a set of captions).
• Summarizing a long document so you can decide whether to read it.
• Generating twenty subject-line variants when you are stuck on one.

It only pretends to be good at a second category, and this is where teams get burned. It will confidently produce a brand strategy, a market-defining tagline, or a “trend” it half-remembers. The output looks finished. It reads with total authority. And it is often wrong, or generic, or subtly off-brand in a way you will not catch until it is public.

Here is the distinction I hold in my head. The model is good at execution inside a frame you set. It is weak at setting the frame, and it is weak at the last five percent that makes work feel like it came from a person.

The middle of the funnel is where it earns its keep

Think of any piece of marketing work as three phases. There is the strategic decision at the start (who is this for, and what are we actually saying that no one else is). There is the tedious middle (research, structure, drafts, variants). And there is the finish (the voice, the taste, the final call on whether this is good).

The middle is the part worth handing over. It is also, conveniently, where marketers already report the most value. HubSpot’s data shows research is the single most common use of generative AI at 48%, with marketers saving one to two hours a day. This is not the AI writing your campaign. It is one to two hours a day clawed back from the grind. That is the realistic prize, and it is a good one.

Concretely, the middle-of-the-funnel jobs I trust the model with:

• Research and synthesis. Pull the main arguments out of ten sources so I can form my own view faster.
• First drafts. Get words on the page so I am editing instead of staring at a blank document.
• Repurposing. One long asset becomes the raw material for a week of smaller ones.
• Variants at volume. Ten headlines, five hooks, three cold-open lines, so I have something to react against.

Notice the pattern. In every case a human sets the direction before and judges the result after. The model does the middle mile.

How to prompt so the output does not read like everyone else’s

Most bad GPT output comes from lazy input. “Write me a LinkedIn post about our new feature” gives you the average of every LinkedIn post ever written, which is exactly the beige nobody stops scrolling for. A usable prompt has four parts, and it helps to run them as a checklist before you hit enter.

Part	What to give it
Persona	Who is writing and who is reading. "A skeptical CFO writing to other skeptical CFOs."
Context	The real facts. Paste the actual feature notes, the actual objection, the actual audience.
Format	Length, structure, medium. "Six short lines, no hashtags, one question at the end."
Tone	Show it, do not name it. Paste two examples of your voice and say "match this."

The tone row is the one people skip, and it is the one that matters most. Telling the model “be conversational” produces the model’s idea of conversational, which is everyone’s idea of conversational. Pasting three paragraphs of your own writing and saying “write in this voice” produces something closer to yours. Show, do not tell.

A few clean templates you can adapt. Fill the brackets with your own specifics.

You are [role] writing for [audience]. Here are the facts: [paste
real notes]. Write [format]. Avoid marketing cliches, no
buzzwords, no rule-of-three lists. Match the voice in this
sample: [paste your own writing].

Here is a long-form piece: [paste article]. Repurpose it into
[channel] for [audience]. Keep the strongest single idea, cut
everything else, and end on a concrete takeaway, not a summary.

Give me 15 variants of [headline / hook / subject line] for
[audience]. Make five plain, five specific with a number, five
that open a loop. No exclamation points.

The goal is not a perfect first output. It is raw material specific enough that your editing makes it yours.

The two things you should never let it own

There are two ends of the work I keep entirely human, and I am strict about it with my team.

The first is the strategic call at the start. What we are saying, and why it should matter to the specific person reading it right now. That decision carries the brand. A model that averages everything on the internet is structurally the wrong tool for a choice that is supposed to be distinctive. If GPT decides your positioning, your positioning becomes the median of your category.

The second is the final voice. The last edit, the line that has a point of view, the joke that lands because a specific person with taste wrote it. This is the five percent that makes an audience feel something, and it is the first thing to vanish when you let the model finish the piece. I have written before about why AI is not about to replace the people who actually make the work, and this is the core of it. The craft lives at the two ends.

Here is the split I give people, laid out plainly.

Hand to GPT	Keep human
Research and source synthesis	The strategic angle and positioning
Structuring a messy first draft	The final voice and the last edit
Repurposing one asset into many	Any claim about your product or numbers
Generating variants to react against	The decision on whether the work is good

Fact-checking is not optional

If you take one operational habit from this piece, take this one. Treat every factual claim the model produces as unverified until you check it against a real source.

The reason is that these models are wrong in the most dangerous possible way. Confidently and fluently. Hallucination rates swing wildly by task. On grounded summarization the best models sit around one percent, which is fine. On harder reasoning, error rates climb past a third. Stanford researchers found that on specific legal questions, leading models hallucinated between 69% and 88% of the time. Read that again. The model does not know it is wrong, and nothing in the tone will warn you.

For marketing that means a hard rule. Any statistic, any date, any comparison, any claim about your own product goes through a human check before it ships. GPT can draft the sentence. It cannot be the source. I have watched a clean-looking paragraph carry an invented number straight toward a client deck, and the only thing that caught it was a person who stopped and asked “where did this figure come from?”

The hidden tax when your audience can tell

There is a cost that does not show up in your workflow at all. It shows up in how people receive the work.

NielsenIQ studied consumer reactions to advertising and found that labeling an ad as AI-generated made people see it as less natural and less useful, which lowered their willingness to research or buy. A separate study from the Nuremberg Institute for Market Decisions isolated the effect even further. AI disclosure lowered attitudes toward an ad even when the content was identical. Same words, same image, worse response, purely because people knew a machine made it.

You will not always disclose, obviously. But the finding underneath is what matters. When work reads as machine-made, trust drops. And audiences are getting better at spotting it every month. The generic cadence, the tone that could belong to any brand in the category, these are becoming tells. If your content is indistinguishable from what your competitor’s model produced, you have not saved money. You have paid a tax you cannot see, in attention you did not earn.

This is exactly why I keep the human at the finish. The human finish is what stops the work from reading as machine-made in the first place, and that is worth more than the minutes it costs.

A workflow that keeps the machine in the middle

Put it all together and you get a shape that is simple to run.

1. A human sets the frame. The strategy, the audience, the single thing this piece has to say. No model input here.
2. GPT does the middle. Research, structure, a first draft, variants, repurposing. Prompted properly, with your context and your voice sample pasted in.
3. A human fact-checks. Every claim, number, and product statement verified against a real source before anything moves forward.
4. A human finishes. The last edit, the voice, the line with a point of view. This is the part that makes it yours, so a person does it.

Steps two and three are where the hours come back. Steps one and four are where the brand lives. Lose that order and you become one of the 80% getting no real value, running a faster machine that produces more of the same average.

If you are building this muscle inside a marketing team, the same logic applies to how you structure an in-house creative team. Tools in the middle, judgment at the edges. At Moonb, that is close to how we actually work. Real creative directors own the frame and the finish, with AI carrying the tedious middle so the people can spend their time on the parts that only people can do. If that division of labor is what you are trying to figure out, here is how we run it day to day.