Marketing was the first job AI got useful for, and it is still where most teams see the fastest return. It is also where the output is most obviously machine-written if you are careless, which costs you more than the time you saved.
Here is what works, what to avoid, and the prompts that actually produce something usable.
What it is genuinely good at
Volume variations. Ten subject lines, five ad headlines, three versions of the same paragraph at different lengths. Generating options is the single highest-value marketing use, because picking is easier than writing.
Repurposing. Turning one long piece into a newsletter, five social posts and a summary. The source material already exists, so the model is reformatting rather than inventing.
Editing your own drafts. Shorter, clearer, different tone. Reliably good, and low-risk.
Briefs and outlines. Structure before you write.
Audience reframing. "Rewrite this for a technical buyer" / "for a non-technical owner." Genuinely useful and hard to do quickly by hand.
What it is bad at
Anything factual about your product. It does not know your pricing, your features or your policies, and it will invent plausible versions of all three.
Claims and statistics. It will produce a confident number with a confident-sounding source, and both may be fabricated. Every figure needs checking.
Your brand voice, unprompted. Out of the box it writes in the same mid-Atlantic marketing register as everyone else's, which is precisely the problem.
Anything legally sensitive. Comparative claims, health or financial statements, guarantees.
Prompts that actually work
Instead of "write a product description for my software" Use: paste your actual feature list, then "Write a 60-word product description for a small-business owner who is not technical. Plain language, no superlatives, no exclamation marks. Lead with the problem it solves."
The pattern that matters: give it your material, name the audience, set the length, and ban the things you do not want. Those four elements do more than any amount of role-play preamble.
For variations: "Give me 10 subject lines. Vary the angle, not just the wording — some curiosity, some direct benefit, some urgency. Under 50 characters."
For brand voice: paste three things you have written that you liked, then "Match this voice. Notice the sentence length and how little jargon there is." Examples beat adjectives — "professional but friendly" means nothing to a model.
For editing: "Cut this by a third without losing any of the three data points. Keep the opening line."
The tell, and how to remove it
AI marketing copy has a signature: tricolon sentences ("fast, simple, and powerful"), em-dash asides everywhere, "In today's landscape", "it's not just X, it's Y", and a closing paragraph that restates the opening.
The fastest fix is to ban them explicitly in your prompt, then read the result aloud. Anything you would not say out loud, cut.
The second fastest fix is to write the first and last sentence yourself. Those carry most of the voice, and the middle matters less than you think.
Which model
Claude for anything a customer reads — its default prose needs the least de-robotising. Gemini when the piece depends on current facts. GPT when you want consistent, predictable structure across many items.
For campaign work, running the same brief through two models and taking the better half of each is genuinely faster than iterating with one.
A workflow that holds up
- You write the brief: audience, one message, what to avoid.
- Model produces 5–10 options.
- You pick the direction — not the words, the direction.
- Model develops it.
- You rewrite the opening and closing, check every fact and figure.
The human steps are first, third and last. That is not a limitation; it is what keeps the output sounding like your company.
Common questions
Can ChatGPT write marketing copy? Yes, well — especially variations and repurposing. It needs your material and your editing to sound like you rather than like everyone else.
Will Google penalise AI-written marketing content? Google penalises unhelpful content, not AI content specifically. Thin, mass-produced pages are the risk; genuinely useful pages are fine regardless of how they were drafted.
How do I make AI copy sound like my brand? Paste three examples of your own writing and ask it to match them. Examples work; adjectives do not.
What should I never use it for? Product facts, statistics, comparative claims, and anything with legal exposure — unless you verify every line.
Which AI is best for marketing? Claude for customer-facing prose, Gemini for anything needing current information, GPT for consistent bulk output.
How do I stop it sounding like AI? Ban the stock constructions in your prompt, write the first and last sentences yourself, and read it aloud before publishing.