Marketing prompts built around the thing most AI copy misses: a specific customer, a specific problem and a claim you can actually support.
5 copy-ready prompts
Positioning to ad copy
Research included
The prompts
Prompt 1
Positioning statement from real customer language
Before writing any copy at all.
I sell [product] to [customer type].
Here is what real customers have said about it:
"""
[paste reviews, support tickets, sales call notes or survey answers]
"""
Using only the language in those quotes:
1. List the problems customers describe, ordered by how often they come up.
2. List the words and phrases they use that we do not use in our own marketing.
3. Write three positioning statements in the format: For [who] who [problem], [product] is the [category] that [benefit], unlike [alternative].
Do not add benefits that no customer mentioned.
Why it works: The third instruction is the important one. Marketing copy invented without customer input is a guess dressed up as a claim.
Prompt 2
Landing page structure with a single job per section
Building or rewriting a product page.
Write the copy for a landing page selling [product] to [audience].
Structure it as: hero, problem, how it works (3 steps), proof, objection handling, pricing framing, final call to action.
For each section give me: the heading, the body copy, and one sentence explaining what that section must achieve.
Constraints:
- The hero must say what the product is in plain words before saying why it is good.
- Every benefit must be tied to a concrete feature.
- Objection handling must address these specific objections: [list your real objections].
- No superlatives I cannot prove.
Why it works: Requiring an explanation per section exposes filler. If a section cannot state its job, it does not belong on the page.
Prompt 3
Ad variations across different angles
Testing messaging rather than testing adjectives.
Write 8 ad variations for [product] aimed at [audience] on [platform].
Use a different angle for each: problem-first, outcome-first, objection-first, comparison, specific number, customer quote style, curiosity, and direct offer.
For each: headline (max [x] characters), body (max [y] characters), call to action, and the angle name.
Keep every factual claim to things I can support: [list what is actually true].
Why it works: Eight variations of the same angle tell you nothing. Testing different angles tells you what your market actually responds to.
Prompt 4
Customer research interview guide
Getting the raw material the other prompts need.
I want to interview customers of [product] to understand why they bought and what nearly stopped them.
Write a 20-minute interview guide with:
- An opening that gets them talking about their situation before mentioning my product.
- Questions about the last time they faced [problem], asking what they actually did.
- Questions about what they tried before finding us.
- Questions that surface hesitation without being leading.
Mark any question that risks leading the answer, and suggest a neutral replacement.
Why it works: Ask about past behaviour, not future intentions. What someone did last month is evidence; what they say they would do is a prediction.
Prompt 5
Rewrite feature copy as customer outcomes
Turning a feature list into something persuasive.
Here is a list of features for [product]:
"""
[paste your feature list]
"""
For each feature give me a table row with: the feature, the job it does for the customer, the outcome they get, and a one-sentence version I could put on a website.
Flag any feature where the outcome is weak or unclear - those are candidates for cutting from the page rather than rewriting.
Why it works: The flagging step matters. Not every feature deserves page space, and a shorter page usually converts better than a complete one.
Using AI for marketing without producing forgettable copy
AI is very good at producing marketing copy that sounds like marketing copy. That is the problem. The prompts here are structured to force in the specifics that make a message land: real customer language, real constraints, real proof.
Research first, copy second
The most common mistake is asking a model to write the headline before anyone has established what the customer actually struggles with. The model will happily produce something polished, and it will be built on assumptions nobody has checked. Feeding real material into the prompt - reviews, support tickets, sales call notes, survey answers - changes the output completely, because the model is now recombining your customers' words rather than the internet's average landing page.
You usually have more of this material than you think. Thirty support tickets and a dozen reviews are enough to see a pattern in the language people use for their problem, and that pattern is worth more than any brainstorm.
Test angles, not adjectives
When teams ask for twenty ad variations they usually get twenty versions of the same idea with different verbs. That tells you which synonym performs marginally better, not which message resonates. Angle testing - problem-first against outcome-first against a direct comparison - tells you something you can use everywhere else, including in your sales conversations and your product page.
Common messaging angles and when they work
Angle
Best when
Risk
Problem-first
The audience feels the pain but has not looked for a solution
Can feel negative if overdone
Outcome-first
The result is easy to picture and desirable
Sounds like every competitor
Comparison
You are displacing a known alternative
Invites feature-by-feature arguments
Specific number
You have a real, verifiable statistic
Useless if the number is vague
Objection-first
A known hesitation blocks most buyers
Only works if the objection is real
Claims have to be defensible
A model asked for persuasive copy will invent statistics, testimonials and comparative claims without hesitation. Every prompt on this page includes a constraint limiting claims to what you can support, and that constraint should survive into your own versions. Beyond the obvious legal and advertising-standards issues, fabricated proof is the fastest way to lose a customer who checks.
Where AI genuinely saves time
Volume and variation - producing ten angles to choose between rather than committing to your first idea.
Format conversion - turning one strong piece into an email, a short post and an ad without rewriting from scratch.
Structure - laying out a page or a sequence so you are editing a draft rather than staring at nothing.
Synthesis - reading fifty pieces of customer feedback and reporting what actually recurs.
Where it does not
Strategy, pricing and brand voice all require judgement about trade-offs a model cannot see. It does not know your margins, your roadmap, the customer segment you have quietly decided to stop serving, or which competitor is about to undercut you. Treat the output as a first draft produced by a fast, well-read assistant who has never met your customers.
Measure the message, not the output
Producing ten times more copy is only useful if you can tell which version worked. Before a test, write down what result would make you change the page: a click-through difference of a given size, a signup rate, a reply rate on outbound. Then give the test enough traffic to say something - small differences on small samples are noise, and acting on them is how teams end up rewriting a page every fortnight without improving it. The prompts here are for generating options; the decision still belongs to the data.
Related prompt packs
The ChatGPT Writing Prompts cover editing and voice-matching, which pairs well with anything here. If you are marketing a technical product, the AI Coding Prompts help with documentation, and teams hiring for marketing roles will find the AI Resume Prompts useful from the other side of the table.
Frequently asked questions
Can AI write marketing copy that converts?
It can write competent drafts quickly. Conversion comes from accurate positioning and real customer language, so feed the model genuine research and treat its output as a starting point for testing.
What should I never let an AI model write unchecked?
Statistics, testimonials, comparative claims about competitors and anything with regulatory implications. Models generate plausible versions of all four.
