Most marketing fails before a word is written
Not because the writing is bad. Because it answers a question nobody asked, in language nobody uses. You write about your artisan sourdough process, and your customers were only ever wondering whether you open early enough for the school run.
AI is genuinely good at fixing this. Just not in the way most people try to use it.
In plain English
- Voice of customer:
- The actual words your customers use about their problem, collected from real sources rather than guessed at.
- Persona:
- A short profile of a typical customer, used to keep everyone writing for the same person.
- Pain point:
- A specific frustration your customer has that your product or service can remove.
- Objection:
- The reason someone almost buys and then does not. Usually price, trust, or effort.
The trap: asking AI to invent your customer
Type into ChatGPT: create a detailed customer persona for my bakery. You will get Sarah, 34, a marketing manager who values authenticity, sustainability and supporting local business. She has a busy lifestyle. She follows food accounts on Instagram.
Sarah does not exist. Sarah is an average of everything ever written about bakery customers, sanded down until she offends nobody and describes nobody. Build a quarter of content on Sarah and you will produce a quarter of content aimed at a statistical ghost.
AI cannot tell you who your customers are. It has never met them. Use it to analyse real input, never to imagine customers from nothing. Everything in this lesson depends on you bringing real material to the conversation.
Where the real language lives
You are already sitting on more customer research than most companies pay for. It is just messy and unread.
- Your reviews, and your competitors' reviews. Competitor one-star reviews are the single richest source there is. People explain in furious detail exactly what they wanted and did not get.
- Support tickets, your inbox and your DMs. Six months of these will tell you what genuinely confuses people.
- Forums and groups. Reddit threads, Facebook groups, trade forums, the comments under industry articles.
- Sales conversations. Whatever you scribbled down after a call that did not close.
You paste that raw material in. AI reads it faster than you can and spots the pattern you are too close to see.
Below are customer reviews. Analyse them and give me:
- The five most frequently mentioned problems, ordered by how often they appear.
- For each one, two or three direct quotes, copied word for word.
- What people praise most often.
- Anything mentioned that surprised you, given this is a [type of business].
Only use what is in the reviews. If something appears once, label it as a single
mention rather than a pattern. Do not add anything from general knowledge.
REVIEWS:
[paste them]
That instruction to copy quotes word for word is the whole game. Summaries drift into marketing language. Direct quotes stay in customer language, and customer language is what you want on the page.
Checkpoint
AI is a pattern reader, not a customer. Feed it real reviews, tickets and threads, and ask it to quote rather than summarise.
Mining language, not just topics
There is a difference between knowing that people worry about cost, and knowing that people say it feels like a lot for something I might not even use. The second one can go straight into a headline. The first one becomes another limp paragraph about affordability.
From the text below, extract the exact phrases people use to describe:
- the problem, in their words
- what they tried before
- what nearly stopped them buying
- how they describe the result once it worked
Give me the raw phrases in a list. No paraphrasing, no tidying up the grammar.
Mark anything that appears more than twice.
TEXT:
[paste reviews, tickets, forum threads or call notes]
โ Weak prompt
Prompt
Create a detailed customer persona for my online plant shop.
Output
Meet Emma, 29, an urban professional who values wellness, sustainability and bringing nature into her small apartment. She is time-poor but design-conscious.
Invented from nothing. Emma is an average of the internet, not your customer. Every plant shop on earth gets a version of this same person.
โ Good prompt
Prompt
Here are 60 reviews of my shop and 40 of my two competitors. What do buyers repeatedly worry about before ordering, in their own words?
Output
Recurring worry: plants arriving damaged. Quotes include it turned up snapped in half, I was scared to order a big one by post, and the last shop sent one that looked half dead.
Grounded in real text. Gives you a delivery guarantee worth writing about, and the exact phrasing to write it in.
From findings to content
Research that sits in a document is not research, it is homework. Turn it into angles the same day.
Here is my audience research: recurring problems, direct quotes and objections.
Give me 12 content ideas that each address one specific item from this list.
For each idea, tell me which finding it comes from and which exact quote it answers.
Do not suggest anything that is not traceable to the research.
Rank them by how often the underlying problem appeared.
RESEARCH:
[paste your findings]
Fifty reviews is a pattern. Five is an anecdote. If you cannot gather much, say so in the prompt and ask for observations to check rather than conclusions to act on.
Do this once properly and it lasts you months. The quotes go into your ad copy, the objections become your FAQ, and the ranked problem list becomes your content calendar in lesson 6. Next, we make sure all of it actually sounds like you.
๐ Quiz
Question 1 of 4Why is an AI-invented customer persona a poor foundation for marketing?