Instead of · 5 min read
Using ChatGPT for startup marketing
ChatGPT is genuinely good at marketing advice for about one turn. It stops being useful the moment the job needs memory: it does not know what you already tried, it does not know what worked, and it gives you twelve ideas when you need one.
Every founder who has shipped something and got nowhere has already done this: opened a chat window, pasted their URL, and asked what they should do about marketing.
It is a completely reasonable thing to do, and the answer you get back is usually not bad. So this page is not an argument that you should stop. It is an argument about where the useful part ends, because knowing that line saves you a fortnight.
What it is genuinely good at
Being fair about this matters, because the failure mode of every “why not just use ChatGPT” page is to pretend a very good tool is a bad one.
- Reading your page as a stranger. Paste the text of your homepage and ask what it thinks the product does, without telling it. The gap between its answer and the truth is real, useful information, and it is the cheapest positioning test available.
- Producing variants. Fifteen headlines, eight subject lines, three ways to open a cold email. Volume on demand, no ego about the output.
- Explaining the vocabulary. Payback period, activation, ICP, the difference between a channel and a tactic. It is a patient teacher and it never makes you feel stupid for asking.
- Arguing against you, if you ask properly. “Give me the three strongest reasons this positioning is wrong” gets better results than any request for ideas.
If you use it for exactly those four things, you are getting real value and you should keep going.
The four things it cannot do
These are not prompt problems. No wording fixes them, because they are properties of a conversation that starts from nothing every time.
1. It does not know what is happening right now
The single most valuable marketing input for an early product is somebody, this week, describing your problem in public and asking what to use. That is a live object: a thread posted four days ago, a competitor’s launch comments, a search result that only exists this month.
A model answering from training data will give you the shape of the advice (“go where your audience discusses the problem”) without the instance. The shape is the part you already knew.
2. It does not remember what you tried
Three weeks ago you posted in two subreddits and got nothing. Last week you rewrote the homepage. Ask again today and you will be told to consider posting in subreddits and tightening your homepage copy.
Advice that cannot see your history will keep sending you back to the thing that already failed.
You can paste a summary in each time. Almost nobody does, past the second week, because maintaining the context by hand is more work than the advice saves.
3. It does not know what worked
This is the one that compounds. Marketing for a new product is a sequence of small bets, and the only thing that makes the tenth bet better than the first is knowing how the previous nine went.
A chat window has no idea whether the email you sent produced a signup. So its advice on week twelve is exactly as good as its advice on day one, which is another way of saying you have learned nothing you did not learn yourself.
4. It will not decide
Ask what to do and you get a numbered list of options, each reasonable, usually with a note that it depends on your goals. That is the correct behaviour for a general assistant and it is useless for somebody who already has too many options.
If you have thirty minutes and no marketing instinct, twelve good ideas and one good idea are not close to equivalent. Twelve is how you end up doing none of them and building another feature instead.
How to get more out of it anyway
If a chat window is what you have, four habits make it meaningfully better:
- Keep your own log. One file. What you tried, the date, what happened. Paste it in at the start of every session. This single habit fixes problems two and three by hand.
- Ask for one thing, and forbid a list. “Pick one. Tell me why it beats the others. I am not asking for options.”
- Bring the evidence yourself. Go and find three real threads where people describe your problem, paste them in, and ask which one is worth answering. The reasoning is good when the inputs are real.
- Write down the success measure before you act. What would make this worth having done? Decide that first, or you will judge the result by how it felt.
What we do differently
Do This Next is a persistent, opinionated wrapper around exactly that gap. The reasoning is not the differentiator. The memory and the refusal to hand you a list are.
- It reads your actual site and works from what is on it, and where it is inferring rather than quoting, it says so.
- It commits to one next move, tells you why that one and not the runners-up, and names what would count as it having worked.
- It records what happened, and the next recommendation is shaped by that record rather than starting from zero.
In one sentence: a chat window does not know what you already tried, does not know what worked, and gives you twelve ideas when you need one.
The Product Check is free and needs no account, so the cheapest way to judge the difference is to run it and compare what comes back against what a chat window told you.