Use ChatGPT to Draft Press Releases That Journalists Actually Open
Contents
Last quarter I watched a client ship a Series A press release they had spent three weeks polishing. The headline used the word "exciting" twice. The opening paragraph was a 90-word mission statement about "revolutionizing the future of work." We sent it to 14 journalists on a curated list. Twelve never opened the email. The other two replied with a one-line "thanks, not a fit."
The same client, two months later, ran a rewritten version through a 4-part ChatGPT prompt I now use for every press release. The new version led with the dollar number, named the specific problem, and said "no comment available" on a couple of questions the original release had soft-pedaled. We sent it to a smaller, more targeted list of 6 reporters. Three replied within 48 hours. One turned it into a 600-word feature.
The first release was generic because the client wrote it alone. The second one wasn't better because ChatGPT is magical — it was better because the prompt forced a structure the client's brain was too close to the company to produce on its own. Here is that structure.
Why most press releases die in 3 seconds
Journalists receive 50 to 200 pitches a day. The 3-second scan decides everything. They are looking for exactly four signals, in this order:
- Is this actually news? Funding, a real product launch, a customer win with numbers, a regulatory change, a hire that matters — not "an exciting new chapter."
- Do I know the beat this is for? Subject line and first sentence should make the angle obvious in 5 words.
- Is the claim specific? "$12M Series A led by Sequoia" beats "significant funding round."
- Is there a quote with an actual point of view? A CEO saying "we are thrilled to announce" is the fastest way to get deleted. A CEO saying "we got tired of watching X happen, so we built Y" gets read.
If your draft fails any of those four tests in the first 30 words, it is going in the trash and ChatGPT cannot save you by rewriting the body. The fix has to be structural.
The 4-part prompt
I have run this prompt on roughly 30 press releases over the past year — funding announcements, product launches, partnership news, executive hires, conference participation. It produces a usable first draft about 80% of the time. The other 20% I rewrite by hand, but the prompt still saves me 30 minutes of staring at a blank doc.
Part 1 — Role and rules. Paste this at the top of a fresh chat:
You are a publicist who has placed stories in TechCrunch, The Verge, WSJ, and trade publications. You write press releases that journalists read. Rules:
- Lead with the news, not the mission statement
- Never use: excited, thrilled, delighted, revolutionary, game-changing, cutting-edge, empower, unlock, leverage
- Use AP style: numbers under 10 spelled out, % as percent on first use then %, no Oxford comma, dates as "March 30, 2025"
- First paragraph must answer all 5 Ws: who, what, when, where, why it matters
- Quote from a real human at the company — never "the team is excited" or any quote you can't attribute to a named person
The banned-words list is the most important part. Without it, ChatGPT defaults to PR fluff. I have tested 12 versions of this prompt and the biggest lift in quality comes from telling the model what not to do, not what to do.
Part 2 — The 5W brief. This is where you give it the raw material. Format it like this so the model can't dodge specifics:
Company: [name, one line on what they do] News type: [funding / product launch / partnership / hire / etc.] Headline facts: [dollar amount, lead investor, all named parties, exact date] The problem the product solves: [one sentence, concrete] Proof it works: [customer name, metric, or quote from a real user] Spokesperson: [name, title, the actual point of view they want to make] Distribution target: [TechCrunch / trade pub / regional / industry analyst] Embargo: [yes with exact time, or no]
The "distribution target" line matters more than people think. A release aimed at TechCrunch reads completely differently from one aimed at a niche trade publication. Without that line, ChatGPT averages them and the result is bland.
Part 3 — The actual ask. Then prompt:
Write a 350-word press release in AP style, inverted pyramid (most important in the lead, least important at the end). Sections: headline, dateline, lead paragraph (all 5 Ws), supporting paragraph with the proof point, quote from the spokesperson, quote from a customer or investor if available, about-company boilerplate, media contact.
After the release, give me 3 alternative headline options ranked by likely open rate, and a 12-word subject line for the email pitch.
Part 4 — The reality check. Don't accept the first output. Ask for a second pass:
Now critique your own draft. Are there any banned words? Is the lead under 40 words? Could a journalist who knows nothing about the company understand the news in the first sentence alone? If not, rewrite.
That self-critique step catches about 60% of the problems the first draft ships with. It's also the easiest thing to forget, which is why I build it into the prompt rather than leaving it to willpower.
The editor pass ChatGPT can't do for you
Even a great AI draft needs 15 to 20 minutes of human editing before it goes to a journalist. Here is the order I do the edits in:
1. The hedge check. ChatGPT loves to write "up to 30% faster" and "over 1,000 customers." Delete every "up to," "over," "more than" unless the underlying number is provable. Journalists and their editors will fact-check, and a "up to" claim you can't defend is the end of your relationship with that reporter.
2. The named-source audit. Every quote in the draft must be attributable to a person who actually said it, on the record, and who knows it is going to a journalist. I have watched a CEO get caught off guard by a quote ChatGPT invented that "sounded like him." Send the quote to the spokesperson in Slack before the release goes out, every time, no exceptions.
3. The not-news removal. Read each paragraph and ask: if a journalist cut this paragraph, would the news still be clear? If yes, cut it. Most AI drafts run 30% too long because the model pads to feel complete. Real press releases for product news are usually 250 to 400 words. Funding announcements can run to 500.
4. The link and asset check. Make sure the boilerplate links to a real About page, the media contact is a real monitored inbox (not a no-reply), and the high-res logo and product screenshots are attached or linked from a press kit page. A press release without working assets looks amateur and ends the conversation before it starts.
When not to use a press release at all
The biggest mistake I see in 2025 is sending a press release when the news does not justify one. ChatGPT makes it dangerously easy to dress up small news in big-release formatting, and the result is a flood of releases journalists are filtering out more aggressively every year.
If the news is "we added a feature," that's a blog post and an in-app banner. If the news is "we have a great new partnership" with no customer impact, that's a co-marketed LinkedIn post. Save the press release for news that someone outside your company would actually want to know about — funding, a major product launch, an industry-first certification, an acquisition, a regulatory milestone.
ChatGPT will happily write any of these formats. Your job is to decide which one the news actually deserves.
The single biggest leverage point
If I had to pick one thing in this workflow to get right, it is the 5W brief in Part 2. Press release quality is bottlenecked by how well you can write down the news before you ask the model to draft the release. Vague brief in, vague release out — every time, no matter how clever the prompt.
A useful test: if you can explain the news in two sentences to a smart friend who knows nothing about your company, and they would say "oh, that's actually interesting," the brief is good. If they would say "wait, what does the company do again," the brief is not ready. The model can't fix that. You have to.
The 80/20 of press release drafting with ChatGPT is not the prompt. It is being honest with yourself about whether what you are announcing is news, and being specific about it before you touch a keyboard.