I Sent the Same Cold Email Brief to ChatGPT, Claude, and Gemini. Only One Got a Reply.

Last month I tested something most AI content skips entirely: which model actually writes cold emails that get replies, not just emails that sound good in a screenshot. I gave ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro the exact same brief — same target, same offer, same context — and sent all three variations to real strangers. Only one email got a genuine reply from someone who didn't know me. If you're using AI to write outreach and getting ignored, this article shows you exactly why, and which model to actually use.

The Brief I Gave All Three Models (And What Each One Did With It)

Here's the exact prompt I used across all three tools: "Write a cold email to a marketing director at a mid-size SaaS company, offering a free audit of their email deliverability. Keep it under 120 words, no fluff, sound like a real person not a salesperson."

ChatGPT came back fast and polished. It opened with "I hope this email finds you well" — which, if you've ever received a cold email, you know is an instant delete trigger. It hit the word count, structured everything perfectly, and read like every other cold email sitting in that director's spam folder.

Gemini took a different angle, leading with a statistic about email deliverability rates. Technically accurate, but it read like a stat pulled from a listicle, not something a real person would casually drop into a first message. It felt researched, not personal.

Claude did something neither of the others did: it asked a clarifying question before writing anything. It came back with "Quick question before I draft this — do you know anything specific about their current email setup, or should I write this as a cold, no-context outreach?" I said cold, no context. Claude then wrote a three-sentence email that opened with a specific, plausible observation about deliverability issues SaaS companies commonly face, made one clear ask, and ended without a hard sell.

That third email is the one that got a reply. Not because Claude is "smarter" — because it understood that cold email lives or dies on specificity and restraint, not polish.

Why Most AI-Written Cold Emails Fail (Even When They're Well-Written)

Here's the insight nobody talks about: the problem isn't grammar, structure, or even tone. It's that all three models default to writing emails that sound like they're trying to sell something, because that's the pattern they've absorbed from millions of training examples of marketing copy.

Cold emails that actually work do the opposite. They sound like a person noticed something specific and is casually flagging it — not like a company is "reaching out to explore synergies." The AI models don't know this instinctively. You have to force it.

The mental model that changed my results: treat the AI like a ghostwriter, not a copywriter. A copywriter is trying to persuade. A ghostwriter is trying to sound like a specific human wrote this specific message to this specific person. Those are completely different jobs, and most people's prompts accidentally ask for the first one.

Here's the technique that fixed it for me: instead of asking for "a cold email," I ask for "a two-line message this person might get from a former colleague who noticed something and wanted to mention it." That single framing shift changes everything the model produces — shorter sentences, less structure, no pitch-y language.

I also stopped asking for a full email in one shot. I now ask each model to give me three different opening lines first, before writing the body. The opening line is 80% of whether a cold email gets opened or deleted, and testing it separately gets dramatically better results than asking for the whole thing at once.

How to Actually Do This Today (Step by Step)

Open Claude first — based on my testing, it's the strongest at nuance and restraint, which matters more than raw fluency for this specific task. If you don't have Claude, ChatGPT-4o works, but you'll need to fight its "polished salesperson" instinct harder.

Step 1: Don't ask for the email yet. Ask this: "I need to write a cold email to [specific person/role] about [specific offer]. Before you write anything, ask me 3 questions that would help you write this better." This forces the model to gather the specificity it needs instead of guessing.

Step 2: Answer those questions honestly, even if your answers are short. The more specific your context — even one real detail about the recipient's company — the less generic the output.

Step 3: Ask for three opening lines only, not the full email: "Give me 3 different opening lines for this email. Each should sound like something a real person would casually notice, not a marketing hook." Pick the one that sounds least like an ad.

Step 4: Build the rest of the email around that opening: "Now write the rest of the email around this opening line. Keep it under 90 words, one clear ask, no sign-off like 'Best regards' — just a first name."

Step 5: Read it out loud before sending. If it sounds like something you'd actually text a person, send it. If it sounds like a LinkedIn ad, go back to Step 3 and try a different opening.

The Part Most People Get Wrong

Most people ask AI to write "a professional cold email" and then send whatever comes back without editing a single word. That's the mistake. Professional is the exact wrong instruction — it's the word that triggers the stiff, salesy tone that gets deleted in two seconds.

The second mistake is treating length as a virtue. People ask for 150-200 word emails because it feels thorough, but every cold email study on open rates says the same thing: shorter emails get more replies, because they respect the reader's time instead of asking for it.

The third mistake, and the biggest one: people don't edit the AI's output at all. Every model I tested produced a solid first draft, but the version that actually got a reply was the one I cut down by 30% after Claude wrote it. AI gets you 80% of the way. The last 20% — cutting the fluff, adding one real detail only you'd know — is what makes a stranger respond.

Key Takeaways

• Claude wins for tone control: It defaults to restraint and specificity, which matters more than polish for cold outreach.
• Avoid the word "professional" in your prompt: It triggers stiff, salesy language that readers instantly recognize and delete.
• Separate the opening line from the rest of the email: Test 3 openings before writing the full draft — the first line decides everything.
• Shorter always wins: Cap your cold email at 90-100 words regardless of which model writes it.
• AI gives you the draft, not the send-ready version: Always cut 20-30% of the output before it goes out.

What to Do Right Now

Open Claude or ChatGPT right now and paste this exact prompt: "I need to write a cold email to [insert real person/role you want to reach]. Before you write anything, ask me 3 questions that would help you write this better." Answer honestly, then follow the 5-step process above before you send your next outreach message.