I Gave ChatGPT, Claude, and Gemini the Same Cold Email Brief — Here's What Happened
I gave all three of the biggest AI tools the exact same cold email brief and told them nothing except what you'd find in a real sales scenario. One wrote something I'd actually send. One was decent but forgettable. And one produced the kind of generic, soulless copy that gets you marked as spam before a human even reads it. If you're using AI to write cold emails — or you're about to start — what I found matters more than any benchmark test you'll read on a tech blog.
The Exact Brief I Used (And Why It Was a Fair Fight)
Every test needs rules. I gave all three tools the exact same prompt with zero extra coaching or system instructions.
Here's the brief I used:
"Write a cold email to a marketing director at a mid-size e-commerce brand. I'm selling an AI-powered email analytics tool that helps reduce churn. The email should be under 150 words, feel human, open with a hook that isn't a compliment or a question, and end with a low-friction CTA. No buzzwords."
That's a real brief. It has constraints, a target persona, a specific product benefit, and a tone instruction. Any decent copywriter could work with it — and any AI worth using should too.
I ran each test on ChatGPT (GPT-4o), Claude 3.5 Sonnet, and Gemini 1.5 Pro on the same day, using the same browser, same account types (free and Pro where I have access). No cherry-picking runs. First output only.
The differences were immediate. Not subtle — immediately obvious within the first sentence of each response.
What Each AI Actually Wrote (The Results, Unfiltered)
Let's start with ChatGPT (GPT-4o). Its opening line was:
"Most e-commerce brands discover they have a churn problem the same way — by losing customers they didn't realize were already leaving."
That's a strong hook. It's specific, it creates a pain point without being dramatic, and it doesn't start with "I hope this email finds you well." The rest of the email stayed tight — under 150 words, one clear benefit mentioned, and a CTA that read: "Worth a 15-minute call this week?" Low friction. No pressure. It felt like a real human wrote it after doing their homework.
Claude 3.5 Sonnet came out swinging differently. Its opening line was:
"You're probably sitting on a goldmine of customer signals your current tools aren't reading."
More provocative. More confident. Claude's email was slightly punchier — shorter sentences, bolder claims — but it also felt a touch more like a pitch and slightly less like a conversation. The CTA was solid: "I can show you what that looks like for your store in 20 minutes." Still good. Still sendable.
Gemini 1.5 Pro opened with:
"In today's competitive e-commerce landscape, retaining customers is more important than ever."
Dead on arrival. That's the kind of opening line that gets deleted before the second sentence. The rest of the email didn't recover — it used the word "synergies," included three separate CTAs, and came in at 210 words after I specifically asked for under 150. Gemini ignored more of the brief than it followed.
The ranking was clear: ChatGPT first, Claude second, Gemini a distant third.
The Hidden Metric Nobody Talks About: Brief Compliance
Winning a cold email test isn't just about which output sounds best — it's about which AI actually listened.
Brief compliance is how well an AI follows every specific instruction you gave, not just the ones that are easy. Word count. Tone. Format. Things you explicitly said to avoid. This is the metric that separates a useful writing tool from a fancy autocomplete.
ChatGPT followed 5 out of 5 instructions. Under 150 words: yes. No opener that's a compliment or question: yes. Human tone: yes. No buzzwords: yes. Low-friction CTA: yes.
Claude followed 4 out of 5. The tone veered slightly into pitch territory, which I flagged as "no buzzwords" being loosely interpreted — it used the phrase "actionable insights" once. Still, 4/5 is strong.
Gemini followed 2 out of 5. Wrong word count, multiple CTAs, and opened with one of the most overused sentences in B2B marketing. That's not a minor miss — that's a tool that's generating text without actually processing constraints.
Here's why this matters for you: when you use AI for outreach at scale, brief compliance determines whether you can trust the output without reviewing every single email. If you're sending 50 personalized cold emails a week, you can't babysit every output. You need an AI that does what you said.
A useful mental model: treat your AI prompt like a job posting. If a freelancer ignored half the requirements in your job posting, you wouldn't hire them. Same logic applies here.
How to Build a Cold Email System Using the Winner (Today)
You don't need a fancy workflow. Here's a simple three-step process using ChatGPT GPT-4o that you can set up in the next 30 minutes.
Step 1: Build your master prompt template. Don't start from scratch every time. Create a saved prompt that includes your product, your ICP (ideal customer profile), your tone rules, and your constraints. Here's a starting structure:
"Write a cold email to a [job title] at a [company type]. I sell [product] that helps [specific benefit]. Keep it under [word count] words. Open with a pattern-interrupt hook — no compliments, no questions. End with a single, low-friction CTA. Avoid: [list your banned words/phrases]. Tone: direct and human, not salesy."
Fill in the brackets once and save it somewhere you can paste it fast — Notion, Apple Notes, anywhere.
Step 2: Generate 3 variations in one session. After your first output, follow up with: "Give me two more versions — one more aggressive, one more conversational." You now have three angles for A/B testing without writing three separate briefs.
Step 3: Run a self-audit before sending. Paste the final email back into ChatGPT and ask: "Does this email sound like it was written by AI? What's the weakest sentence? What's the most generic phrase?" This takes 30 seconds and catches the lines that slipped through.
This system works because you're using ChatGPT's brief compliance strength and its self-critique capability together. Most people only use step one. Steps two and three are where the real quality jump happens.
The Part Most People Get Wrong
Most people treat AI cold email tools like a magic button — type in your product name and hit send. That's wrong, and it's why so many AI-written cold emails end up in spam folders or get ignored.
The quality of your AI output is almost entirely determined by the quality of your brief. A vague prompt like "write a cold email selling my analytics tool" will get you a vague, generic email every single time — from any AI. ChatGPT's edge only shows when you give it constraints to work with.
The second mistake: people don't iterate. They take the first output and either send it or abandon it. The real workflow is first output → critique → one revision → send. That extra step takes two minutes and closes 80% of the gap between "AI-sounding" and "actually human."
The third mistake is assuming the best AI for cold email is also the best AI for everything else. Claude 3.5 Sonnet is arguably better than ChatGPT at long-form writing and nuanced reasoning. Gemini has real strengths in research and multimodal tasks. This test was specifically about cold email brief compliance — don't generalize the result beyond that.
Key Takeaways
- Brief compliance: ChatGPT GPT-4o followed every instruction in this test — that's the most important factor for cold email at scale.
- Claude 3.5 Sonnet: A strong second — punchier tone, but slightly less precise on constraints. Worth testing for your specific voice.
- Gemini 1.5 Pro: Struggled with word count, CTAs, and tone in this specific use case — not the right tool for cold outreach right now.
- Prompt quality: Your brief is 80% of the result — specific constraints, persona details, and banned phrases are non-negotiable.
- The revision step: Always run a self-audit prompt after your first output — it takes 30 seconds and dramatically improves the final email.
What to Do Right Now
Open ChatGPT and paste this prompt in the next 10 minutes:
"Write a cold email to a [your target job title] at a [your target company type]. I sell [your product] that helps [one specific benefit]. Keep it under 150 words. Open with a pattern-interrupt hook — no compliments, no questions. One CTA only, low-friction. No buzzwords."
Read the output, then immediately follow up with: "What's the weakest sentence in this email and why?" That two-prompt combination will get you further in 10 minutes than most people get in a week of tinkering.