← All posts

Why ChatGPT replies fail on LinkedIn

April 22, 2026 · 6 min read

I ran the test in February. 200 cold LinkedIn DMs to founders with a stock GPT-4 prompt ("write a friendly cold outreach about X"), then 200 written in my actual voice with the same opening hook. Same target list, same sender, same week.

Reply rate, GPT prompt: 4.5%. Reply rate, my voice: 18%.

That gap isn't because GPT writes badly. It writes fluently. The replies were grammatically perfect, polite, structured. They just didn't sound like a person — they sounded like every cold outreach the recipient had archived that week.

The pattern recipients learned to filter

LinkedIn inboxes have been getting flooded with AI-generated outreach since late 2022. Three years in, your average mid-senior IC can spot it in two seconds. The signals they clock without consciously thinking about it:

Why GPT defaults to this

Two reasons, neither of which is the model's fault.

First, the training data. Most professional writing on the internet is corporate English — earnings calls, sales emails, press releases, LinkedIn posts that got engagement. The base statistical pull of "write a professional message" lands somewhere in the middle of that distribution. The middle of corporate English is the exact register humans now identify as "AI-flavored."

Second, the prompts most people use. "Write a friendly outreach about X" is a generic instruction with no voice anchor. The model has nothing to differentiate on, so it regresses to the safest middle. You can prompt-engineer your way to a better register ("write like a 28-year-old founder who uses em-dashes and never says 'circle back'") but you'll spend more energy on the prompt than you would just writing the thing yourself.

What works instead

The thing that moved my reply rate from 4.5% to 18% wasn't a smarter prompt. It was giving the model 5 examples of how I actually write. Past replies, varied contexts — thanking a mutual for an intro, declining a partnership, asking a stranger for advice. Then asking it to draft a new reply matching that style.

Cost-wise it's barely different from a stock prompt. Quality-wise it's a different product. The reply suddenly has my contractions, my cadence, my way of opening. It doesn't say "I hope this finds you well" because I don't say that.

I built Reply Coach around this realization. You paste 5–10 of your past LinkedIn messages once, and every reply it drafts gets those samples as a style anchor in the system prompt. The samples never go to a fine-tune — they stay on your device, get re-sent each generation.

The boring-but-true takeaway

AI outreach isn't dead. Generic AI outreach is dead. The new floor for LinkedIn DMs is "indistinguishable from a real human who writes well" — and the way you get there isn't a smarter model, it's giving the model your actual voice.

Most tools haven't caught up to this yet. They optimize for volume and templates, which is the opposite direction. The tools that survive 2026 will be the ones that lean into personalization — not because it's morally better, but because it's the only thing that works once recipients learn to filter.


Reply Coach is a Chrome extension that does the voice-mimicking thing I described above on every LinkedIn DM. 5 free generations a day, $9/month for unlimited. See how it works →