Your deliverability numbers probably look fine on paper. Open rates, click rates — all tracking normally. The problem is what's happening before the open: your emails aren't reaching the primary inbox anymore.
Google and Microsoft updated their spam filters in late 2025. They're not just checking domain reputation or DKIM records now — they're running their own LLMs to analyze the semantic content of incoming mail. If the email pattern looks like it came from ChatGPT, it routes to Promotions. Or spam. Automatically.
If you're using ChatGPT to write cold emails and blasting them through Apollo or Instantly, your pipeline didn't get worse. You got filtered.
Here's how to fix it.
#1. The Semantic Fingerprint Problem
LLMs generate text with highly predictable token distributions. Words like "delve," "testament," "unlock," and structured transitions like "Furthermore" or "In conclusion" have become statistical fingerprints for AI-generated content — and modern spam filters know exactly what to look for.
The fix isn't to use less AI. It's to constrain what the AI produces.
Bad prompt: "Write a cold email to this CTO about our security product."
Better prompt: "Write a 3-sentence email to this CTO. Use a 6th-grade reading level. No corporate jargon. No adverbs. Make it sound like a quick note typed on an iPhone while walking."
Same tool. Completely different output. One goes to spam, one doesn't.
The Cold Email Deliverability Checklist
Get the exact 12-step technical checklist we use to bypass spam filters and land in the primary inbox, delivered instantly.
Where should we send it?
You'll also get the weekly briefing.
#2. Stop Sending the Same Email 500 Times
Old spintax — {Hi|Hello|Hey} {First Name} — is dead. Spam filters parse it in milliseconds. The approach that's actually working: route every lead through an agent before sending, and have the agent rewrite the entire email structure from scratch.
The prompt looks something like: "Rewrite this core offer specifically for [Lead Name] at [Company]. Change the opening hook to reference their recent [Specific Event]. Ensure the sentence structure is completely different from the base template."
Every email in your sequence ends up with a unique cryptographic hash. Not just different words — different structure, different rhythm, different length. That's what gets you past fingerprinting.
Yes, this takes more setup. But you're doing it once, and it's the difference between landing in primary and burning your sending domain in 30 days.
#3. Plain Text Outperforms HTML
HTML-heavy emails with tracking pixels, images, and branded templates are being penalized. Google's heuristic is simple: if it looks like a marketing brochure, it belongs in Promotions.
Revert to raw text.
Strip all HTML formatting. Remove the open-tracking pixel — it hurts deliverability more than the metric helps you. Remove your calendar link from the first email too. Every link in the initial message decreases deliverability. Your goal for email number one is a reply, not a booked meeting.
When they reply, your domain is now trusted. Then you send the link.
#4. Never Use Your Main Domain for Cold Outreach
This is the one that still surprises founders. If your primary domain gets blacklisted, your transactional emails, investor updates, and internal comms go down with it.
The standard setup:
- Buy 5-10 secondary domains (
tryfounderbrief.com,founderbrief.io,getfounderbrief.com) - Set up Google Workspace or Microsoft 365 inboxes on each
- Warm them for 14-21 days using Smartlead or Instantly before sending a single cold email
- Cap volume at 30-40 emails per inbox per day — never more
Yes, this takes three weeks before you can send anything. Do it anyway. The founders who skip this step end up with a blacklisted main domain and a support ticket to Google that goes nowhere.
Cold email isn't dead. Lazy AI cold email is dead. The founders who build the systems to get past the filters will own the channel.