AI-written cold emails don't go to spam because an AI wrote them. The big mailbox providers publish what they ask of senders, and the list is about authentication, complaints, bounces and unsubscribe links, not who typed the words. What does hurt is the thing AI drafts do by default: sound the same as every other draft. Identical copy sent in volume looks like a pattern to a filter, and generic phrasing looks like a stranger to a reader. Below are the five things filters measure, what readers notice instead, and a read-aloud test to run before you send.
One caution: nobody outside Google, Microsoft or Yahoo can see inside their filters, so anyone who tells you exactly how they work is guessing. What follows is based on their published sender guidance and on what happens in practice.
Five signals a filter can actually measure
| Signal | What it is | What you control |
|---|---|---|
| 1. Authentication | SPF, DKIM and DMARC records proving the mail is really from your domain | Set them up once; see cold email deliverability basics |
| 2. Spam complaints | The share of recipients who click "report spam". Google's published guidance for bulk senders is to stay under 0.10 percent and never reach 0.30 percent | Write to people who have a reason to care; make opting out easy |
| 3. Bounces | Mail sent to addresses that don't exist | Verify addresses before sending |
| 4. Sending pattern | Sudden jumps in volume, or hundreds of near-identical messages from one new account | Ramp slowly; vary the message by real facts |
| 5. Links and format | Many links, shortened URLs, attachments, heavy HTML, images with little text | Plain text, one link at most, no attachments in the first email |
Notice that none of the five asks who wrote the text. A draft from ChatGPT and a draft from you are treated the same if you pass all five. A draft you wrote yourself and sent identically to 500 people on day one of a new account is the one that gets caught.
What readers notice instead
The reader is a better detector than any filter. They won't think "AI wrote this." They'll think "this could have come to anyone," and they'll delete it or, worse, press the spam button. That click is signal number 2, which is how a generic message turns into a deliverability problem: not because it's machine-written, but because it makes people click.
Hana runs a window-tinting shop and used a chatbot to draft 30 emails to property managers. The first draft:
I hope this email finds you well. I wanted to reach out because tailored window solutions can transform your properties and deliver cutting-edge energy savings for your portfolio.
Nothing in it is false, and nothing in it is about the reader. After editing, with one fact Hana found on the manager's own site:
Your Harbor Street building lists south-facing offices on the top two floors. Those are the rooms that get unbearable by 2 p.m. in July. We tint commercial glass, and a single floor takes us about a day.
That version has a place, a time, a number, and a sentence only a person who looked would write. It also differs from the next manager's email in the part that matters, which is why it doesn't read as a template.
Why identical copy at volume hurts
If 200 emails go out in an hour with the same body, a filter that compares messages sees the same text arriving from one sender. Swapping in the first name doesn't change that. The fix is not to shuffle synonyms with a tool, which makes the writing worse. The fix is to send fewer emails with a real, different fact in each, which is the approach in personalization at low volume. Where to find the fact is covered in cold email opening lines.
A read-aloud test before you send
Read the draft out loud, at normal speed, and check five things. It takes about a minute per email.
- Could I send this to the next name on my list unchanged? If yes, it needs a fact.
- Is there a sentence I would never say to this person's face? Cut it.
- Did I stumble on any phrase like "I hope this finds you well," "reach out," or "tailored solutions"? Replace it with what you mean.
- Are all the sentences about the same length? Break up the rhythm, add a short one.
- Does it contain one thing only I would know, such as a street, a price, or a past job?
Fail two or more and rewrite before sending. How to write the whole message is in how to write a cold email that gets replies.
Keeping your own voice in the draft
Give the model three things you've actually written to customers and ask it to match them. Cut anything that sounds like a brochure. Use contractions. Keep the ask to one plain sentence. If you're wondering whether to say a machine helped, that's a separate question, taken up in do you have to tell people an AI sent your email.
Sending from your own Gmail
The sending address matters as much as the text. Mail from your own Gmail or Google account has a history, and complaints and replies attach to you, so there is no shared sender whose reputation you're borrowing. Rocketship sends from your own Gmail. The rest of the picture is in why cold email is not spam.
Rocketship finds the buyers, emails them from your own Gmail, follows up, and books the meeting on its own. Plans that get you customers start at $24.99 a month (Launch), and both Launch and Frontier start with a 7-day free trial, card required. See the AI workers.