How to Write a Cold Email Using Gemini That Actually Gets Replies

Why most AI cold emails get ignored

AI-written outreach fails for one of two reasons: it sounds like marketing, or it asks for something before it gives anything. Left to its own devices, Gemini will do both — it has read a great deal of sales copy and assumes that’s the register you want.

The register problem

Buyers can spot promotional language in about a second and a half. The giveaway isn’t a single word, it’s the density of them: enthusiasm words, abstraction words, and an exclamation mark carrying weight the sentence hasn’t earned.

Asking before giving

The second failure is structural and more expensive. A meeting request is a withdrawal from an account you haven’t paid into. The recipient doesn’t know you, so the honest reading of “do you have thirty minutes?” is “will you give a stranger half an hour of your day?”

Reverse the order and the maths changes. Hand over something small and genuinely useful first, and the ask that follows costs the reader almost nothing to grant.

The Signal-Offer-Ask Method

The Signal-Offer-Ask Method is the structure we teach at PromptPeakAI for cold outreach: one signal you actually observed, one offer that stands on its own, one ask that takes under a minute to answer. Three parts, in that order, inside 90 words.

Each slot has a job, and the job explains the word budget:

• Signal. Proof you looked. A job posting, a product change, a policy update, a number in their annual report. Specific enough that a competitor couldn’t have sent the same line to four hundred people.
• Offer. A small finished thing — a two-page teardown, a checklist, a benchmark you already have. Not a demo, not a “quick chat”, not a discovery call.
• Ask. A yes-or-no question about the offer. “Want me to send it?” is the whole ask. Nothing about your calendar belongs here.

The structure matters more than the wording, which is exactly why it survives being handed to an AI. You’re not asking Gemini to be persuasive. You’re asking it to fill three clearly defined slots without wandering off, which is a much easier instruction to follow. This is the same principle as Role, Goal, Context, Format applied to one narrow task, and it works for the same reason: constraints beat adjectives.

Where Gemini fits into your inbox

Here’s the thing most guides skip. “Write a cold email using Gemini” describes at least three different products with different prices and different availability, and picking the wrong one is how people conclude the workflow doesn’t work.

The practical read: the Gemini app is where you should do this work. It’s free, it’s available almost everywhere, and it’s the only surface that lets you save a Gem. Gmail’s in-line drafting is a convenience for the final polish, not the engine. If you want the full picture of what each tier buys, we’ve broken down Gemini Free versus Google AI Pro and Ultra separately.

How to write a cold email using Gemini

Writing a cold email using Gemini takes about three minutes per prospect once the setup is done. The setup itself is a one-off twenty minutes, and it’s the part that determines whether every email after it is usable or binnable.

Spend five minutes on the prospect’s careers page, their newsroom, their pricing page, or their most recent LinkedIn post. You’re hunting for one concrete, checkable fact — not a vibe. Write it down in a single sentence.

Do this first, deliberately, because the alternative is asking Gemini to invent a reason you’re writing. It will oblige, and the invention will be plausible and wrong. Our guide to researching a prospect before a sales call covers how to make this pass faster.

A Gem is a saved set of instructions you can return to. Building one means you never re-explain your rules, your banned words, or your word count again. In the Gemini app, open Gems, create a new one, and paste the instruction set below.

You help me write short cold outreach emails. You never write anything else.

MY CONTEXT
I am a [your role] at [your company]. We help [who] with [what problem].
The assets I can offer for free are: [list two or three, e.g. a two-page teardown, a benchmark sheet].

STRUCTURE — always exactly three parts, in this order:
1. SIGNAL (about 20 words): the specific thing I observed about them. Use ONLY the detail I give you.
2. OFFER (about 35 words): one of my free assets, described in terms of what they get from it.
3. ASK (about 15 words): a yes-or-no question about sending the asset.

HARD RULES
- Total length 70 to 90 words. Count before you answer.
- No greeting beyond "Hi [first name] —".
- No exclamation marks anywhere.
- Do not use these words: leverage, synergy, supercharge, game-changing, revolutionary, cutting-edge, seamless, transform, delighted, thrilled, excited, robust, innovative.
- Do not write "I hope this email finds you well", "I came across", "I wanted to reach out", "quick chat", "circle back", "touch base", "pick your brain".
- Never request a call, a demo, or a meeting.
- Never include a link, an attachment, or a calendar invite.
- If a detail is missing, write [MISSING: what you need] instead of guessing. Never invent facts about their company.
- Write the way a busy colleague writes their fourth email of the day: plain, direct, slightly clipped.

OUTPUT
Give me the subject line on one line, then the email body. Nothing else. No commentary.

That banned-word block is doing most of the work here. Telling a model what not to do is consistently more reliable than telling it to “be conversational” — a technique we cover in depth in negative constraints in prompting. If Gems are new to you, the full walkthrough of creating a Gem takes about ten minutes.

Open the Gem and give it only what it needs. Resist the urge to paste the whole website — more context here produces longer, blurrier emails, not sharper ones.

First name: [Sarah]
Company: [Northfield Logistics]
Their role: [Head of Operations]

The signal I found:
[Their careers page lists three dispatcher roles in Leeds, all night shift, posted this month.]

The asset I'm offering:
[A two-page write-up of what cut overtime spend for two regional carriers we worked with last year.]

Write the email.

First drafts run long even with a word cap. One refinement pass is almost always enough — and if you find yourself on pass four, the problem is the signal you supplied, not the model.

Cut this to under 75 words without losing the specific detail in the first sentence.

Then check it against these and fix anything that fails:
- Does sentence one name something only this company would recognise?
- Does the offer stand on its own if they never reply?
- Is the ask answerable with one word?
- Would a colleague send this, or does it read like marketing?

Return only the corrected email.

Subject lines written before the body tend to over-promise. Written after, they can just describe what’s inside.

Write five subject lines for the email above.

Rules: two to four words. Lowercase, as if typed quickly. No colons, no questions, no numbers, no company name.
They should read like an internal note between colleagues, not a campaign.

Paste the finished draft into a normal Gmail compose window and send it from your own address. If you’re on a paid Google AI plan or a Workspace account, Gemini inside Gmail can do the final tidy without a copy-paste — but the draft that arrives there should already be right.

Choosing a signal worth mentioning

The quality of the whole email is decided in the first sentence, and the most common mistake is confusing personalisation with relevance. Knowing where someone went to university is personalisation. Knowing they’re hiring three night-shift dispatchers is relevance. A cold email using Gemini lives or dies on which one you hand it.

Gemini will happily manufacture a signal if you don’t supply one, and manufactured signals are the fastest way to get marked as spam by a human rather than a filter. This is the same failure mode behind AI hallucination generally: a model asked to be specific without being given specifics will produce specificity anyway.

If you’re unsure whether what you found is strong enough, test it before you draft:

Here is the fact I found about a prospect: [paste your one sentence].

Answer three questions in one line each:
1. Could this same sentence be sent to fifty other companies? Yes or no.
2. Does it point at an operational problem, or at a person? Which one.
3. If a stranger opened with this, would it read as homework or as surveillance?

If the answer to question one is yes, tell me what to look for instead.

Two rewrites, before and after

Structure is easier to trust once you’ve watched it survive contact with a real example. Both of these started life as an ordinary cold email using Gemini, and were fixed by supplying a signal and enforcing the three slots — not by asking for better writing.

Notice what changed and what didn’t. The good version is shorter, contains no adjectives about the recipient, and gives away the thing a recruiter would normally hold back — salary data — before asking for anything. The ask itself shrank to two words.

The same structure holds across tools, incidentally. If your team lives in a different stack, the ChatGPT version of this workflow follows the same three slots with different mechanics.

Deliverability and privacy, done honestly

Two things get badly misreported in cold email guides, and both matter more than the copy. Let’s take them in order.

What Google’s rules actually cover

You’ll read that Google penalises AI-written email. It doesn’t — a cold email using Gemini is treated exactly like one you typed yourself. According to Google’s own email sender guidelines, the requirements cover authentication (SPF, DKIM, DMARC), keeping your user-reported spam rate below 0.1% and never letting it reach 0.3%, and offering one-click unsubscribe on promotional mail. Nothing in there inspects whether a human or a model wrote the words.

Two details worth knowing:

• The bulk-sender requirements kick in at roughly 5,000 messages a day to personal Gmail addresses. Sending twenty considered emails a week does not make you a bulk sender.
• Those guidelines don’t apply at all to mail sent to Google Workspace accounts — which is where most B2B recipients actually sit.

What does hurt you is the thing the Signal-Offer-Ask Method is designed to prevent: recipients hitting “report spam” because your message was irrelevant. Separately, if you’re sending commercial email at any volume, the FTC’s CAN-SPAM compliance guide sets out the rules that genuinely are legal obligations in the US, including honest subject lines and a working opt-out.

What you paste into Gemini

Prospect research is other people’s data, and Gemini’s consumer terms are not the same as a Workspace account’s. Google states plainly that human reviewers see some conversations, and that chats selected for review can be retained for up to three years even after you delete your activity. The Keep Activity setting is where you turn that off.

What Gemini can’t do for you

Every workflow has a boundary, and pretending otherwise is how people end up disappointed in month two.

When the draft comes out wrong

Most problems trace back to the input, not the model. This table covers what you’ll actually hit in the first week.

One pattern worth naming: if the same fix keeps coming up, edit the Gem rather than correcting the output each time. That is the whole point of building one — every correction you make permanent is a correction you never make again. The context you give the model up front is doing more work than any follow-up instruction ever will.

Your first ten emails

Build the Gem once this week, then run ten prospects through it before you change anything. Ten is enough to see which part of the structure your industry pushes back on, and small enough that a weak offer costs you nothing much to discover.

Writing a cold email using Gemini stops being a copy problem the moment you treat it as a research problem with a writing step attached. The tool was never the bottleneck. Finding one true thing to say was.