The best evidence we have today
Three pre-registered, randomised controlled experiments ran on the real newsletters of a US wine retailer, each with 27,500 customers, two weeks apiece. The arms: a human copywriter, a language model, and the two combined. Only the body copy differed — the selected products, the button and the image were identical everywhere.
Two results came out of it, and both matter.
One: the email itself is worth a lot. All three mailing arms doubled gross margin compared with the control group that got no email. That is the largest effect in the whole experiment, and it has nothing to do with AI.
Two: no detectable difference between AI and human. Not in purchase rate, not in revenue, not in open rate — in any of the three experiments. The AI arms still came out ahead on annual net profit, because the model licence was on the order of a thousand dollars a year while three copywriters cost $375,000.
So the right conclusion is not that AI writes better copy. It is that it writes copy just as good at a fraction of the cost — and commercially that is quite enough.
One qualification the people quoting this tend to leave out: the experiment varied the body copy of the email, not the subject line. On AI-generated subject lines specifically we found no large-sample measurement with published methodology. If someone promises you one, ask for the data behind it.
What AI does not solve: deliverability
This part is independent of who wrote the email, and in 2026 it is tougher than it used to be.
Gmail. For senders mailing more than 5,000 Gmail addresses a day it is mandatory to have: SPF and DKIM set up; a DMARC record on the sending domain (the enforcement policy may be as loose as p=none); alignment between the sender's domain and the SPF or DKIM domain; a TLS connection; valid forward and reverse DNS; and, for marketing mail, one-click unsubscribe with the right headers. Complaint rate: keep it below 0.10%, and never let it reach 0.30%.
The significant change: from November 2025 Gmail tightened enforcement. Traffic that does not meet the requirements runs into temporary and then permanent rejection. The complaint rate is calculated daily, and a sender above 0.3% is not eligible for mitigation.
Yahoo. Since February 2024: SPF and DKIM, a valid DMARC record with at least p=none, a working one-click unsubscribe, unsubscribes honoured within two days, and a complaint rate below 0.3%.
Microsoft. At 5,000 or more messages a day both the SPF and the DKIM check have to pass, a DMARC record is required, and at least one of them has to align with the sender's domain. If not: a 550 5.7.515 rejection.
And the question everyone actually wants answered: do the providers filter machine-written email? Not one major provider's official sender requirements mention AI-generated content. No prohibition, no labelling duty, no separate filtering criterion. What counts is whether the sender is authenticated, whether the recipient asked for the mail, and how many people report it as spam.
Open rate no longer measures opens
Apple's mail protection, by its own description, "downloads remote content in the background by default — regardless of whether you engaged with the message". In other words, the tracking pixel loads even if the recipient never touched the email.
How big is that distortion? By a July 2026 measurement based on more than a billion opens, 62.26% of recorded opens come from an Apple client, 27.03% from Gmail and 5.83% from Outlook. So roughly six tenths of "opens" come from an environment where images load automatically.
That does not mean the open rate is inflated by six tenths — the data does not break out how many of the Apple opens fall under the protection. It means the open has become an unreliable signal, and no decision should be built on it.
What to look at instead:
- clicks as a share of messages sent (not relative to opens — that puts the same spoiled number in the denominator),
- conversion and revenue per email,
- unsubscribe rate,
- complaint rate — which is also the life expectancy of your deliverability.
Segmentation: this is where AI delivers most
At writing copy, the machine is measurably no better. At segmentation, though, it does something you cannot do by hand: it turns free text into categories.
WHAT YOU CANNOT DO BY HAND, THE MACHINE CAN:
- 4,000 support tickets -> "what the problem was" in 12 categories
- 800 reviews -> which product draws which recurring complaint
- free text from the signup form -> groups by intent
- text notes on past orders -> how the product is used
THIS IS WHAT A SEGMENT COMES FROM. NOT from
"female, 35-44, Budapest".
Prompt template:
"Here are 200 tickets (one batch out of the 4,000). Group them by
WHAT OBSTACLE the customer ran into. Do not hand me
pre-invented categories: they have to come out of the text.
For each group, say how many tickets it holds and quote two verbatim.
Whatever fits nowhere, leave in a 'mixed' group."
Getting that raw material out has a post of its own: working through reviews and support tickets takes it apart at the level of method and of limitation alike.
A segment is good if it justifies a different email. If you would send the same thing to two groups, those are not two segments, just two lists.
The copy: what to ask for and what not to
- Feed it real customer sentences. The quality of generated copy rests on the input, not on a prompting trick.
- One email = one action. With three buttons in it, the measurement is uninterpretable.
- Do not ask for an "attention-grabbing" subject line. Ask for a subject line that says what is in the email. The first kind can bring opens, the second avoids unsubscribes.
- Check the factual claims. Price, deadline, guarantee, stock: these are the ones that cost you the most when they are wrong.
- Shorter is not a handicap. In the experiment cited, the AI emails were on average 96 words shorter — and brought in just as much.
If a landing page hangs off the email, the sample-size rules for testing variants apply there too: landing page copy variants is the post that gives you the numbers.
Common misreadings
- "AI copy converts better." In the most thorough measurement it brought in exactly the same. The advantage is the price.
- "Providers filter out AI-written email." Not one major provider's requirements mention AI. Authentication and complaint rate are what count.
- "The open rate went up, so the subject line is better." Most opens come from automatic image loading. Look at clicks.
- "Segmentation equals demographics." Demographics rarely justify a different email. Intent and obstacle do.
- "Let us just send it and see." Without authentication, "and see" means it never arrives.
What to take away
In email marketing AI does not win on the copy, it wins on cost and on segmentation. It writes copy as well as a good copywriter, far more cheaply; and out of free-text raw material it forms groups you could not form by hand. What it does not solve: authentication, list hygiene, and the fact that open rate is no longer a metric.