Your Email Gets Summarized Before It Gets Read: How to Write for the AI Inbox

Back in August I argued you should treat open rate as a mood ring. This one's about something less comfortable for your subject line: for a growing share of your subscribers, an AI is deciding what your email says before a human has read a word of it.

Google announced a batch of Gemini features for Gmail in January, Apple has been putting summaries on lock screens for a while, and the honest state of the evidence is "we're all guessing a little." So I read what the vendors and deliverability teams have actually published, separated what's confirmed from what's folklore, and boiled it down to what I'd change in a send this week.

Here's the question worth sitting with: if a model reads your email first and hands your subscriber a two-line version, which two lines would it pick – and would you be happy with them?

What actually changed

On January 8, 2026, Google announced three Gemini-powered features for Gmail: AI Overviews (summaries of long threads, plus the ability to ask questions about them), upgraded Suggested Replies, and an AI Inbox that highlights things it thinks need action – bills, appointments, unanswered messages. Klaviyo's read on it is the cleanest one-liner I've seen: Gmail is trying to turn the inbox into a prioritized task list.

On the Apple side, Apple Intelligence shows a short summary of a message in the inbox list, on the lock screen and in notification center. According to MailPro's May 2026 write-up, that summary replaces the preheader in those spots rather than sitting beside it.

What we know, and what we really don't

Here's the part that most of the "how to win in the AI inbox" posts skip: nobody outside Google and Apple knows how these systems weigh your subject line, your preheader and your body copy. Klaviyo says so in its own article – "We don't yet know exactly how Gmail's AI will summarize, categorize, or prioritize brand emails at scale." Rollout is uneven too: some features are on by default, some are tied to paid Google plans.

One more fact worth keeping in your back pocket: Google's own sender guidance says it doesn't track open rates or verify third-party open-rate accuracy. Which is one more reason the metric you've been leaning on was already shaky before any of this arrived.

What I'd change: be specific, early, and in plain words

Every source I read, whatever its angle, lands in the same place, and it happens to be good writing advice anyway: say the specific thing, say it first, and say it in words that survive being shortened. "20% off linen until Sunday" survives a summary. "Spring has sprung" gets flattened into nothing.

If you're sending acquisition email, this matters more

A subscriber who knows you can skim past a vague opener. A stranger can't, and neither can a model summarizing for a stranger. So in a first-touch or welcome-adjacent email, spend sentence one on who you are and why this message exists: "Hi Sam – you downloaded our deliverability checklist last week; here's the one check most senders skip." That reads fine to a person and summarizes cleanly for a machine.

Make the email readable as text

Klaviyo's practical advice is worth borrowing wholesale: image-only emails are harder for AI to interpret, so use alt text that describes the action and value ("Shop the winter sale"), keep navigation, footers and offer banners as live text, and consider plain-text or lightly formatted emails for launches, founder notes and follow-ups. Explicit action language – confirm, register, complete, download – may be more likely to get surfaced as something actionable, though that's an expectation, not a tested result.

How to know if it's working

Not with opens. Judge a change on unique click-to-delivered, click-to-open, conversions and unsubscribe rate. The cleanest test is small: split your list randomly, keep send time and offer identical, change only the first sentence and subject line, and compare clicks. Then run the results through the Campaign Post-Mortem so you're comparing against benchmarks instead of your gut.

What I'd do this week

1. Take your last five sends and read only the first two sentences of each. If the offer or news isn't in there, that's your rewrite list.
2. Rewrite subject lines so they name the topic or offer – and the deadline, if there is one.
3. Cut the throat-clearing opener. "Hope you're well" has never once been the news.
4. Write the preheader as if it will be shown, but don't depend on it – it's still the fallback in Outlook, older clients and spam filters.
5. Replace image-only heroes with live text and give every image alt text that describes the action.
6. One job per email, one button, one specific verb on the button.
7. Record your click and click-to-open baseline now, before the rollout fully lands, so you have something to compare against later.

None of this requires believing any particular theory of how Gemini ranks mail. It's the same advice good copywriters have given for decades – lead with the point, be specific, make the next step obvious – with a new reader in the room. If the AI inbox turns out to matter less than the hype, you've still written better emails. If it matters more, you're already ahead.

Stephen Taunt