Summarized Before You're Seen: Writing Newsletters for Inboxes That Read First
Photo: Gannu03, CC BY-SA 4.0, via Wikimedia Commons
The reader who arrives first
Something has changed in the design of the American inbox, and most publishers are still writing as though it hadn't. When a subscriber opens Gmail on an Android phone, the message they encounter is not necessarily the message that was sent. Increasingly it is a short paraphrase produced by the provider's own assistant, whether that is Gemini inside Gmail, Apple Intelligence on iOS and macOS, or Copilot in Outlook. That paraphrase is not a preview of the newsletter. For a growing share of recipients, it is the newsletter.
This creates an editorial constraint that has no precedent in print. Your issue must now function both as a piece of writing and as a source document for compression. Those are different jobs, and compression is not neutral. When an assistant decides which sentences to keep, it is deciding, on your behalf, what your publication is about.
What a summarizer is really doing
Nothing mystical happens inside these systems. Given a message, an assistant extracts candidate sentences, scores them for topical centrality and specificity, and produces a short abstract in the voice of a helpful third party. Three consequences follow, and each one should shape how you draft.
First, summarizers favor declarative specificity over build-up. A sentence carrying a number, a name, or a date survives far more often than a sentence that promises something later. Compare 'The filing deadline moves to April 2 and the fee rises to $340' with 'We have a lot to cover this week.' Only one of those can be quoted.
Second, assistants work from extracted text. Decorative structures are invisible to them. An image-led opening, a headline rendered as a graphic, a call to action locked inside a button: these either vanish or get described in the flattest possible terms. If your lead is a photograph with two words of overlay type, the machine sees neither the photo's meaning nor the type.
Third, most summarizers weight the opening heavily and then sample from the middle. Your first hundred words are therefore read twice, once by a human skimming a notification and once by a model looking for evidence.
The preheader is no longer a subtitle
For a decade the preheader was treated as ornament, a place to restate the subject line with a different adjective or to announce that the email can be viewed in a browser. In a summarized inbox it becomes the only sustained prose the assistant sees before your body content, and it becomes the scaffolding for the paraphrase.
Write it as a full sentence that makes a claim. 'Three publishers left their platform this month; here is what their invoices look like now' gives the model a subject and a promise. 'March news, plus platform updates' gives it mood and nothing else, and mood compresses into mush.
Structure that survives a compression pass
Newsletters that summarize cleanly tend to share a handful of habits, and none of them require rewriting your voice.
- One idea per paragraph, stated in the first sentence. Long paragraphs with delayed reveals collapse into their opening clause, so any paragraph that buries its point permanently loses it.
- Subheadings written as assertions rather than atmosphere. 'Churn rose in the sixty-to-ninety-day cohort' tells a model what the section contains. 'The long goodbye' does not.
- Numbers in the running text, not locked inside images or infographics. A statistic in a PNG is a statistic that will not be quoted.
- A short takeaway line near the top of each major section, phrased as a standalone sentence. Summarizers love sentences that stand alone.
- Restrained nesting. Bulleted lists inside boxed modules inside multi-column layouts get flattened in extraction order, and the reading sequence a subscriber sees is not the sequence the model receives.
- Descriptive link text. 'Read the 2026 rate card' survives as a fragment of meaning; 'click here' survives as noise.
Testing what your issue looks like after it has been reduced
The practical problem is that you cannot see your own summary the way a subscriber does, because summaries are generated per user and per client. You can still approximate the result with a short quarterly exercise.
Send each issue to a seed inbox across Gmail, Outlook, and Apple Mail, then capture whatever summary the client produces and compare it against the one-sentence thesis you intended. Separately, paste the plain-text version into a general-purpose model and ask for a three-sentence abstract. Build a small table: issue number, intended takeaway, machine takeaway, delta.
Two failure patterns show up almost immediately. Either your thesis softens into something bland and unobjectionable, or the model latches onto a minor item from the middle of the issue because it was the only concrete thing available. Both are fixable, and both are invisible without the test.
The habits that make it worse
There has been a predictable rush toward what vendors call AI-friendly copy: keyword-laden sentences, paragraphs restructured into staccato fragments, and occasional attempts to hide instructions in white text. The last of these is a deliverability hazard, not a growth tactic. Filters treat concealed text as deception, and the reputational cost at the major mailbox providers is not worth any theoretical gain. Keyword stuffing is simply ineffective, because summarizers abstract rather than index.
What to change this week
Rewrite the preheader as a claim. Move the single most concrete fact of the issue into the first sentence of the body. Rename your subheads so each one states what the section proves. Then send a test and read what comes back.
If the paraphrase misrepresents your newsletter, your subscribers are already reading a different publication than the one you wrote. They will not email to tell you.