Writing email for AI inbox summaries: what is proven and what is not

By Charley Wilcock · Published 2026-10-17 · Last reviewed 2026-10-17

Quick summary: Apple Mail and Gmail now summarize some messages with AI, so the subject line, the first sentences and real text matter more than preview text and image banners. Keep one clear goal per email, put key facts in live text, keep the plain-text version consistent, and authenticate your mail. Schema.org markup works in Gmail only for registered senders, and there is no public evidence that other assistants use it.

What changed

Two behaviors are worth knowing, as reported by industry testing and documentation reviews in 2025 and 2026. Apple Mail can show an AI-written summary in the message list, in place of the preview text you wrote. Gmail’s Gemini summaries are reported to be generated after a message is opened, so they do not decide whether it gets opened. Both are generated from the message content, and both can be wrong, which is why testers recommend checking how your own emails come out. These features change quickly, so confirm the current behavior in each provider’s documentation before relying on any detail here.

What you can do that does not depend on guesswork

  • Make the subject and first two sentences stand on their own. If someone read only those, would they know what the email is and what to do? Preview text can no longer be relied on to finish the thought.
  • One clear goal per email. Testers report that when a message makes several asks, a summary may pick one, and not always the one you wanted.
  • Put critical facts in live text, not images. Dates, totals, codes and deadlines flattened into banners cannot be summarized. The copy checker flags image-heavy messages.
  • Send a real plain-text version with the same facts and the same link as the HTML. Do not leave the default “view in browser” boilerplate.
  • Use sound structure. Real headings and paragraphs, and role="presentation" on layout tables, make email easier for people and software to read.
  • Never use hidden text. Security researchers have shown that text hidden with zero font size or white-on-white can steer an AI summarizer. Clients may defend against this, and mail that does it is likely to be treated as abusive.

Structured data (JSON-LD): what is true

Gmail does support Schema.org markup in emails, in JSON-LD or microdata, for types such as orders, parcel deliveries and reservations. But it only shows these cards for authenticated senders that have registered with Google, and unregistered markup is often ignored or stripped. Gmail’s supported types and rules change, so check Google’s current Gmail email markup documentation. I found no public evidence that Apple Intelligence or other assistants read JSON-LD in emails. Treat claims that markup “guarantees” correct AI summaries as unproven. It may be worthwhile for high-volume transactional senders who qualify for Gmail’s cards, but it is not a general fix.

Authentication still matters

Inbox providers decide first whether your mail is trustworthy. Reports on Apple’s email features say they expect DKIM-authenticated mail and DMARC compliance, and Gmail, Yahoo and Microsoft have all tightened sender requirements. Check yours with the SPF, DKIM and DMARC lookup, and see moving DMARC to reject.

A test you can run yourself

  1. Prepare two versions of one email: your current design and a version that follows the checklist above.
  2. Send both to test accounts on devices that have the features: Apple Mail with Apple Intelligence summaries on, and Gmail with Gemini.
  3. Record the subject, the preview or summary shown, and what Gemini says after opening. Take screenshots with the date and software versions.
  4. Repeat with a few real emails and several days of results, since summaries vary.

Publish what you find with the method. Original, documented tests are far more useful, and far more likely to be cited, than general advice.

What is not known

  • How much any of these changes affect opens, clicks or revenue is not established in public data I could find.
  • How the summarizers weigh HTML structure, plain text and metadata is not publicly documented, and it changes with each release.

Sources checked

Need hands-on help with this? I freelance on email platforms, automation logic and migrations.

Start a project