--- title: "Write emails that AI inbox summaries get right" description: "What Gmail's Gemini and Apple Intelligence do with your email, per their own help pages, and how to write mail that their summaries get right." url: https://www.unitpost.com/blog/agents/emails-ai-summaries-read-correctly section: Blog updated: 2026-11-02 --- # Write emails that AI inbox summaries get right Gmail and Apple Mail now summarize email before people read it, and neither says how its model decides what to keep. Write for the summary the way you would for a busy reader: the point, the date and the action in the first lines, as live text rather than images, a plain-text part that says what the HTML says, and nothing hidden. Then read the summaries on accounts you own. ## Write emails that AI inbox summaries get right > What Gmail's Gemini and Apple Intelligence do with your email, per their own help pages, and how to write mail that their summaries get right. Gmail and Apple Mail now put a model between your email and the person you sent it to. Gmail can lay an AI Overview over a thread, and Apple Mail can show a one-line summary where the first lines of the message used to be. The reader decides whether those are on, and neither company publishes how its model picks what to keep. What you control is the email. ### What do Gmail's Gemini summaries do with your email? Gmail's help pages describe several places where Gemini condenses mail before anyone reads it: - AI Overview conversation summaries. Gemini can "generate an AI Overview at the top of the thread that synthesizes key points and replies in the conversation". On the web the reader clicks Summarize this email at the top of a thread (Gmail Help (https://support.google.com/mail/answer/16561387)). Google's feature list offers them on web, iOS and Android, to personal and paid accounts, in eight languages (Gemini features in Gmail (https://support.google.com/mail/answer/16831098)). - Reminders from due dates. When a thread "mentions a due date with a task required from you", the reader can create a reminder straight from the AI Overview. - Summary cards. Cards at the top of the inbox for order tracking, events, travel, bills and promotions. When a card is missing or thin, the help page says "there may not be enough details in the email" (summary cards (https://support.google.com/mail/answer/15195630)). - AI Inbox. A beta for US English accounts on eligible plans. Its Suggested to-dos show "high-priority items from incoming emails that need your attention". It reads only the Primary tab, filters out spam, and does not support attachments (AI Inbox (https://support.google.com/mail/answer/16845247)). Summaries depend on smart features, which the reader can turn off. Google's own caveat: "Gemini features may suggest inaccurate or inappropriate information." ### What does Apple Intelligence do in Mail? Apple's Mail guide for iPhone and Mac describes three things (iPhone (https://support.apple.com/guide/iphone/iph9ae667055/26.0/ios/26.0), Mac (https://support.apple.com/guide/mac-help/mchlb2dbea8f/26/mac/26)): - Summaries in the message list. "A short summary automatically appears under each email in your inbox" (on the Mac, under each unread email). A reader who wants the first lines of the message back turns off Summarize Message Previews. With summaries on, the preview text you wrote may never be shown. - Summarize. A summary at the top of an open email. "If the message is part of an email thread, the contents of the thread are summarized." - Priority Messages. "Your most urgent emails, like a same-day dinner invitation or boarding pass", shown at the top of the inbox. The page gives examples, not rules. Apple's caveat: "Apple Intelligence uses generative models and outputs may vary." ### Why put the point in the first lines? Both features compress, and a model can only keep what it can find. An email that opens with a greeting, a hero image and two sentences of context makes the summary work for the point. Lead with the fact the reader needs. "Your order 4821 ships Thursday" beats "Thanks for shopping with us". Write dates as dates with a time zone, amounts with a currency, and one action. Keep the subject and the first line saying the same thing. Don't write urgency you don't mean in the hope of landing in Priority. The summary reads the body too, and a reader who sees an urgent subject over a routine update learns to skip your mail. ### Why do the facts need to be live text? A summary is built from what the model can read. Neither help page says images are read, and Gmail says outright that AI Inbox does not support attachments. If the order number, the date or the code is in a banner image or only in an attached PDF, assume the summary won't have it. Put every fact in HTML text, and give each meaningful image `alt` text that says what it shows. When Unitpost derives a plain-text part from your HTML, an image with `alt` becomes `[image: alt]` and an image without one leaves nothing behind. ### Write the action as a sentence "Click the button below" gives a summary nothing to carry. "Confirm your new address by Friday, 17:00 UTC" does, and it is the shape Gmail's reminders look for: a task and a due date. Put the action and its deadline in a sentence near the top, then let the button repeat it. One email, one main action. If a receipt also asks for a review and announces a feature, the summary has to choose, and it may not choose the receipt. In a thread, the summary covers every message. When a later email changes something, say so in words: "New time: 15:00 UTC (was 14:00)". Otherwise the summary may report both times, or the old one. ### Why send a real plain-text part? Neither Google nor Apple says which MIME part its model reads, so make both parts say the same thing. On Unitpost, `text` is a real `text/plain` body and is sent unchanged. If you send only `html`, Unitpost derives the text part for you: it removes the markup, keeps readable link destinations, and leaves out hidden elements such as a hidden preheader (`display:none`, `visibility:hidden`, `mso-hide:all`). Write `text` yourself when you need exact control, and keep it matching the HTML. If an AI agent drafts your mail, have it produce both parts from one draft. Two separate prompts give you two emails that drift apart. ### Why should nothing in the email be hidden? Some senders have tried text that people can't see, aimed at the reader's AI. Don't. Hidden instructions in a `display:none` span or white-on-white text are exactly how prompt injection against email agents (/blog/agents/protect-ai-email-agent-prompt-injection) works, so mail that tries to steer a model looks like an attack to anything built to stop one. It also breaks the deal with the reader: the summary and the email should agree. The same goes for preview text. A preheader that tells a different story from the body is hidden text with a nicer name. Keep it a true one-line version of the point. ### How do you check what the summary says? Read the summaries the way your recipients will: - Send to accounts you own. Use Gmail on the web and on a phone with smart features on, and Apple Mail on a device with Apple Intelligence and Summarize Message Previews on. On Unitpost, `POST /email/templates/{id}/test-send` sends one test of a template to an address you choose. - Compare the summary with the email. Does it name the point, the date and the action? When it gets a fact wrong, find the sentence it came from and rewrite that sentence. - Read the plain-text part. `POST /email/templates/{id}/render` returns the subject, HTML and text a send would carry, with `missing_variables`, and sends nothing. Read the text on its own: if it doesn't make the point, a model reading it won't either. - Load it with images off. Whatever disappears is what a model may not see. - Reply to it once. Then check the thread summary, because that is what the reader sees when the conversation continues. A summary can't help mail that never reaches the inbox, and Gmail's AI Inbox filters out spam. Check SPF, DKIM and DMARC with the email header analyzer (/tools/email-header-analyzer), and see why transactional email goes to spam (/blog/deliverability/email-going-to-spam) if it isn't arriving. ## FAQ ### Can a sender stop Gmail or Apple Mail from summarizing their email? Not in any way either company documents. The switches belong to the reader: Gmail's summaries depend on its smart features settings, and Apple Mail has a Summarize Message Previews setting. What a sender controls is the email, so write it so the summary comes out right. ### Does Gmail summarize every email? No. Gmail's help describes AI Overview summaries at the top of a thread, summary cards for orders, events, travel and bills, and a beta AI Inbox that lists to-dos from the Primary tab for US English accounts on eligible plans. Which ones a reader sees depends on their account, language and settings. ### Can hidden text make an AI summary say what I want? Neither Google nor Apple documents how its model treats hidden HTML, and hiding instructions for a model is how prompt injection against email agents works. Mail that tries to steer the reader's AI looks like an attack to anything defending one. Put what you want said in the visible text. ## Related - [Gmail sending limits and API scopes for AI agents](https://www.unitpost.com/blog/agents/gmail-limits-for-ai-agents): What an AI agent sending through Gmail runs into: daily sending limits, Gmail API quota units, and restricted scopes that need a yearly security review. - [How email webhooks work (they're POST)](https://www.unitpost.com/blog/concepts/how-email-webhooks-work): An email webhook is an HTTP POST to your URL when a message is delivered, bounced, opened, or clicked — not polling, not IMAP. - [OAuth or API key: how to connect an AI agent to an MCP server](https://www.unitpost.com/blog/agents/mcp-oauth-vs-api-key): What the MCP spec says about OAuth, how Unitpost's hosted server signs a person in, what consent grants, and when a scoped API key fits better.