You write the email yourself. You pick the subject line, lead with the thing that matters, trim the fluff, and send it to a customer who asked to hear from you. For most of email's history, the next thing that happened was simple: it landed in an inbox and a human read it, or it didn't land at all. There were two outcomes — inbox or spam — and a filter that decided between them.
That model is gone. In January 2026, Google moved Gmail into what it called the Gemini era: AI Overviews that summarize threads, Help Me Write, Suggested Replies, and a new AI Inbox that filters and prioritizes mail before the human sees it. The practical effect is that a large language model now reads your email first. It summarizes what you said, decides how important it is, and presents the human with its interpretation — or, increasingly, decides the human does not need to see your message prominently at all. Folderly's tracking puts it bluntly: up to 40% of emails that technically reach a Gmail inbox are being deprioritized by AI filtering. They pass the old spam check and still lose.
This is the most consequential change to email since Gmail introduced tabs in 2013, and most teams are still writing for the human.
The reader changed, and nobody updated the writing
For twenty years, email writing advice optimized for one thing: getting a person to open, read, and act. Subject lines were written to earn a click from a human scanning a list. Preview text was written to entice. The body could ramble a little, because a person would skim it and find the point.
Now there is a reader in between. Before your customer opens the message, Gemini may have already condensed your three paragraphs into a one-line summary in the AI Inbox, or answered the customer's question about your email without them opening it at all. Gmail's AI Overviews will tell a user "your order ships Tuesday" without the user ever reading the shipping email you spent design hours on. The machine extracted the fact and discarded the packaging.
Two metrics from the first months of the rollout show the shape of it. Open rates went up — Folderly reports figures near 45%, partly because Gmail auto-opens messages to generate summaries, which makes the open-rate metric increasingly meaningless. Click-through rates went down, from around 4.35% to 3.93%, because users extract the value from the AI summary and never click through to the source. If that pattern sounds familiar, it should: it is exactly what happened to web search, where AI Overviews collapsed referral traffic by harvesting the answer and removing the reason to click. Email is now getting the same treatment inside the inbox.
The implication is structural. You are no longer writing for one reader. You are writing for two: the AI that reads first and decides what to surface, and the human who reads the AI's version of you.
What the machine reader rewards
The good news is that the AI reader is not adversarial. It is trying to extract value and present it clearly. The senders who lose are the ones whose emails are hard to extract value from — buried leads, vague subjects, decorative filler. The senders who win write in a way the model can parse cleanly. Across the deliverability analyses published this spring, the optimization advice converges on a short, almost old-fashioned list.
| What the AI reader rewards | What it punishes |
|---|---|
| The point in the first sentence | A warm-up paragraph before the ask |
| Specific, literal subject lines | Curiosity-gap or clickbait subjects |
| Structure: labeled sections, bullets, clear dates | A wall of prose |
| One obvious purpose per email | Five topics crammed into one send |
| Plain language, low filler | Pleasantries, hedging, marketing adjectives |
| Clean, valid HTML | Heavy nested tables, image-only emails |
The first 100–200 characters are now inbox-critical. That window is what the AI is most likely to use when it summarizes your message into a single line, so the benefit to the reader has to be obvious inside it. "Hi there! We hope you're having a great week. We wanted to reach out because..." is dead weight — the model will strip every word of it and may conclude there is nothing worth surfacing. "Your invoice #4021 is due June 3. Pay or update your card here." survives summarization intact, because there is nothing to strip.
Note what is not on the list: tone, brand voice, clever copy. Those still matter to the human on the second read, but they are invisible to the machine on the first. The hierarchy inverted. Clarity and structure used to be hygiene; now they are the difference between being surfaced and being summarized into oblivion.
And none of this matters if you fail the prerequisite. Before Gemini evaluates whether your email is relevant, Gmail checks whether it is trustworthy. SPF, DKIM, and DMARC are the gate that decides whether your message reaches the AI layer at all — fail them and you are filtered before the model ever reads a word. We covered why this got strict in the 2026 deliverability breakdown; the short version is that authentication is now a hard pass/fail and the grace period is over. Writing for the machine reader is the second step. Authenticating to reach it is the first.
The problem optimization can't solve
Here is the uncomfortable part. You can do all of it — perfect authentication, front-loaded value, clean structure — and the fundamental position has still changed. There is now an AI standing between you and your customer, and its job is to give the customer less of you, not more. A summary instead of your message. A one-line answer instead of a click. The better the AI gets at its job, the less of your carefully built email the customer actually experiences.
This is the same dynamic that hit publishers and SaaS landing pages when search went AI-first. The intermediary is not malicious; it is just optimizing for the user's time, and your message is friction in that optimization. You can write to be summarized well, but you cannot write your way out of being summarized. The channel itself now has a layer that mediates, compresses, and deprioritizes everything you send through it.
So the strategic question is not only "how do I write for the AI reader?" It is "where do I have a channel with no AI reader in between?"
The channel with no intermediary
There is exactly one place where you talk to a customer with no summarizer, no priority filter, and no deprioritization: a real-time conversation on a surface you own. When a visitor is on your website and opens a chat, there is no Gemini condensing your answer into a line and deciding whether to show it. The customer asked a question; your chatbot answers it; the customer reads the actual answer. The full message gets through because there is no inbox algorithm sitting in the middle.
That is why the email-to-conversation handoff is becoming the dominant pattern for anything that needs a real exchange. Email is excellent at reaching someone — it is the universal address. It is increasingly poor at conversing with them, because the AI inbox compresses the conversation. So you use email for what it is now good at, getting attention, and move the actual interaction to a channel where the whole message survives. A transactional email front-loads the one fact the AI will surface, then offers a direct line — "questions about this order? chat with us" — that drops the customer into a real-time conversation where nothing gets summarized away.
On the Agentkit side, this is what teams are building: a chatbot trained on their product, policies, and order data, embedded on the site and linked from email, that resolves the question in the conversation instead of in a thread an AI will compress. The customer who would have replied to an email — and had that reply summarized, deprioritized, or blackholed — instead gets a complete answer in the channel you control. A trend post about Gmail's AI does not need to oversell that; the point is narrower and harder to argue with: the on-site conversation is the one customer-facing channel left that has no algorithm deciding how much of your message the customer is allowed to see.
What to actually change this quarter
The work splits cleanly into two tracks: write better for the AI reader, and reduce your dependence on it.
Rewrite your highest-volume emails for summarization. Take your top five transactional and lifecycle emails — order confirmation, shipping notice, password reset, invoice, renewal — and rewrite each so the single most important fact is in the first sentence and the subject line is literal. Assume the customer will only ever read the AI's one-line version. If that one line is correct and complete, you win the AI inbox. If the one line would be "a marketing email from BrandName," you lose.
Fix authentication before anything else. If SPF, DKIM, and DMARC are not all passing, the rewriting is wasted, because you are being filtered before the model reads you. This is a checklist, not a strategy — pass it.
Move the conversation off email. For every email that exists to start a back-and-forth — support follow-ups, "reply with questions," onboarding nudges — replace the reply path with a link to a real-time channel. The reply-to-email loop is now a summarization loop. A link to a conversation is not.
Measure the right thing. Open rate is now noise — AI auto-opens inflate it. Track click-through to your owned channel and, more importantly, what happens after the click: did the customer get their answer, complete the action, convert? The metric that matters is no longer "did they open it" but "did they reach a place where your full message could actually land." The chatbot KPIs guide covers how to instrument the after-the-click side.
The shift underneath the shift
It is tempting to treat Gmail's AI as a deliverability problem — one more filter to optimize around, like the bulk-sender rules before it. It is bigger than that. For the first time, an AI is the primary reader of your email, and it is actively designed to give your customer a compressed, deprioritized version of what you sent. Optimization buys you a better summary. It does not buy you the unmediated attention you used to have.
The senders who adapt will do two things at once. They will write for the machine — literal, structured, front-loaded — so the AI's version of them is accurate and gets surfaced. And they will move every interaction that actually matters onto a channel where no AI gets to decide how much the customer sees. Email becomes the way you reach people; a real-time conversation on your own site becomes the way you talk to them. The teams still writing warm-up paragraphs for a human who will never read them are optimizing for an inbox that no longer exists.
No credit card required.
