You have probably had the conversation. Deliverability is fine. Domains are warm, DMARC passes, the seed tests come back green, inbox placement sits north of 90 percent. And replies are flat. Somebody suggests new copy. Somebody else suggests more volume. Nobody suggests that the email arrived, sat in the right folder, and was never actually read by a human being.
That third possibility is the one that got materially more likely this year, and the Gmail AI inbox is why.
On January 8, 2026, Google started rolling Gmail into what it calls the Gemini era. Conversation summaries went out to everyone at no cost. Ask-your-inbox question answering went to AI Pro and Ultra subscribers. And a feature called the AI inbox, which filters low-priority messages so the important ones stand out, entered trusted testing with a broader rollout promised in the following months. All of it running on Gemini 3, US and English first. Google set the whole thing out in its own announcement, which is worth reading as a sender rather than as a user, because the framing throughout is about protecting the recipient's attention from mail like yours.
None of this was aimed at cold email specifically, which is precisely why it is going to be effective against it. Anti-spam systems can be studied, gamed and eventually beaten, because they are adversarial by design and both sides know the rules. An attention-ranking system is not adversarial. It is just a model doing its best to work out what a person cares about, and a message from a stranger about a problem they have not told anyone they have is a legitimately hard case for it to rank highly.
A spam filter asks whether this message is legitimate. A priority filter asks whether this message is worth your time. Those are not the same question, and the second one is much harder to pass with a template.
The metric everyone optimizes stops one step short
Inbox placement was always a proxy. It was a good proxy, for a long time, because in a chronological list of undifferentiated messages, arriving in the primary tab was most of the battle. Get past the filter, and the subject line did the rest.
Chronological lists are what is going away. Not the inbox, the list. When software sorts by predicted importance rather than arrival time, placement guarantees existence and nothing else. Your message is in there. It is on page two of a view nobody scrolls.
This is the uncomfortable part for anyone who has spent two years building infrastructure. All of that work still matters. It just stopped being sufficient, and sufficiency was the thing it used to have. We wrote about the spam threshold half of this problem when Gmail's filtering thresholds shifted under cold email programs, and the vendor comparison work on inbox placement across the major sending platforms holds up fine. Neither of them measures whether a person laid eyes on the message.
What the Gmail AI inbox actually changes for senders
Three things, and they compound.
The first is triage. A model ranks what is important. Bills, appointments, replies to threads the user started, messages from people they have corresponded with before. A first-touch email from a stranger sits at the bottom of every one of those heuristics by construction. Not because it is spam. Because it is genuinely, measurably, less likely to matter than the invoice due Thursday.
The second is summarization. Where a thread gets summarized, the words that reach the reader are not necessarily your words. They are a model's compression of your words. Two paragraphs of positioning become one clause. TechCrunch reported back in May 2025 that Gemini would summarize long emails automatically unless the user opted out, and the direction of travel since has been more summarization rather than less.
The third is retrieval. Ask-your-inbox turns the mailbox into a searchable corpus. Your email stops being a moment and becomes a record that may or may not surface later when someone asks a question it happens to answer. That is a slower, stranger channel than the one outbound teams are used to, and it rewards completely different writing.
Four gates, not one
It helps to draw the funnel properly, because most outbound dashboards draw two boxes where there are now four.
| GATE | DECIDED BY | WHAT PASSES IT | CAN YOU MEASURE IT |
|---|---|---|---|
| Accepted | Receiving server, authentication and reputation | Correct SPF, DKIM, DMARC, warm domain, sane volume | Yes, bounce and acceptance logs |
| Placed | Spam classification | Content that does not look like bulk mail, healthy sender history | Partly, via seed tests and placement tooling |
| Surfaced | Priority model deciding what the person sees first | Relevance to this specific recipient, right now | No, not from the sender side |
| Read and understood | The person, often via a summary | A concrete claim that survives compression | Only indirectly, through replies |
Two of those four gates have no sender-side telemetry at all. That is the actual state of outbound measurement in 2026, and pretending otherwise is how teams end up scaling a sequence that stopped working two months ago. The honest response is not better tracking, because there is nothing to track. It is shorter feedback loops on the one number that still means something.
Illustrative funnel for 1,000 sends under a priority-sorting inbox, expressed as a percentage of messages accepted. Figures are an illustrative scenario for shaping expectations, not measured data.
Run your own numbers rather than borrowing those. The point of the shape is that the two biggest drop-offs have moved to the two stages you cannot see, and every optimization habit the industry built is pointed at the two you can.
Writing for a reader that skims on your prospect's behalf
Here is the strange thing about writing for a summarizer. It wants almost exactly what a busy human wants, only more strictly, and with no patience for craft.
Front-load the specific. Not the pleasantry, not the flattering observation about their recent funding round, not the setup. The concrete claim, in the first sentence, with the number or the named thing in it. A summary of an email whose first line is a compliment produces a summary that says someone paid you a compliment.
Name entities, not categories. We help companies like yours improve efficiency compresses to nothing. We cut onboarding time for three mid-market claims processors from eleven days to four survives compression, because every element of it is a fact that has to be carried through to stay true.
Ask one question. A message with three asks compresses into an ambiguous summary, and ambiguity reads as low priority to a model and to a person. One question, answerable in one line, is both good manners and good machine legibility.
Keep it short enough that summarization is pointless. If the whole email is four sentences and each one carries a fact, there is nothing to compress and the reader gets your words rather than a paraphrase. Brevity used to be a courtesy. It is now a control mechanism.
“Brevity used to be a courtesy. Under a summarizer it is a control mechanism, because a message short enough to need no compression is the only kind that reaches the reader in your own words.”
What to stop doing this quarter
Stop reporting open rates internally. They were degraded before this and they are now actively misleading, because a model reading a message on a person's behalf is not a person reading a message. Every hour spent explaining an open rate movement to a leadership team is an hour not spent on the number that survived.
Stop treating inbox placement as a finish line. Keep the infrastructure work, keep the seed tests, keep the domain hygiene. Just move the goal post to reply rate per sequence and stop celebrating the halfway mark.
Stop adding volume to fix a relevance problem. The maths changed. Adding sends to a list that is not tightly matched now costs you at the reputation gate and the priority gate simultaneously, and the second cost does not show up anywhere you are looking. The work goes into the list and the first sentence instead, which is slower, harder and considerably more likely to produce a meeting. The performance data we looked at on cadence versus copy in machine-written outbound points the same direction: structural decisions beat surface polish.
And stop assuming this is a Gmail-only story. Priority sorting and assistant summaries are being built into every major mail client. Gmail is simply the one that told everybody the date.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.