When an AI reads your email for them: what 'open rate' even means now

Search went through this exact transition first. For years, “did they click through” was the measure of whether content worked. Then AI-generated answers started satisfying the question directly, on the results page, and a huge share of searches stopped producing a click at all — not because the content failed, but because the answer arrived without it. Email is now entering the early stage of the same shift, and it’s worth asking honestly what “open rate” and “reply” even measure once an AI can summarize a message before a human looks at the original.

The adjacent, well-documented case: zero-click search

This isn’t a hypothetical pattern. In search, it’s already been measured for years, and the trend accelerated sharply once AI-generated answers became common on results pages.

The reason this matters for email isn’t that the mechanism is identical — it isn’t. A search “answer” and an email “summary” are different products serving different intents. What transfers is the underlying dynamic: when an AI layer sits between the content and the person, and does a competent job of extracting what matters, the downstream action (a click, an open, a reply) stops being a reliable proxy for “did this reach and register with a human.” Search proved that dynamic can move fast — a nearly 8-point shift in two years — once the AI layer is good enough and widely deployed enough.

What’s actually documented for email, right now

Unlike the search case, we don’t yet have a multi-year, precisely measured trend line for “how often is a cold email summarized instead of opened.” What we do have is confirmation that the underlying mechanism — an AI reading and summarizing messages before a human necessarily does — is live, not theoretical, in the two biggest mailbox providers. The full rundown of what’s shipped in Gmail and Outlook is here; the short version is that Gmail’s AI summary and “AI Inbox” features are on by default for US accounts across a user base reported in the billions, and Outlook’s Copilot thread summarization completed its rollout to eligible users by the end of January 2026.

The broader context: AI mediation is rising fast, generally

It’s worth being careful here not to overstate what’s specific to email. What’s genuinely documented is a broader, faster shift in how people interact with AI assistants generally: roughly half of US adults (49%) now report having used an AI chatbot (ChatGPT, Gemini, Copilot, or Meta AI), up from 33% in 2024, per a Pew Research Center survey fielded in February 2026 — adoption up sharply in under two years. That statistic is not about email specifically, and shouldn’t be read as evidence about email open or reply behavior on its own. What it does establish is the pace at which AI-mediated interaction has become normal, ordinary behavior for a majority-adjacent share of the population, in the same window that Gmail and Outlook shipped their own inbox-specific summarization features. The email-specific mechanism and the general-adoption trend are two separate facts pointing in a consistent direction, not one fact standing in for the other.

What this plausibly means for “open” as a metric

An email open, as currently tracked, is a tracking-pixel fetch — and we’ve written at length about how much of that signal is already noise: Apple’s Mail Privacy Protection pre-fetching, Gmail’s image proxy caching, and corporate security scanners all fire the pixel without a human present. AI summarization adds a new, different kind of ambiguity on top of that existing one. Consider the plausible outcomes for a recipient whose mail client shows them an AI summary card before the original message:

• The pixel never fires, because the summary is generated server-side or from cached content, and the recipient never renders the original HTML at all. Your dashboard shows this as “not opened” — which is technically true and practically misleading, since the recipient did register the content of your message, just not through the channel your tracking measures.
• The pixel fires later, or not at the moment of genuine attention, if the recipient opens the original message only after deciding, from the summary, that it’s worth a closer look — inverting the usual assumption that an open precedes engagement.
• A reply gets triggered from a suggested-reply interface built on the summary, rather than from the recipient composing a response inside the opened message — which may or may not register cleanly in reply-tracking depending on how the platform threads it.

None of these are measured phenomena yet. They’re the reasonable set of ways a documented mechanism (AI summarization sitting upstream of the open event) could plausibly distort a metric (open rate) that was already, separately, unreliable. The honest position is that we don’t know the size of this effect yet — only that the mechanism enabling it is now live at scale, which is more than could be said a year ago.

How to think about measurement without overcorrecting

The instinct to declare “open rate is dead, replace it with X” is exactly the kind of overclaim this piece is trying to avoid. A more grounded approach:

1. Keep leaning on reply, not open, as the primary signal. This was already the right call before AI summarization entered the picture, because opens were already dominated by automated fetches. AI summarization is a second, independent reason to distrust the open event, not a new problem requiring a new solution — the existing fix (measure replies) still applies.
2. Watch for a rise in “no open, but reply” patterns as a leading indicator. If a growing share of your replies come from contacts your tracking never recorded as having opened the message, that’s a concrete, measurable signal that summarization-before-open is happening in your own list — worth watching in honest analytics that separate real human signals from noise, rather than assuming.
3. Don’t rewrite your copy strategy around a hypothesis you can’t yet measure. The concrete, already-actionable response to AI summarization is writing specific, front-loaded copy that survives compression — that’s good practice regardless of how big the zero-click effect turns out to be, which is exactly why it’s the right thing to act on now, versus a metrics overhaul built on a number nobody has measured yet.

Search’s zero-click shift took roughly two years to become undeniable in the data, once the AI layer was in place. Email’s AI-summarization layer has been in place, broadly, for well under two years. If the analogy holds even loosely, the measurement conversation this piece is having now is early — which is the right time to be honest about what’s documented and what’s still analysis, rather than waiting for a definitive number that, in search’s case, only became clear in hindsight.

Why reply is likely to hold up better than open, even here

There’s one structural reason to expect reply-based metrics to age better than open-based ones through this shift, and it’s worth stating plainly rather than leaving it implied. An “open” is a passive byproduct of rendering — it happens whether or not the content registered, which is exactly why a summarizer sitting upstream of it is such a clean way to break the signal further. A reply is not passive in the same way: it requires the recipient (or, per the speculative scenario above, an agent acting with some delegation from the recipient) to decide the message was worth a response and to produce one. That decision-and-production step is a meaningfully higher bar than an image rendering, and it’s much harder for an AI summarization layer to manufacture accidentally the way a tracking pixel gets fired accidentally today. This isn’t a guarantee reply stays perfectly clean forever — the speculative “AI drafts and sends a routine reply on the recipient’s behalf” scenario would complicate it too — but it’s a reason to expect the open-rate erosion documented in search, and now plausible in email, to hit “open” harder and sooner than it hits “reply.”

The question worth sitting with

None of this argues for panic, or for a wholesale replacement of email metrics before there’s data to justify one. It argues for something more modest: treating “open rate,” which was already compromised by bot pre-fetching before AI summarization entered the picture, as a metric on borrowed time rather than a stable baseline — and building measurement habits now, around replies and verified human signals, that don’t depend on an assumption (a human directly rendered this message) that’s getting less safe to make every year this decade. Our own benchmark work already treats reply, not open, as the honest scoreboard — the AI-summarization shift isn’t a reason to change that conclusion, it’s one more reason it was right in the first place.