On iOS 27, a lot of your users never read the push you wrote. They read what Apple Intelligence rewrote it into. The lock screen used to show your notification roughly as sent. Now a model reads it, weighs it, condenses it to a line, and sometimes decides not to show it at all. Your copy is an input to that model. It is not what the user sees.
A quick note on names, because the coverage muddles them. The AI digest that rewrites your notifications is Apple’s Notification Summaries feature (sometimes searched as the “Notify Me” digest; the actual Notify Me is a separate Safari tool in iOS 27). This article is about the summarization layer, whatever you call it, and what it does to your push copy.
None of this is brand new. Notification summaries and the Reduce Interruptions Focus arrived with Apple Intelligence back on iOS 18. What iOS 27 changes is how good and how confident the model is. A smarter Siri AI summarizes more aggressively and makes the keep-or-hide call more often, so copy that used to squeak through now gets flattened or filtered. This is a guide to writing push that survives that layer. Pushwoosh is a customer engagement platform, and the AI copy tools and payload controls it gives you map directly onto the levers below.
What the AI layer actually does to your message
3 things happen to a notification before your user consciously reads it, and each one is a different problem for your copy.
Summarized. On Apple Intelligence devices, iOS can rewrite your notification into a single AI-generated line. Feed it clean, self-contained copy and the summary stays faithful. Feed it something vague or context-dependent and the model fills the gap with a guess, and now the user is reading something you never wrote. You cannot dictate the summary. You can only give it copy that is hard to misread.
Grouped. iOS bundles related notifications into one thread and summarizes the set. You steer this with the thread-id key in your payload. Group deliberately and the model condenses a tidy, related set. Leave it to chance and it may stitch unrelated alerts together, which is exactly where meaning blurs.
Filtered. The Reduce Interruptions Focus does not reorder your message. It decides whether the user sees it at all. This is the one to watch, because a smarter model in iOS 27 makes that call more often, and generic promo copy is the first thing it holds back.
The through-line: the lock screen is now a lossy, AI-mediated surface. Treat it like one.
Write for the first few words, because that is what survives
When a summary collapses your message, it works from the top. The opening words carry the meaning that lives; everything after is at risk of being compressed away. Front-load the specific, and never bury the point behind a warm-up.
Take the same promo, written two ways. “We’ve got some exciting news for you this week, so open the app to find out what’s waiting” summarizes to noise, because there is no concrete fact in it for the model to keep. “Your $15 reward expires tonight at midnight” survives any rewrite, because the specifics are the message.
The test is simple. Read only the first 5 or 6 words of your push. If they do not carry the actual value, rewrite until they do.
The payload levers you actually control
You cannot edit the summary the model writes. But 3 fields in your notification payload shape how it gets treated, and most marketers never set them on purpose. They are a one-time setup with your dev team, not a per-send fight.
| Lever | What it does | Use it for |
|---|---|---|
Interruption level (time-sensitive) | Breaks through Focus and the scheduled summary | Genuinely urgent messages only. Users can switch it off per app, so overuse gets your app muted. |
| Relevance score (0 to 1) | Sets where your notification lands in the summary stack | Floating the message that matters most when several of yours arrive together |
Thread ID (thread-id) | Controls how iOS groups related notifications | Grouping on purpose, so the model summarizes a tidy set instead of blurring unrelated alerts into one line |
time-sensitive)thread-id)Set these once, correctly, and every campaign after that inherits sane behavior on the lock screen instead of leaving it to the model’s guess.
An AI-proof copy checklist
The payload handles grouping and breakthrough. Copy has to carry everything the summary actually reads. Before a campaign goes out, run it against this:
- Lead with the concrete. A number, a name, a deadline, a status. Put the fact the user needs in the first clause, not the last.
- Make each notification self-contained. Assume it will be read alone, stripped of the notification before and after it. Copy that only makes sense in a sequence loses in a summary.
- Cut the warm-up. “We’re thrilled to share” and “Don’t miss out” carry no information and get compressed to nothing. The model keeps facts, not enthusiasm.
- Put specifics in text, not tone. Numbers, names, and times survive a rewrite. Vibes do not.
- Do not make the lock screen your only shot. If a message matters, make sure it also lands in the app and, where it fits, over email. A silent push that updates in-app content, backed by an in-app message, does not depend on surviving summarization at all.
Focus filters are a segmentation problem, not just a copy problem
There is a second angle here that most copy advice misses. When users can put your app behind a Focus filter, the ones who keep you breaking through are telling you something: they want to hear from you. The ones who filter you out are telling you the opposite, usually because the last several messages were not worth the interruption.
That is a segmentation signal, and it is worth more than it looks. If you are sending the same broadcast to everyone, you train your most filter-happy users to silence you, and you burn your breakthrough budget on people who did not want it. We’ve watched an app keep raising send volume to hold its numbers up, right up until the OS quietly decided its promos were interruptions, and then the numbers fell anyway. Tighter segmentation, so each message is relevant enough to earn its place, is what keeps you on the right side of that filter. You can act on this with dynamic content that adapts each message to the user, and with behavior-based segments so a promo only reaches the slice it actually fits.
The uncomfortable part: the AI layer rewards relevance and punishes volume, and it does this automatically now, at the OS level, outside your dashboard. The old game of sending more to lift absolute numbers works against you here. Sending less, but making each one land, is not just better practice on iOS 27. It is how you stay visible at all.
Make your push copy work harder with Pushwoosh
Writing self-contained, front-loaded copy at scale, personalized per user and per segment, is not something you want to do by hand across a million-user base. Pushwoosh gives you AI-assisted copy, dynamic content, and the segmentation to match message to audience, so more of what you send survives the summary and reaches the user intact.
