Priority Inbox predicts how likely you are to do something with a message. It learns that from the way you’ve treated mail like it before. That’s a real prediction, and a narrow one. Every complaint people have about Priority Inbox comes out of the distance between what it predicts and what you meant by important.
- It’s one yes-or-no, reused in several places. The Priority Inbox layout, the Important first layout, the yellow marker,
is:importantand your notification settings all read the same flag. - The prediction is about you. Google’s own published description says the feature “ranks mail by the probability that the user will perform an action on that mail.”
- Interesting keeps beating important, and Google said so first. Its research note names opening a message as a strong signal, then concedes that people open a lot of mail that is interesting rather than important.
- A first message from a stranger has nothing behind it. Most of the model’s strongest signals measure a relationship, and with a stranger that relationship starts today.
- Correcting a marker moves a threshold. You can make the promoted section bigger or smaller. The question it answers stays the same.
This page is about that one flag and where it gives out. Choosing between Gmail’s six inbox layouts, and everything else that shortens a busy list, is how to organize a busy Gmail inbox. The label set you apply yourself is a separate system from the importance flag, and it’s making Gmail labels work for you. Pushing a thread into next week is how to snooze and follow up in Gmail. Which of the remaining limits belong to Gmail and which belong to email itself is what Gmail can’t do on its own.
The one flag underneath all of it
Gmail keeps a single yes-or-no on every message: important, or not. Everything in this piece sits downstream of that one bit.
You meet it in four places. It’s the yellow marker beside a message. It’s a search term, and Google’s help page says that to see everything marked this way you “search Gmail for is:important” (Gmail Help, checked September 6, 2026). It’s a reserved system label in the API, where IMPORTANT sits in the same table as INBOX and STARRED, recorded as one a program may apply by hand (Google for Developers, checked September 6, 2026). And it’s a layout.
That last one is what people mean by Priority Inbox. Google describes it as splitting “your inbox into multiple sections. You can choose which sections you want to show, including ‘Important and unread,’ ‘Starred,’ ‘Everything else,’ or a label that you have made” (Gmail Help, checked September 6, 2026).
Read that carefully and the first useful thing falls out. Priority Inbox and Important first run the same judgment, and Priority Inbox gives it more shelves. Whatever is true about the model is true about both layouts, about the marker, about is:important, and about anything you build on top of them.
The flag reaches past the inbox too. Desktop notifications offer “Important mail notifications on,” and Google notes that when you do that “you get importance markers in Gmail” (Gmail Help, checked September 6, 2026). On Android the same choice is called “High priority only.” Google’s caution there is that “high priority emails will override any other notifications settings for certain labels” (Gmail Help, checked September 6, 2026). So the flag that decides what sits at the top of your list also decides what buzzes your phone at 9:40 on a Saturday.
What the model is actually predicting
Google published a research note about this system in 2010, written by three of its engineers and titled “The Learning Behind Gmail Priority Inbox.” It’s old, and sixteen years of model changes sit between it and the Gmail you’re looking at. It’s also the only first-party technical description of the mechanism that exists. The signal list on Gmail’s current help page lines up with the paper’s feature families almost one for one, so read together, the two say the same thing.
Here’s the paper’s first sentence about the feature. “The Priority Inbox feature of Gmail ranks mail by the probability that the user will perform an action on that mail.” The precise version sits in the section defining the target. “Our goal is to predict the probability that the user will interact with the mail within T seconds of delivery, providing the rank of the mail” (Google Research, 2010).
Notice what’s being predicted. Whether you’ll touch this message, and whether you’ll touch it soon.
The features that go into that prediction fall into four families, in the paper’s own words. “Social features are based on the degree of interaction between sender and recipient, e.g. the percentage of a sender’s mail that is read by the recipient. Content features attempt to identify headers and recent terms that are highly correlated with the recipient acting (or not) on the mail, e.g. the presence of a recent term in the subject.” Thread features “note the user’s interaction with the thread so far.” Label features “examine the labels that the user applies to mail using filters” (Google Research, 2010).
Gmail’s help page for importance markers names today’s signals (Gmail Help, checked September 6, 2026):
- “Whom you email, and how often you email them”
- “Which emails you open”
- “Which emails you reply to”
- “Keywords that are in emails you usually read”
- “Which emails you star, archive, or delete”
Put the two lists side by side. Every item on both of them is a fact about behavior, mostly yours and some the sender’s. Each one describes how mail has been treated before. What the message is asking you to do sits outside all of them.
That’s the whole shape of the thing, and everything below follows from it.
Interesting is not the same as important
The clearest statement of the limit is Google’s own. It appears in the part of the paper about where to put the cutoff between important and not. “Opening a mail is a strong signal of importance for our metric, but many users open a lot of mail that is ‘interesting’ rather than ‘important’” (Google Research, 2010).
For somebody running a small firm out of one mailbox, that sentence describes a specific and familiar failure.
The trade newsletter you read with coffee every Tuesday gets promoted, because you open it every Tuesday. The practice-management tool that emails you whenever anything happens gets promoted, because you click through to the portal most times. The client who sends you six short messages a day gets promoted too. Your open rate with them is nearly perfect, and that percentage is the model’s strongest social feature.
Meanwhile the message from the person who only writes when something has gone wrong sits lower. You’ve historically taken a day to open those. The model is reading your behavior correctly there, and your behavior is what it was asked to predict.
This is what people mean when they say Gmail marks the wrong things important, and they’re reading it right. The feature is working as built. Engagement and obligation line up most of the time, which is why the feature works at all, and they come apart precisely where the stakes are highest. Mattering and being due are two different properties, and one flag can hold only one of them: important versus urgent.
The message with no history behind it
Social features measure “the degree of interaction between sender and recipient.” Thread features note your interaction with the thread so far. Both of those measure a relationship.
A first email from an address you’ve never seen arrives with no relationship behind it, and a brand new thread arrives with no interaction so far. What’s left for the model is the content features and the global model that sits under every user’s personal one, which is to say general patterns learned across everybody. That’s a reasonable basis for judging a newsletter. It’s a thin one for judging a new client inquiry, a first note from opposing counsel, or a message from the accountant your client just switched to.
The paper adds a second gap, quieter and easy to miss. Training data is only collected inside a window it calls Tmax, which is “measured in days,” and the note is blunt about the consequence, that “users with interaction periods greater than Tmax will not generate training data” (Google Research, 2010). Read that as an owner. The mail you take a week to get to teaches the model nothing at all. Slow, deliberate, expensive correspondence is exactly the part of your inbox where the learning signal is weakest.
There’s a related surprise in the escape hatch people reach for. A filter that pins a known sender catches each message that matches it, and a reply is its own case. Google’s own caution is that “when someone replies to a message you’ve filtered, the reply will only be filtered if it meets the same search criteria” (Gmail Help, checked September 6, 2026).
What correcting a marker actually changes
The reasonable reaction to a wrong marker is to fix it, and Google invites exactly that. Its help page says that changing the marker “also helps Gmail learn which emails you think are important,” and that you can “hover over the importance marker” to see why a message was marked at all (Gmail Help, checked September 6, 2026). That transparency is genuinely more than most ranked feeds offer.
Here’s what the correction does, precisely, because it’s less than it sounds.
Two things happen. Your marking is fed back as training data, weighted more heavily than a passive signal. And it moves your personal cutoff. The paper describes a “per user threshold” for classifying each message, chosen this way because “it is difficult to algorithmically determine the threshold that will make a user happy.” It adds: “When a user marks messages in a consistent direction, we perform a real-time increment to their threshold” (Google Research, 2010).
A threshold is a volume control. Marking consistently in one direction changes how much mail clears the bar. What the bar measures stays exactly where it was. The score is a prediction of engagement, so a stricter cutoff gets you less predicted engagement. Months of patient correcting will make the promoted section smaller, and it stays a section about what you click.
Two levers do have a hard effect, and they’re worth naming because they’re the honest ceiling of what tuning can do.
The first is a rule. Google’s own page on categories notes that for specific senders you can set up a filter that “marks their emails as important” (Gmail Help, checked September 6, 2026). That’s deterministic, with no learning involved, and it works. Its limit is the limit every rule has. It covers the senders you thought of in advance, which is the opposite of the case that hurts you.
The second is the off switch. Gmail’s Inbox settings carry an “Importance markers” section with the options “Don’t use my past actions to predict which messages are important” and “No markers” (Gmail Help, checked September 6, 2026). Google notes the setting “can’t be changed from the Gmail app, but the settings you choose on your computer will apply to your app too.” Worth knowing if the promotions have become actively misleading, since a ranking you’ve stopped believing is worse than a list in plain arrival order.
How often it is wrong, in Google’s own numbers
Three figures come from the 2010 note. All of them were measured on the system as it stood then, and all of them are scored against Google’s own implicit definition of importance rather than yours. Treat them as indicative rather than current. They’re also the only published numbers anyone has.
Accuracy was around 80 percent. In the paper’s words: “Based on our implicit importance definition, our accuracy (tp + tn)/messages is approximately 80±5% on a control group” (Google Research, 2010). Read that phrase exactly. It’s how often the model correctly predicted whether you would interact with a message, not how often it was right about what mattered. The paper is careful about the difference.
The errors point one way. Because of how the cutoff is tuned, the note reports a false negative rate running three to four times the false positive rate (Google Research, 2010). The system leans toward leaving things out. For most inboxes that’s the right trade, since a promoted section full of junk is worth nothing. For a reader whose real fear is the message they found on Thursday that arrived on Monday, the errors point the wrong way.
Personalization helps, and then stops helping. The numbers come from a set of 160,000 messages users had corrected by hand. The global model alone was wrong 45 percent of the time. Personal models cut that to 38 percent, and personal models with personalized thresholds cut it to 31 percent (Google Research, 2010). That set is by construction the mail the model had already got wrong, so those rates describe a hard sample rather than an ordinary inbox. The shape is worth having anyway. Personalizing the model took roughly a third off the error, and roughly a third of it stayed where it was.
What this layout costs you elsewhere
Choosing Priority Inbox is a trade, and three parts of the trade are documented.
You give up the category tabs. Gmail sorts arriving mail into Primary, Social, Promotions, Updates and Forums only “when you use the ‘Default’ inbox type,” and Google’s instruction for switching categories off is to change your inbox type (Gmail Help, checked September 6, 2026). So you’re trading one thing for another. The tabs need no learning at all, and they reliably strip out bulk and marketing mail. The ranking depends entirely on learning. If your inbox’s problem is volume rather than stakes, that’s often the wrong way round. Which layout suits which problem is worked through in how to organize a busy Gmail inbox.
The section doing the work is “Important and unread.” Both halves of that name are conditions. Reading a message takes it out of the section that was holding it, whether or not you did anything about it. So the single action that most strongly tells the model a message was important is also the action that clears it from view. A message you opened at 8:30 and meant to come back to is no longer anywhere special.
The layout that could carry your own definition stays on the computer. Multiple Inboxes builds extra sections out of your own search operators, which is the one place in Gmail where you get to say what needs you. Google lists it among the inbox layouts on a computer. The layouts Google lists in the Gmail app on Android and on iPhone and iPad are Default Inbox, Important first, Unread first, Starred first and Priority Inbox (Gmail Help, checked September 6, 2026). Priority Inbox comes with you to the phone. Multiple Inboxes stays where you set it up.
Where Priority Inbox earns its place
The feature works, it costs nothing, and it asks nothing of you. That deserves saying plainly on a page about its limits.
Google measured this on its own staff. Averaging over employees receiving similar volumes of mail, about 2,000 Priority Inbox users “spent 6% less time reading mail overall, and 13% less time reading unimportant mail,” and were “more confident to bulk archive or delete email” (Google Research, 2010). Those are Google employees in 2010, self-selected into the feature, which is about as favorable a sample as exists. The direction is still real, and it matches what people report. A ranking that’s right most of the time beats arrival order, which is right by accident.
It’s the right answer when most of your mail comes from people you already write to regularly. It’s the right answer when the problem is how long the list is rather than what one miss would cost. And it’s the right answer when you’d rather have something imperfect running for free than maintain a system you have to keep true.
It’s the wrong answer when the misses cluster on senders you have no history with. It’s the wrong answer when a single message going unseen for three days costs real money or a client. And it’s the wrong answer when the reason you went looking for a fix was one specific email you found late.
The question the model was never asked
Worth stating in one place, because it’s the whole of it.
Every signal Google publishes, in the 2010 note and on the help page today, is a fact about behavior. Who you write to. What you open. What you reply to. What you star, archive or delete. The words that tend to appear in mail you read. Those are real signals, and a model built on them will get most of an ordinary week roughly right.
What they reach is your habits. The state of an obligation sits outside them. Is there an ask in this message. Is there a date attached to it. Have I already answered it. Is somebody waiting on me right now. Did the deadline in the fourth paragraph move by a week. Is this warm, chatty note from a client I hear from constantly the one that needs a signature before Friday.
Those are facts about what a message contains and where a commitment stands, and they’re readable. They’re a different question from the one a ranking trained on clicks was built to answer. A model of your habits is a good approximation of your obligations right up until the week it matters that they’re different things.
When the ranking reads the message
Point is an AI email client that sits on top of the mailbox you already run. Nothing moves. The address is the one you have, the archive stays where you left it, and there’s no migration step.
The difference to this page’s subject is what the ranking is built from. Point weighs each message for importance and for urgency at the moment it arrives, so the thing you can’t afford to miss is at the top rather than the thing you usually click. Every thread arrives with a plain summary attached, so you know what it wants before you open it. A request buried three paragraphs into a long message comes out as a dated task. And a first message from an address with no history behind it is judged on what it says, which is the case where a model of your habits has the least to go on.
Where Priority Inbox lets you nudge a threshold, Point takes the instruction in words. You can say what to watch for and be told when it arrives. You can give more weight to the part of your life that needs you this month and let the rest wait. What Point has learned about your habits is kept as individual notes you can read and edit, rather than as a hidden weight. That’s the part that’s genuinely hard to do with a marker you can only agree or disagree with one message at a time. Two pages go further into what that feels like day to day, only what needs you, at the top for the ranking and nothing falls through the cracks for the catching. Everything else Point does is listed on the benefits page.
That judgment travels out to the mailbox as real Gmail labels, so clearing a thread in one place shows it archived in the other, and nothing you labeled yourself is touched. How that write-back behaves is how Point works as your Gmail label manager.
How much Point does on its own is set per kind of action. Every kind carries its own level, from suggest-only up through review to fully handled, and review is where they all begin. The work gets prepared, then stops. A type you’ve satisfied yourself about can be raised on its own, and dropped back down later. Actions are logged in order, and undo reaches back across that log. The exception applies to every email product ever made, and it’s a message already sitting on somebody else’s server.
The limits, said straight. Point leaves Gmail’s importance markers exactly as they are, has no opinion about your filters, and makes no difference to how Priority Inbox performs. Point replaces the ranking rather than tuning it. That’s worth doing when the ranking is your problem, and it’s worth nothing when something else is. The narrower comparison is Point vs Gmail alone.
Common questions
How does Gmail decide which emails are important?
By predicting your behavior. Google’s published description of Priority Inbox says it “ranks mail by the probability that the user will perform an action on that mail.” Gmail’s current help page names the signals. The people you write to and how often. The mail you open. The mail you answer. The words that recur in what you read. What you star, archive or delete. Every one of those is a fact about how you’ve treated mail before. What a message is asking you to do, and when it’s due, sits outside the list.
Why does Gmail mark unimportant emails as important?
Because opening something is the model’s strongest signal, and you open plenty of mail that’s interesting rather than important. Google’s own research note says exactly that: opening a mail is a strong signal for its metric, “but many users open a lot of mail that is ‘interesting’ rather than ‘important’.” So the newsletter you read every week gets promoted on the strength of your click history. So does the notification service you always click through. That history is the behavior the model was asked to predict.
Why does Gmail keep missing my important emails?
Usually because the message came from somebody you have no history with. The strongest features in the model measure the degree of interaction between a sender and you, and a first email from a new address has none to measure. Google’s note also reports that the system’s false negative rate runs three to four times its false positive rate. That’s a system tuned to leave things out rather than let them in. Both effects push the same way, on exactly the mail you most want caught.
Can I train Gmail’s importance markers?
You can shift them, within limits. Correcting a marker feeds the model. Per Google’s research note, it also triggers a real-time increment to your personal threshold when you mark consistently in one direction. That threshold is a volume control. It changes how much mail clears the bar, and what the bar measures stays put. For a hard result on a sender you already know about, a filter that marks their mail important is deterministic where the model is probabilistic. There’s also a setting that stops Gmail using your past actions to predict importance at all, and one that hides the markers. Both are changed from a browser rather than from the Gmail app.
Is Priority Inbox better than Important first?
They run on the same judgment, so the two are equally accurate. Important first splits the inbox into two sections. Priority Inbox splits it into several, and lets you choose which ones appear, including “Important and unread,” “Starred,” “Everything else,” or a label you made. Choose between them on how many shelves you want. Both inherit what the importance model knows and what it misses.
Does using Priority Inbox mean losing the category tabs?
Yes. Gmail sorts mail into Primary, Social, Promotions, Updates and Forums only under the Default inbox type, and Google’s own instruction for switching categories off is to change your inbox type. That makes the choice a genuine trade. The tabs strip out bulk and marketing mail on their own, without depending on anything the model has learned about you, while Priority Inbox depends on it entirely. If your inbox’s real problem is volume, the tabs may be doing more for you than the ranking is.
Does Priority Inbox work in the Gmail app?
Yes, on both Android and iPhone. The layout the app leaves out is Multiple Inboxes, which Google lists among the inbox layouts on a computer. That’s the one worth flagging, because Multiple Inboxes is where you define sections with your own search operators. The prediction travels to your phone, and your own definition of what needs you stays on the computer. Importance also reaches your notifications there, through the “High priority only” setting, which Google warns will override other per-label notification settings.
The short version
Priority Inbox is a layout built on a single yes-or-no flag. That same flag drives Important first, the yellow marker, is:important and your notification settings. Understand the flag and you understand all of them.
The flag is a prediction about you. Google’s own published description ranks mail by how likely you are to act on it soon. Every signal on both Google’s 2010 note and today’s help page is a measure of behavior. Who you write to, what you open, what you reply to, what you star or archive. That prediction is right most of the time, and where it goes wrong it goes wrong in a pattern. Mail you always open floats up whether or not it matters. Senders you have never met sink, because the relationship starts today. The correspondence you take a week to answer barely trains the model at all.
Correcting the markers moves your threshold, which changes how much gets promoted rather than what promotion means. A filter can hard-wire a sender you already know about. Both leave the gap where it is, because the gap is a question problem rather than an accuracy problem.
The model was asked about your clicks. Whether there’s an ask in this message, whether a date is attached to it, whether somebody is waiting on you right now: those questions sit outside it. They’re also the facts that decide what a day costs you, and a good model of your habits is a different thing from them.