Editor’s Draft, 19 March 2026
More details about this documentCopyright © 2026 World Wide Web Consortium. W3C® liability, trademark and permissive document license rules apply.
This document defines an API that provides web page authors with insights into the latency of certain events triggered by user interactions.
This is a public copy of the editors’ draft. It is provided for discussion only and may change at any moment. Its publication here does not imply endorsement of its contents by W3C. Don’t cite this document other than as work in progress.
GitHub Issues are preferred for discussion of this specification.
This document is governed by the 18 August 2025 W3C Process Document.
This section is non-normative.
When a user engages with a website, they expect their actions to cause changes to the website quickly. In fact, research suggests that any user input that is not handled within 100ms is considered slow. Therefore, it is important to surface performance timing information about input events that could not achieve those guidelines.
A common way to monitor event latency consists of registering an event listener. The timestamp at which the event was created can be obtained via the event’s timeStamp. In addition, performance.now() could be called both at the beginning and at the end of the event handler logic. By subtracting the hardware timestamp from the timestamp obtained at the beginning of the event handler, the developer can compute the input delay: the time it takes for an input to start being processed. By subtracting the timestamp obtained at the beginning of the event handler from the timestamp obtained at the end of the event handler, the developer can compute the amount of synchronous work performed in the event handler. Finally, when inputs are handled synchronously, the duration from event hardware timestamp to the next paint after the event is handled is a useful user experience metric.
This approach has several fundamental flaws. First, requiring event listeners precludes measuring event latency very early in the page load because listeners might not be registered yet. Second, developers who are only interested in input delay might be forced to add new listeners to events that originally did not have them. This adds unnecessary performance overhead to the event latency calculation. And lastly, it would be very hard to measure asynchronous work caused by the event via this approach.
This specification provides an alternative to event latency monitoring that solves some of these problems. Since the user agent computes the timestamps, there is no need for event listeners in order to measure performance. This means that even events that occur very early in the page load can be captured. This also enables visibility into slow events without requiring analytics providers to attempt to patch and subscribe to every conceivable event. In addition to this, the website’s performance will not suffer from the overhead of unneeded event listeners. Finally, this specification allows developers to obtain detailed information about the timing of the rendering that occurs right after the event has been processed. This can be useful to measure the overhead of website modifications that are triggered by events.
This section is non-normative.
A single user Interaction (sometimes called a Gesture) is typically made up of multiple physical hardware input events. Each physical input event might cause the User Agent to dispatch several UIEvents, and each of those might trigger multiple custom event listeners, or trigger distinct default actions.
For example, a single user "tap" interaction with a touchscreen device is actually made up of a sequence of physical input events:
a touch start,
a tiny amount of touch movement,
a touch end.
Those physical input events might dispatch a series of UIEvents:
...and potentially some Focus events, Input events, etc, in between.
These individual UIEvents will each become candidates for their own PerformanceEventTiming entry reporting, which is useful for detailed timing.
Note: pointermove and touchmove are not currently considered for Event Timing.
However, this specification also defines a mechanism for grouping related PerformanceEventTimings into Interactions via an interactionId. This mechanism can be used to define a page responsiveness metric called Interaction to Next Paint (INP).
This section is non-normative.
The very first user Interaction typically has a disproportionate impact on user experience, and is also often disproportionately slow.
To that effect, the Event Timing API exposes timing information about the first input of a Window, defined as the first PerformanceEventTiming entry with a non-0 interactionId.
Unlike most PerformanceEventTimings, the first input entry is reported even if it does not exceed a provided durationThreshold, and is buffered even if it does not exceed the default duration threshold of 104ms. This mechanism can be used to define a page responsiveness metric called First Input Delay (FID).
This also allows developers to better measure percentiles and performance improvements, by including data even from pages which are always very responsive, without having to register event handlers.
The Event Timing API exposes timing information only for certain events.
If event’s isTrusted attribute value is set to false, return false.
If event’s type is one of the following: auxclick, click, contextmenu, dblclick, mousedown, mouseenter, mouseleave, mouseout, mouseover, mouseup, pointerover, pointerenter, pointerdown, pointerup, pointercancel, pointerout, pointerleave, gotpointercapture, lostpointercapture, touchstart, touchend, touchcancel, keydown, keypress, keyup, beforeinput, input, compositionstart, compositionupdate, compositionend, dragstart, dragend, dragenter, dragleave, dragover, drop, return true.
Return false.
Note: mousemove, pointermove, pointerrawupdate, touchmove, wheel, and drag are excluded because these are "continuous" events. The current API does not have enough guidance on how to count and aggregate these events to obtain meaningful performance metrics based on entries. Therefore, these event types are not exposed.
This section is non-normative. It explains at a high level the information that is exposed in the § 3 Processing model section.
Event timing information is only exposed for certain events, and only when the time difference between user input and paint operations that follow input processing exceeds a certain duration threshold.
The Event Timing API exposes a duration value, which is meant to be the time from when the physical user input occurs (estimated via the Event’s timeStamp) to the next time the rendering of the Event’s relevant global object’s associated Document is updated. This value is provided with 8 millisecond granularity.
By default, the Event Timing API buffers and exposes entries when the duration is 104 or greater, but a developer can set up a PerformanceObserver to observe future entries with a different threshold. Note that this does not change the entries that are buffered and hence the buffered flag only enables receiving past entries with duration greater than or equal to the default threshold.
An Event’s delay is the difference between the time when the browser is about to run event handlers for the event and the Event’s timeStamp. The former point in time is exposed as the PerformanceEventTiming’s processingStart, whereas the latter is exposed as PerformanceEventTiming’s startTime. Therefore, an Event’s delay can be computed as processingStart - startTime.
Note that the Event Timing API creates entries for events regardless of whether they have any event listeners. In particular, the first click or the first key might not be the user actually trying to interact with the page functionality; many users do things like select text while they’re reading or click in blank areas to control what has focus. This is a design choice to capture problems with pages which register their event listeners too late and to capture performance of inputs that are meaningful despite not having event listeners, such as hover effects. Developers can choose to ignore such entries by ignoring those with essentially zero values of processingEnd - processingStart, as processingEnd is the time when the event dispatch algorithm algorithm has concluded.
The following example computes a dictionary mapping interactionId to the maximum duration of any of its events. This dictionary can later be aggregated and reported to analytics.
let maxDurations = {}; new PerformanceObserver(list => { for (let entry of list.getEntries()) { if (entry.interactionId > 0) { let id = entry.interactionId; if (!maxDurations[id]) { maxDurations[id] = entry.duration; } else { maxDurations[id] = Math.max(maxDurations[id], entry.duration); } } } }).observe({ type: 'event', buffered: true, durationThreshold: 16 });The following are sample use cases that could be achieved by using this API:
Gather first input delay data on a website and track its performance over time.
Clicking a button changes the sorting order on a table. Measure how long it takes from the click until we display reordered content.
A user drags a slider to control volume. Measure the latency to drag the slider.
Hovering a menu item triggers a flyout menu. Measure the latency for the flyout to appear.
Measure the 75’th percentile of the latency of the first user click (whenever click happens to be the first user interaction).
Event Timing adds the following interfaces:
A PerformanceEventTiming object reports timing information about one associated Event.
Each PerformanceEventTiming object has these associated concepts, all of which are initially set to null:
An eventTarget containing the associated Node.
The target attribute’s getter must perform the following steps:
If this’s eventTarget is not exposed for paint timing given null, return null.
Return this’s eventTarget.
Note: A user agent implementing the Event Timing API would need to include "first-input" and "event" in supportedEntryTypes for Window contexts. This allows developers to detect support for event timing.
This remainder of this section is non-normative. The values of the attributes of PerformanceEventTiming are set in the processing model in § 3 Processing model. This section provides an informative summary of how they will be set.
PerformanceEventTiming extends the following attributes of the PerformanceEntry interface:
name The name attribute’s getter provides the associated event’s type. entryType The entryType attribute’s getter returns "event" (for long events) or "first-input" (for the first user interaction). startTime The startTime attribute’s getter returns the associated event’s timeStamp. duration The duration attribute’s getter returns the difference between the next time the update the rendering steps are completed for the associated event’s Document after the associated event has been dispatched, and the startTime, rounded to the nearest 8ms.PerformanceEventTiming has the following additional attributes:
processingStart The processingStart attribute’s getter returns a timestamp captured at the beginning of the event dispatch algorithm. This is when event handlers are about to be executed. processingEnd The processingEnd attribute’s getter returns a timestamp captured at the end of the event dispatch algorithm. This is when event handlers have finished executing. It’s equal to processingStart when there are no such event handlers. cancelable The cancelable attribute’s getter returns the associated event’s cancelable attribute value. target The target attribute’s getter returns the associated event’s last target when such Node is not disconnected nor in the shadow DOM. targetSelector The targetSelector attribute’s getter returns a string that identifies the associated event’s last target. interactionId The interactionId attribute’s getter returns a number that uniquely identifies the user Interaction which triggered the associated event. This attribute is 0 unless the associated event’s type attribute value is one of:A pointerdown, pointerup, or click, and belongs to a tap or drag gesture. Note that pointerdown that ends in scroll is excluded.
The EventCounts object is a map where the keys are event types and the values are the number of events that have been dispatched that are of that type. Only events whose type is supported by PerformanceEventTiming entries (see section § 1.4 Events exposed) are counted via this map.
The eventCounts attribute’s getter returns this’s relevant global object’s eventCounts.
The interactionCount attribute’s getter returns this’s relevant global object’s interactionCount.
This section will be removed once [DOM] has been modified.
Right after step 1, we add the following steps:
Let interactionId be the result of computing interactionId given event.
Let timingEntry be the result of initializing event timing given event, the current high resolution time, and interactionId.
Right before the returning step of that algorithm, add the following step:
Finalize event timing passing timingEntry, event, target, and the current high resolution time as inputs.
Note: If a user agent skips the event dispatch algorithm, it can still choose to include an entry for that Event. In this case, it will estimate the value of processingStart and set the processingEnd to the same value.
This section will be removed once [HTML] has been modified.
Each Window has the following associated concepts:
entries to be queued, a list that stores PerformanceEventTiming objects, which will initially be empty.
has dispatched input event, a boolean which is initially set to false.
user interaction value, an integer which is initially set to a random integer between 100 and 10000.
Note: The user interaction value is set to a random integer instead of 0 so that developers do not rely on it to count the number of interactions in the page. By starting at a random value, developers are less likely to use it as the source of truth for the number of interactions that have occurred in the page.
pending key downs, a map of integers to PerformanceEventTimings which is initially empty.
pointer interaction value map, a map of integers which is initially empty.
pending pointer downs, a map of integers to PerformanceEventTimings which is initially empty.
is contextmenu triggered, a boolean which is initially set to false.
eventCounts, a map with entries of the form type → numEvents. This means that there have been numEvents dispatched such that their type attribute value is equal to type. Upon construction of a Performance object whose relevant global object is a Window, its eventCounts must be initialized to a map containing 0s for all event types that the user agent supports from the list described in § 1.4 Events exposed.
interactionCount, an integer which counts the total number of distinct user interactions, for which there was a unique interactionId computed via computing interactionId.
For each fully active Document in docs, invoke the algorithm to dispatch pending Event Timing entries for that Document.
This section will be removed once [PERFORMANCE-TIMELINE-2] had been modified.
The PerformanceObserverInit dictionary is augmented:
partial dictionary PerformanceObserverInit { DOMHighResTimeStamp durationThreshold; };Note: The following algorithm is used in the [PERFORMANCE-TIMELINE-2] specification to determine when a PerformanceEventTiming entry needs to be added to the buffer of a PerformanceObserver or to the performance timeline, as described in the registry.
If entry’s entryType attribute value equals to "first-input", return true.
Assert that entry’s entryType attribute value equals "event".
Let minDuration be computed as follows:
If options is not present or if options’s durationThreshold is not present, let minDuration be 104.
Otherwise, let minDuration be the maximum between 16 and options’s durationThreshold value.
If entry’s duration attribute value is greater than or equal to minDuration, return true.
Otherwise, return false.
Increase window’s user interaction value value by a small number chosen by the user agent.
Let interactionCount be window’s interactionCount.
Set interactionCount to interactionCount + 1.
Note: The user interaction value is increased by a small number chosen by the user agent instead of 1 to discourage developers from considering it as a counter of the number of user interactions that have occurred in the web application. This allows the user agent to choose to eagerly assign a user interaction value (i.e. at pointerdown) and then discard it (i.e. after pointercancel), rather than to lazily compute it.
A user agent may choose to increase it by a small random integer every time, or choose a constant. A user agent must not use a shared global user interaction values for all Windows, because this could introduce cross-origin leaks.
If event’s isTrusted attribute value is false, return 0.
Let type be event’s type attribute value.
If type is not one among keyup, compositionstart, input, pointercancel, pointerup, click, or contextmenu, return 0.
Note: keydown and pointerdown are marked pending in finalize event timing, and then updated later when computing interactionId for future events (like keyup and pointerup).
Let window be event’s relevant global object.
Let pendingKeyDowns be window’s pending key downs.
Let pointerMap be window’s pointer interaction value map.
Let pendingPointerDowns be window’s pending pointer downs.
If type is keyup:
If event’s isComposing attribute value is true, return 0.
Let code be event’s keyCode attribute value.
If pendingKeyDowns[code] does not exist, return 0.
Let entry be pendingKeyDowns[code].
Increase interaction count on window.
Let interactionId be window’s user interaction value value.
Set entry’s interactionId to interactionId.
Add entry to window’s entries to be queued.
Remove pendingKeyDowns[code].
Return interactionId.
If type is compositionstart:
For each entry in the values of pendingKeyDowns:
Append entry to window’s entries to be queued.
Clear pendingKeyDowns.
Return 0.
If type is input:
If event is not an instance of InputEvent, return 0. Note: This check is done to exclude Events for which the type is input but that are not about modified text content.
If event’s isComposing attribute value is false, return 0.
Increase interaction count on window.
Return window’s user interaction value.
Otherwise (type is pointercancel, pointerup, click, or contextmenu):
Let pointerId be event’s pointerId attribute value.
If type is click:
If pointerMap[pointerId] does not exist, return 0.
Let value be pointerMap[pointerId].
Remove pointerMap[pointerId].
Return value.
Assert that type is pointerup, pointercancel, or contextmenu.
If pendingPointerDowns[pointerId] does not exist:
If type is contextmenu, return window’s user interaction value.
If type is pointerup and window’s is contextmenu triggered flag is true:
Set window’s is contextmenu triggered flag to false.
Return window’s user interaction value.
Otherwise, return 0.
Let pointerDownEntry be pendingPointerDowns[pointerId].
Assert that pointerDownEntry is a PerformanceEventTiming entry.
If type is pointerup or contextmenu:
Increase interaction count on window.
Set pointerMap[pointerId] to window’s user interaction value.
Set pointerDownEntry’s interactionId to pointerMap[pointerId].
Append pointerDownEntry to window’s entries to be queued.
Remove pendingPointerDowns[pointerId].
If type is contextmenu, set window’s is contextmenu triggered to true.
If type is pointercancel, return 0.
Return pointerMap[pointerId].
Note: The algorithm attempts to assign events to the corresponding interaction IDs. For keyboard events, a keydown triggers a new interaction ID, whereas a keyup has to match its ID with a previous keydown. For pointer events, when we get a pointerdown we have to wait until pointercancel or pointerup occur to know its interactionId. We try to match click with a previous interaction ID from a pointerdown. If pointercancel or pointerup happens, we’ll be ready to set the interactionId for the stored entry corresponding to pointerdown. If it is pointercancel, this means we do not want to assign a new interaction ID to the pointerdown. If it is pointerup, we compute a new interaction ID and set it on both the pointerdown and the pointerup (and later, the click if it occurs).
If the algorithm to determine if event should be considered for Event Timing returns false, then return null.
Let timingEntry be a new PerformanceEventTiming object with event’s relevant realm.
Set timingEntry’s entryType to "event".
Set timingEntry’s startTime to event’s timeStamp attribute value.
Set timingEntry’s processingStart to processingStart.
Set timingEntry’s cancelable to event’s cancelable attribute value.
Set timingEntry’s interactionId to interactionId.
Return timingEntry.
If timingEntry is null, then return.
Let relevantGlobal be target’s relevant global object.
Set timingEntry’s processingEnd to processingEnd.
Assert that target implements Node.
Note: This assertion holds due to the types of events supported by the Event Timing API.
Set timingEntry’s eventTarget to target.
Note: This will set eventTarget to the last event target. So if retargeting occurs, the last target, closest to the root, will be used.
Set timingEntry’s targetSelector to the result of running the algorithm to generate a CSS selector with target as input.
If event’s type attribute value is pointerdown:
Let pendingPointerDowns be relevantGlobal’s pending pointer downs.
Let pointerId be event’s pointerId.
If pendingPointerDowns[pointerId] exists:
Let previousPointerDownEntry be pendingPointerDowns[pointerId].
Add previousPointerDownEntry to relevantGlobal’s entries to be queued.
Set pendingPointerDowns[pointerId] to timingEntry.
Set relevantGlobal’s is contextmenu triggered to false.
Otherwise, if event’s type attribute value is keydown:
If event’s isComposing attribute value is true:
Append timingEntry to relevantGlobal’s entries to be queued.
Return.
Let pendingKeyDowns be relevantGlobal’s pending key downs.
Let code be event’s keyCode attribute value.
If pendingKeyDowns[code] exists:
Let previousKeyDownEntry be pendingKeyDowns[code].
If code is not 229:
Increase relevantGlobal’s user interaction value value by a small number chosen by the user agent.
Set previousKeyDownEntry’s interactionId to relevantGlobal’s user interaction value.
Note: 229 is a special case since it corresponds to IME keyboard events. Sometimes multiple of these are sent by the user agent, and they do not correspond to holding a key down repeatedly.
Add previousKeyDownEntry to relevantGlobal’s entries to be queued.
Set pendingKeyDowns[code] to timingEntry.
Otherwise:
Append timingEntry to relevantGlobal’s entries to be queued.
Let window be doc’s relevant global object.
Let renderingTimestamp be the current high resolution time.
For each timingEntry in window’s entries to be queued:
Set event timing entry duration passing timingEntry, window, and renderingTimestamp.
If timingEntry’s duration attribute value is greater than or equal to 16, then queue timingEntry.
Set window’s entries to be queued to an empty list.
For each pendingPointerDownEntry in the values from window’s pending pointer downs:
Set event timing entry duration passing pendingPointerDownEntry, window, and renderingTimestamp.
For each pendingKeyDownEntry in the values from window’s pending key downs:
Set event timing entry duration passing pendingKeyDownEntry, window, and renderingTimestamp.
If timingEntry’s duration attribute value is nonzero, return.
Let start be timingEntry’s startTime attribute value.
Set timingEntry’s duration to a DOMHighResTimeStamp resulting from renderingTimestamp - start, with granularity of 8ms or less.
Let name be timingEntry’s name attribute value.
Perform the following steps to update the event counts:
Let eventCounts be window’s eventCounts.
Assert that eventCounts contains name.
Set eventCounts[name] to eventCounts[name] + 1.
If window’s has dispatched input event is false, and timingEntry’s interactionId is not 0, run the following steps:
Let firstInputEntry be a copy of timingEntry.
Set firstInputEntry’s entryType to "first-input".
queue firstInputEntry.
Set window’s has dispatched input event to true.
If target is a Node, run the following steps:
Let selector be a string with an initial value of target’s nodeName.
If target is an Element, run the following steps:
If target has an `id` attribute, set selector to the concatenation of « selector, "#", the value of the `id` attribute ».
Otherwise, if target has a `src` attribute, set selector to the concatenation of « selector, "[src=", the value of the `src` attribute, "]" ».
Return selector.
Otherwise, return an empty string.
We would not like to introduce more high resolution timers to the web platform due to the security concerns entailed by such timers. Event handler timestamps have the same accuracy as performance.now(). Since processingStart and processingEnd could be computed without using this API, exposing these attributes does not produce new attack surfaces. Thus, duration is the only one which requires further consideration.
The duration has an 8 millisecond granularity (it is computed as such by performing rounding). Thus, a high resolution timer cannot be produced from these timestamps. However, it does introduce new information that is not readily available to web developers: the time pixels draw after an event has been processed. We do not find security or privacy concerns with exposing the timestamp, especially given its granularity. In an effort to expose the minimal amount of new information that is useful, we decided to pick 8 milliseconds as the granularity. This allows relatively precise timing even for 120Hz displays.
The choice of 104ms as the default cutoff value for the duration is just the first multiple of 8 greater than 100ms. An event whose rounded duration is greater than or equal to 104ms will have its pre-rounded duration greater than or equal to 100ms. Such events are not handled within 100ms and will likely negatively impact user experience.
The choice of 16ms as the minimum value allowed for durationThreshold is because it enables the typical use-case of making sure that the response is smooth. In 120Hz displays, a response that skips more than a single frame will be at least 16ms, so the entry corresponding to this user input will be surfaced in the API under the minimum value.
Conformance requirements are expressed with a combination of descriptive assertions and RFC 2119 terminology. The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in the normative parts of this document are to be interpreted as described in RFC 2119. However, for readability, these words do not appear in all uppercase letters in this specification.
All of the text of this specification is normative except sections explicitly marked as non-normative, examples, and notes. [RFC2119]
Examples in this specification are introduced with the words “for example” or are set apart from the normative text with class="example", like this:
Informative notes begin with the word “Note” and are set apart from the normative text with class="note", like this:
Note, this is an informative note.
Requirements phrased in the imperative as part of algorithms (such as "strip any leading space characters" or "return false and abort these steps") are to be interpreted with the meaning of the key word ("must", "should", "may", etc) used in introducing the algorithm.
Conformance requirements phrased as algorithms or specific steps can be implemented in any manner, so long as the end result is equivalent. In particular, the algorithms defined in this specification are intended to be easy to understand and are not intended to be performant. Implementers are encouraged to optimize.
PerformanceEventTiming/cancelable
PerformanceEventTiming/interactionId
In only one current engine.
PerformanceEventTiming/processingEnd
PerformanceEventTiming/processingStart