diff --git a/aw-transform/src/merge.rs b/aw-transform/src/merge.rs
index b46c4e3f..f04fbdb5 100644
--- a/aw-transform/src/merge.rs
+++ b/aw-transform/src/merge.rs
@@ -6,7 +6,13 @@ use aw_models::Event;
 ///
 /// Doesn't care about if events are neighbouring or not, this transform merges
 /// all events with the same key.
-/// The timestamp will be the timestamp of the first event with a specific key value
+/// Each merged event keeps the timestamp and the whole data of the first event
+/// with those key values, and the durations are summed. Events missing any of
+/// the keys are dropped, and an empty key list returns no events. The merged
+/// events come out in the order their first event appeared in the input, so
+/// the result is deterministic. That is not necessarily timestamp order: sort
+/// by timestamp before passing the result to union_no_overlap, which expects
+/// chronological input.
 ///
 /// # Example 1
 /// A simple example only using one key
@@ -22,7 +28,6 @@ use aw_models::Event;
 /// output:
 ///   { duration: 3.0, data: { "a": 1 } }
 ///   { duration: 1.0, data: { "a": 2 } }
-///   { duration: 1.0, data: { "b": 1 } }
 /// ```
 ///
 /// # Example 2
@@ -43,7 +48,10 @@ pub fn merge_events_by_keys(events: Vec<Event>, keys: Vec<String>) -> Vec<Event>
     if keys.is_empty() {
         return vec![];
     }
-    let mut merged_events_map: HashMap<String, Event> = HashMap::new();
+    // Index into `merged` by key, so the output keeps first-seen order
+    // instead of HashMap iteration order, which differs between runs.
+    let mut index: HashMap<String, usize> = HashMap::new();
+    let mut merged: Vec<Event> = Vec::new();
     'event: for mut event in events {
         let mut key_values = Vec::new();
         for key in &keys {
@@ -53,17 +61,18 @@ pub fn merge_events_by_keys(events: Vec<Event>, keys: Vec<String>) -> Vec<Event>
             }
         }
         let summed_key = key_values.join(".");
-        match merged_events_map.entry(summed_key) {
-            std::collections::hash_map::Entry::Occupied(mut entry) => {
-                entry.get_mut().duration += event.duration;
+        match index.entry(summed_key) {
+            std::collections::hash_map::Entry::Occupied(entry) => {
+                merged[*entry.get()].duration += event.duration;
             }
             std::collections::hash_map::Entry::Vacant(entry) => {
                 event.id = None;
-                entry.insert(event);
+                entry.insert(merged.len());
+                merged.push(event);
             }
         }
     }
-    merged_events_map.into_values().collect()
+    merged
 }
 
 #[cfg(test)]
@@ -150,4 +159,32 @@ mod tests {
         ];
         assert_eq!(&res2, &expected);
     }
+
+    #[test]
+    fn merge_keeps_first_seen_order() {
+        // HashMap iteration order would vary between runs; the output must
+        // follow the order in which each group first appears.
+        let mk = |secs: &str, app: &str| Event {
+            id: None,
+            timestamp: DateTime::from_str(secs).unwrap(),
+            duration: Duration::seconds(1),
+            data: json_map! {"app": json!(app)},
+        };
+        let apps = ["m", "c", "x", "a", "q", "b", "z", "d", "k", "e"];
+        let mut events = Vec::new();
+        for (i, app) in apps.iter().enumerate() {
+            events.push(mk(&format!("2000-01-01T00:00:{:02}Z", i), app));
+        }
+        // Repeats merge into the first occurrence without reordering
+        events.push(mk("2000-01-01T00:00:30Z", "a"));
+        events.push(mk("2000-01-01T00:00:31Z", "m"));
+        let res = merge_events_by_keys(events, vec!["app".to_string()]);
+        let order: Vec<&str> = res
+            .iter()
+            .map(|e| e.data.get("app").unwrap().as_str().unwrap())
+            .collect();
+        assert_eq!(order, apps);
+        assert_eq!(res[0].duration, Duration::seconds(2));
+        assert_eq!(res[3].duration, Duration::seconds(2));
+    }
 }