Who was there

Play-level audience

One row per ad play, with the audience the camera read while it ran: faces, head pose, gaze, facing angle, dwell, emotion, ambient sound and light, each with the number of camera readings behind it. Every logged play ships, so the file is also a complete play log.

Request a sample All datasets

The question it answers

Who was in front of the screen while my ad was playing?

182,303

plays with a camera reading

165,787

plays with a face on camera

4,610

screens

Measured over the 28 days to October 11, 2026.

The dataset

Grain: One row per ad play.
Cadence: One row per play, one file per day.
History: From June 24, 2026.
Format: Parquet or JSONL.
Delivered to: Your S3 bucket, your GCS bucket, or a time-limited download link, with a manifest and a schema.
Method: How each measure is produced and how to add rows up: the measurement methodology.

What buyers do with it

Attention and audience per creative, flight, screen and hour.
Send your own play logs (screen, start, end) and get the audience of each play back as a file.
Ground truth for attention and outcome models.

The table

Read from the export registry each time this site is built.

`play_audience`

Daily, from June 24, 2026

Grain: One row per play_date, screen_id, play_source and play_key.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 1.

Column	Type	What it holds
`play_date`	date	UTC day of the play (of its kept report's completion).
`screen_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`play_source`	string	Where the play is logged: completed_impressions, or partner_measurement_plays for a play a partner served.
`play_key`	string	With play_date, screen_id and play_source it is the play: billing's partner-aware key (completed_impressions.play_key) per UTC day, or partner_id:play_id for a partner play.
`ad_id`	string	Ad id of the kept report. Null: No ad id, or the pod placeholder.
`creative_id`	string	Creative id of the kept report. Null: No creative id, or the IMA house/fallback creative.
`impression_id`	string	Impression id of the kept report. Null: The report carried none.
`request_id`	string	Our ad request id (joins vast_requests). Null: The report carried none.
`content_sha256`	string	SHA-256 of the media file's bytes that played. Null: The report carried none.
`ingestion_path`	string	The kept report's path (vast_event, proof_of_play, partner_sdk, owned_direct; partner_measurement for a partner play).
`source_provider`	string	Demand source of the play.
`play_started_at`	timestamp	play_completed_at minus the play's length (its longest positive reported duration, else 30 s), UTC.
`play_completed_at`	timestamp	When the kept report says the play completed, UTC.
`duration_imputed`	integer	1 when no report carried a positive duration, so the play's interval is the 30 s default.
`windows_n`	integer	Audience windows attributed to the play: every window overlapping it widened by 5 s each side, any sensing mode and grade. Unit: windows.
`overlap_windows_n`	integer	Of those, the windows inside the logged play itself. Unit: windows.
`overlap_ms_sum`	integer	Milliseconds of window inside the logged play, summed. Unit: ms.
`camera_windows_n`	integer	Overlapping windows that are camera-graded: the base of every camera measure. Unit: windows.
`camera_overlap_ms`	integer	Overlap of those windows: the weight of face_count_mean and head_pose_index_mean. Unit: ms.
`face_windows_n`	integer	Camera windows whose face reading is above 0. Unit: windows.
`scene_frames_n`	integer	Attributed windows carrying a vision-model scene description. Unit: windows.
`evidence_grade_windows`	map	Attributed windows by evidence grade ('' = none recorded); sums to windows_n. Unit: windows.
`face_count_max`	double	Largest per-window face reading over the camera windows. Unit: faces. Null: No camera window.
`face_count_mean`	double	Overlap-weighted mean face reading over the camera windows. Unit: faces_per_window. Null: No camera window.
`head_pose_index_mean`	double	Overlap-weighted mean head-pose index (0.4 yaw + 0.3 pitch + 0.3 eye-open) over every camera window; gaze is gaze_pct_on_screen. Unit: index. Null: No camera window.
`head_pose_windows_n`	integer	Camera windows whose head-pose index is above 0 (its coverage, not its denominator). Unit: windows.
`gaze_windows_n`	integer	Camera windows carrying a gaze reading. Unit: windows.
`gaze_overlap_ms`	integer	Overlap of those windows: gaze_pct_on_screen's weight. Unit: ms.
`gaze_pct_on_screen`	double	Overlap-weighted mean on-screen gaze share 0-1. Unit: share. Null: No gaze window.
`head_yaw_windows_n`	integer	Camera windows whose facing angle came from the head-yaw instrument and is above 0. Unit: windows.
`head_yaw_overlap_ms`	integer	Overlap of those windows. Unit: ms.
`head_yaw_heads_up_share`	double	Overlap-weighted share of those windows facing the screen within 30 degrees. Never pooled with the shoulder instrument. Unit: share. Null: No such window.
`shoulder_windows_n`	integer	Camera windows whose facing angle came from the shoulder (MoveNet) instrument and is above 0. Unit: windows.
`shoulder_overlap_ms`	integer	Overlap of those windows. Unit: ms.
`shoulder_heads_up_share`	double	Overlap-weighted share of those windows within 30 degrees. Unit: share. Null: No such window.
`dwell_windows_n`	integer	Camera windows with a capped dwell (a start time to cap at). Unit: windows.
`dwell_overlap_ms`	integer	Overlap of those windows. Unit: ms.
`dwell_capped_ms_mean`	double	Overlap-weighted mean dwell, each window capped at its own length. Unit: ms. Null: No such window.
`receptivity_windows_n`	integer	Camera windows carrying an ad-receptivity score. Unit: windows.
`receptivity_overlap_ms`	integer	Overlap of those windows. Unit: ms.
`receptivity_mean`	double	Overlap-weighted mean ad-receptivity score 0-1. Unit: index. Null: None carried one.
`emotional_engagement_windows_n`	integer	Camera windows carrying an emotional engagement score. Unit: windows.
`emotional_engagement_overlap_ms`	integer	Overlap of those windows. Unit: ms.
`emotional_engagement_mean`	double	Overlap-weighted mean emotional engagement score 0-1. Unit: index. Null: None carried one.
`emotion_windows_n`	integer	Camera windows with a real dominant-emotion reading. Unit: windows.
`dominant_emotion_overlap_ms`	map	Overlap milliseconds per dominant-emotion value: the play's emotion distribution. Unit: ms.
`dominant_emotion`	string	The emotion with the most overlap, ties to the alphabetically first. Null: No emotion reading.
`audio_windows_n`	integer	Overlapping windows carrying an audio level. Unit: windows.
`audio_db_max`	double	Loudest audio level over those windows. Unit: dB. Null: No audio reading.
`lux_windows_n`	integer	Overlapping windows with an ambient light reading. Unit: windows.
`lux_overlap_ms`	integer	Overlap of those windows. Unit: ms.
`ambient_lux_mean`	double	Overlap-weighted mean ambient light. Unit: lux. Null: No light reading.
`face_rows_n`	integer	Per-face observation rows whose span overlaps the logged play. Unit: face_rows.
`face_tracks_n`	integer	Distinct face tracks over those rows (the edge's pre-merge track id where it sent one); distinct counts do not add across plays. Unit: face_tracks. Null: No face row.
`gaze_face_rows_n`	integer	Face rows carrying gaze seconds. Unit: face_rows.
`gaze_seconds`	double	Gaze seconds of those rows inside the logged play. Unit: seconds. Null: No such row.
`builder_version`	string	Digest of the builder SQL that built the row (pa-<12 hex>).

Request a sample

Tell us the markets and the weeks you want. The sample is cut from production by the same export that makes the full delivery.

Request a sample

All datasets · Next: Mobility flows, daily