Everything underneath

Raw event streams

The tables every product here is built from, one row per event: ad requests and auctions, completed plays, playback and bid-chain events, camera windows, attention and on-device inferences, radio exposure, screen-to-screen transitions, IP geography, and the audience packages and panel cells we curate. Same keys, same delivery and the same manifest as the products.

Request a sample All datasets

The question it answers

Can I have the events themselves and build my own tables?

raw tables

122.6 million

demographic readings

50.6 million

hourly movement rows

Measured over the 28 days to October 11, 2026.

The dataset

Grain: One row per event.
Cadence: Event rows, one file per day. The curation tables ship as current snapshots.
History: From February 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

Build your own aggregates at any grain.
Train models on the full event history.
Reconcile every product row back to the events that made it.

The tables

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

`ip_geo_observations`

Daily, from June 2026

The hashed public IP each screen connected from, per day, with ISP, organisation, city, state, country, connection type and whether it is residential.

Grain: One row per ip_hash, screen_mongo_id and observed_date.
Delivery: One file per day, as Parquet or JSONL.
Basis: Declared by the source.
Definition: Version 2.

Column	Type	What it holds
`observed_date`	date	UTC day the screen's traffic left from the address.
`observed_at`	timestamp	Latest time that day the screen was seen on the address, UTC.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`ip_hash`	string	Hash of the public IP address the screen egressed from (the screen's network, never a visitor's); the address itself is not shipped.
`geohash_6`	string	Geohash-6 cell (about 1.2 x 0.6 km) of the address's IP geolocation. Null: The IP lookup returned no location.
`city`	string	City of the address's IP geolocation. Null: The IP lookup returned no city.
`state`	string	State or region of the address's IP geolocation. Null: The IP lookup returned no state.
`country`	string	Country of the address's IP geolocation. Null: The IP lookup returned no country.
`isp`	string	Internet service provider of the address. Null: The IP lookup named no provider.
`org`	string	Organization the address is registered to. Null: The IP lookup named no organization.
`connection_type`	string	Line type from the IP lookup: residential, commercial, mobile or unknown. Null: The IP lookup did not classify the line.
`is_residential`	integer	1 when the IP lookup classifies the address as residential, else 0.

`observation_demographics`

One row per event, from July 16, 2026

Every attribute reading the cameras make: screen, time, attribute and value.

Grain: One row per observation_id, dimension and value.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 2.

Column	Type	What it holds
`observed_date`	date	UTC day of the reading.
`observed_at`	timestamp	When the camera observation the reading comes from was made, UTC.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`observation_id`	string	The camera observation (one analysed frame window) the reading belongs to.
`dimension`	string	The audience attribute read (config/observationDemographicsDimensions.js): age, gender, mood, device_engagement, activity_state, intent, attire, lifestyle, carried_items and the others declared. head_turn and occupant_position are image-frame values: head_turn is turned_image_left, turned_image_right, not_turned or not_assessable (where the nose points relative to the edges of the picture), occupant_position is image_left_third, image_center_third, image_right_third or driving (the third of the picture the person occupies). Neither is a side of the vehicle or the street: a rearward cabin camera mirrors the vehicle's left and right.
`value`	string	The value read for the dimension. lifestyle values are numeric IAB Audience Taxonomy 1.1 segment ids (segtax=4); the other dimensions use our own vocabulary.
`count`	integer	Readings of the value in the observation: 1 on every row (one row is one reading). Unit: readings.

`proximity_graph_edges_hourly`

Hourly, from July 15, 2026

Screen-to-screen movement by hour and radio signal, anchored on ad plays.

Grain: One row per hour_bucket, source_screen_id, destination_screen_id and source.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 2.

Column	Type	What it holds
`hour_bucket`	timestamp	UTC hour the identifiers were seen at the source screen during an ad play.
`source_screen_id`	string	Screen where the identifiers were seen during the play. Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`destination_screen_id`	string	Screen where the same identifiers were seen within the next 24 hours.
`source`	string	Radio signal source of the identifiers: ble, mdns, ssdp or http_probe.
`repeated_signal_count`	integer	Pairs of (source-screen signal, destination-screen signal) behind the row, summed over its ad-window cells. Unit: signal_pairs.
`unique_identifiers_floor`	integer	Distinct identifiers in the largest single (creative, ad) cell of the hour: the least the hour's distinct count can be. Unit: identifiers.
`identifier_cell_sum`	integer	Sum of the cells' distinct identifiers: the most the hour's distinct count can be (an identifier seen in two cells counts twice here). Unit: identifiers.
`ad_window_cells`	integer	The (creative, ad) cells the row aggregates. Unit: cells.
`median_lag_minutes`	double	Mean over the cells of each cell's median minutes from the source sighting to the destination sighting. Unit: minutes.

`auction_results`

One row per event, from March 2026

Every auction and outbound bid request: floor, bids received, winning bid and clearing price, winning DSP and creative, result and latency.

Grain: One row per auction_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 2.

Column	Type	What it holds
`auction_id`	string	The auction's id: one auction per bid request we ran.
`bid_request_id`	string	OpenRTB bid request id the auction answered. Null: Not recorded on the auction.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`timestamp`	timestamp	When the auction ran, UTC.
`auction_type`	string	Auction mechanics: first_price, or outbound_openrtb_proxy for an auction we forwarded to a demand partner.
`floor_price_cpm`	double	Floor price of the auction, USD CPM. Unit: usd_per_1000_plays. Null: No floor was set.
`total_bids_received`	integer	Bids the auction received. Unit: bids.
`winning_bid_id`	string	Id of the winning bid. Null: No bid won.
`winning_bid_price_cpm`	double	Price the winning bid offered, USD CPM. Unit: usd_per_1000_plays. Null: No bid won.
`winning_clearing_price_cpm`	double	Price the winning bid cleared at, USD CPM. Unit: usd_per_1000_plays. Null: No bid won, or no clearing price was recorded.
`winning_dsp_name`	string	Demand platform of the winning bid. Null: No bid won.
`winning_creative_id`	string	Creative id of the winning bid. Null: No bid won, or the bid named no creative.
`result`	string	Outcome: bid_received, no_bid, timeout or error.
`latency_ms`	integer	Auction duration in milliseconds. Unit: ms. Null: Not recorded.
`distributor_id`	string	Reseller (distributor) id the traffic is attributed to; empty for Trillboards' own.

`completed_impressions`

One row per event, from April 2, 2026

Every completed play as billing counts it: creative, ad, seconds played, CPM, gross revenue, invalid-traffic result and the path that reported it.

Grain: One row per play_key and screen_mongo_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 3.

Column	Type	What it holds
`play_key`	string	The play's identity as our billing policy keys it: the impression id, else the bid request, else the request; the request for a request-level provider and for an SDK callback without a real impression id. With screen_mongo_id it is the key.
`completed_at`	timestamp	When the kept report says the play completed, UTC.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`ad_id`	string	Ad id of the play. Null: No ad id, or the pod-build placeholder ('ima_pending') of an ad whose id arrives later.
`creative_id`	string	Creative id the demand source supplied for the ad. Null: No creative id, or the IMA house/fallback creative ('-1' or '').
`impression_id`	string	Impression id the report carried. Null: The report carried no impression id.
`source_provider`	string	Demand source that served the ad.
`variant_name`	string	Serving variant (the ad-serving configuration arm) of the request or play. Null: Not recorded.
`actual_cpm`	double	CPM the play was priced at, USD; pricing_source says how the price was set. Unit: usd_per_1000_plays. Null: No price was recorded.
`actual_gross_revenue`	double	Gross revenue of the play in USD; pricing_source says how it was set. Unit: usd. Null: No price was recorded.
`pricing_source`	string	How the price was set: openrtb_clearing_cpm, vast_pricing, a _reporting_api value (the price a partner's reporting API reported), placement_purchase and its variants (a direct booking), cpm_estimate or config_default (a configured estimate, not a settled price), or a _zero value for a play that earns nothing.
`currency`	string	Currency of the price.
`ingestion_path`	string	Which report of the play this is: vast_event, proof_of_play, partner_sdk or owned_direct. The kept report is the one billing keeps.
`request_id`	string	Our ad request id (joins vast_requests.request_id). Null: The report carried no request id.
`bid_request_id`	string	OpenRTB bid request id (joins auction_results.bid_request_id). Null: The report carried no bid request id.
`play_id`	string	The player's own id for the play. Null: The report carried no play id.
`duration_seconds`	double	Play length the kept report carried. ad_delivery_hourly.played_seconds uses the longest positive duration over every report of the play. Unit: seconds. Null: The kept report carried no duration (a vast_event report never does).
`givt_passed`	boolean	True when the play passed general invalid-traffic checks.
`display_mode`	string	How the ad was displayed (fullscreen on every row of 2026-08..10).
`display_area_ratio`	double	Share of the screen the ad occupied, 0-1. Unit: share.
`created_at`	timestamp	When we stored the report, UTC.
`impression_multiplier`	double	OpenRTB 2.6 impression multiplier applied to the play's revenue (1 on every row to date).
`signature`	string	Proof-of-play signature over the play's impression, ad, screen, device, time and event (Ed25519, or HMAC on the older version), verifiable against our published key. Null: The report carried no signature.
`content_sha256`	string	SHA-256 of the bytes of the media file that played: the creative's content identity. Null: The report carried no content hash.

`event_tracking`

One row per event, from February 2026

The playback funnel of every ad: start, first quartile, midpoint, third quartile and complete.

Grain: One row per event_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 1.

Column	Type	What it holds
`event_id`	string	The event's id.
`event_type`	string	Player event: impression, start, firstQuartile, midpoint, thirdQuartile, complete, heartbeat or error.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id. Null: The event named no screen.
`ad_mongo_id`	string	Ad id the event belongs to. Null: The event named no ad.
`impression_id`	string	Impression id the event belongs to. Null: The event named no impression.
`creative_id`	string	Creative id the demand source supplied for the ad. '-1' is the IMA house/fallback creative. Null: The event named no creative.
`occurred_at`	timestamp	When the event happened, UTC.
`created_at`	timestamp	When we stored the event, UTC.
`distributor_id`	string	Reseller (distributor) id the traffic is attributed to; empty for Trillboards' own.
`variant_name`	string	Serving variant (the ad-serving configuration arm) of the request or play.

`impression_events`

One row per event, from April 15, 2026

The OpenRTB bid chain behind each impression: bid request, ad, creative, start time and status.

Grain: One row per impression_event_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 2.

Column	Type	What it holds
`impression_event_id`	string	The impression event's id.
`event_id`	string	Id of the player event that opened the impression.
`bid_request_id`	string	OpenRTB bid request id of the impression. Null: Not recorded.
`impression_id`	string	Impression id.
`ad_id`	string	Ad id of the impression. Null: No ad id, or the pod-build placeholder ('ima_pending') of an ad whose id arrives later.
`creative_id`	string	Creative id the demand source supplied for the ad. Null: No creative id, or the IMA house/fallback creative ('-1' or '').
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`timestamp`	timestamp	When the impression was recorded, UTC.
`start_time`	timestamp	When the ad started, UTC.
`status`	string	Impression status (started on every row of 2026-09-01..07). Null: Not recorded.
`distributor_id`	string	Reseller (distributor) id the traffic is attributed to; empty for Trillboards' own.

`omsdk_sessions`

One row per event, from July 15, 2026

Grain: One row per session_id and created_at.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 1.

Column	Type	What it holds
`session_id`	string	The Open Measurement SDK session's id.
`created_at`	timestamp	When the session opened, UTC. With session_id it is the key.
`screen_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id. Null: The session named no screen (16 of 291.7M rows).
`ad_id`	string	Ad id the session opened for.
`impression_id`	string	Impression id the session opened for. Null: The session named no impression (1 row).
`event_count`	integer	OM SDK events the session recorded: 1 on every row to date, the open. Unit: events.
`status`	string	Session status: active on every row to date (no session has recorded a close).
`ad_session_type`	string	OM ad session type: video on every row to date.
`omid_api_version`	string	OMID API version of the session (1.4).
`partner_name`	string	OM integration partner that opened the session: Trillboards, our own integration.
`partner_version`	string	Version of that integration (1.0.0).

`vast_requests`

One row per event, from February 2026

Every ad request, one row per waterfall attempt: outcome, latency, error codes, device, demand source, advertiser, clearing CPM and SDK.

Grain: One row per request_id and requested_at.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 3.

Column	Type	What it holds
`request_id`	string	Our ad request id. With requested_at it is the key; to count requests keep the latest requested_at per request_id.
`auction_request_id`	string	Id of the auction request the ad request triggered. Null: The request ran no auction.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id. Partner-exchange placements carry a non-screen id here.
`creative_id`	string	Creative id the demand source supplied for the ad. Null: No creative id, or the IMA house/fallback creative ('-1' or '').
`variant_name`	string	Serving variant (the ad-serving configuration arm) of the request or play.
`outcome`	string	Latest outcome of the request: requested, bid_received, no_bid, no_fill, filled, redirect, player_served, completed, skipped, timeout or error. Null: Not recorded.
`completed`	boolean	True once the ad completed (a latch: it never reverts).
`requested_at`	timestamp	When the request was made, UTC (a partner-supplied request id is re-stamped at each later update).
`latency_ms`	integer	Time to answer the request, milliseconds. Unit: ms. Null: Not recorded.
`ima_error_code`	integer	Google IMA SDK error code. Null: No IMA error.
`vast_error_code`	integer	VAST error code. Null: No VAST error.
`device_category`	string	Device family of the player (e.g. Amazon Fire TV, Android box, Samsung tablet). Null: Not classified.
`device_environment`	string	Player environment the request declared (e.g. ctv). Null: Not classified.
`device_type_hint`	string	OpenRTB device type the player declared. Null: Not sent.
`source_name`	string	Demand source the request was sent to. Null: Not recorded.
`request_source`	string	How the request was built: vast_wrap, server_constructed, proxy_auction, client_lifecycle or partner_sdk. Null: Not recorded.
`waterfall_position`	integer	Position of the demand source in the screen's waterfall. Null: Not in a waterfall.
`venue_type`	string	Venue category of the screen at request time. Null: Not recorded.
`advertiser_name`	string	Advertiser the ad response named. Null: The response named no advertiser.
`clearing_cpm`	double	Clearing price the response carried, USD CPM. Unit: usd_per_1000_plays. Null: No price was returned.
`content_type`	string	MIME type of the media file returned (empty when none).
`ad_system`	string	VAST AdSystem of the response (empty when none).
`buyer_seat`	string	Buyer seat of the winning bid (empty when none).
`currency`	string	Currency of the price (empty when none).
`distributor_id`	string	Reseller (distributor) id the traffic is attributed to; empty for Trillboards' own.
`sdk_type`	string	Ad SDK of the player (e.g. google-ima). Null: Not sent.
`sdk_version`	string	Version of that SDK. Null: Not sent.
`platform_hint`	string	Operating system the player declared. Null: Not sent.
`vast_type`	string	Shape of the VAST response: inline, inline_resolved, wrapper_passthrough or empty. Null: Not classified.
`openrtb_bid_count`	integer	Bids the OpenRTB auction returned. Unit: bids. Null: The request was not an OpenRTB auction.
`openrtb_seat_count`	integer	Seats that bid in the OpenRTB auction. Unit: seats. Null: The request was not an OpenRTB auction.
`creative_landing_url`	string	Landing page (click-through) of the creative served. Null: The response named no landing page.
`media_url`	string	URL of the creative's media file. Null: The response named no media file.
`schain`	string	OpenRTB supply chain object (JSON) of the request. Null: The request carried no supply chain.
`upstream_extensions`	string	VAST extensions the upstream response carried (JSON). Null: The upstream response carried no extensions.
`experiment_assignments`	string	Serving experiment arms the request was assigned (JSON). Null: The request was in no serving experiment.
`request_signature_hash`	string	SHA-256 of the request's host, path and query parameters (correlator and request id excluded): one value per request shape. Null: Not computed (an older server-built request).

`device_exposure_hourly`

Hourly, from May 24, 2026

Radio exposure per screen, hour and signal: distinct, returning and cross-screen devices, dwell and attention seconds.

Grain: One row per hour, screen_mongo_id and source.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 3.

Column	Type	What it holds
`hour`	timestamp	UTC hour the 6-minute exposure slots of the row started.
`screen_mongo_id`	string	Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`source`	string	Radio signal source: ble, wifi, mdns, ssdp or http_probe.
`visitor_source`	integer	1 when the source is one the canonical visitor count uses (ble, mdns, ssdp, http_probe), 0 for wifi. Visitor headline counts sum the rows with 1.
`n_windows`	integer	Exposure windows: one per device key, screen, source and 6-minute slot. Unit: exposure_windows.
`unique_devices`	integer	Distinct device keys (the canonical visitor id) seen at the screen in the hour on the source. Unit: identifiers.
`returning_devices`	integer	Of those, the keys seen at this screen on an earlier day (30-day look-back). Unit: identifiers.
`cross_screen_devices`	integer	Of those, the keys that also appear at another screen (cross_screen_transitions). Unit: identifiers.
`total_dwell_seconds`	integer	Sum over the windows of each window's dwell (first to last sighting in the slot). Unit: seconds.
`attention_windows`	integer	Windows whose slot carried a camera attention value. Through 2026-09-23 a slot no camera window covered carried none; from 2026-09-24 the windower writes 0 for such a slot, so every window counts. Unit: exposure_windows.
`total_attention_seconds`	integer	Sum over those windows of the slot's camera attention seconds (the slot's high or medium attention camera windows times their length); every window of a slot carries the slot's value. From 2026-09-24 a slot no camera covered counts 0. Unit: seconds. Null: No window of the row carried a camera attention value (through 2026-09-23: no camera window covered its slots).
`p50_dwell_seconds`	integer	Exact median dwell of the windows. Unit: seconds.
`p90_dwell_seconds`	integer	Exact 90th-percentile dwell of the windows. Unit: seconds.

`cross_screen_edges_daily`

Daily, from May 2026

Devices seen at one screen and then at another within the hour, by screen pair and day, with the median lag and distance.

Grain: One row per dt, source_screen_id and destination_screen_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 3.

Column	Type	What it holds
`dt`	date	UTC day of the transitions.
`source_screen_id`	string	Screen the device was seen at first. Trillboards screen id (24-character hex); joins dim_screen.screen_id.
`destination_screen_id`	string	Screen the device was seen at next, 1-60 minutes later.
`edge_count`	integer	Distinct transitions (device key, screens, time, lag) from the source to the destination that day. Unit: transitions.
`unique_devices`	integer	Distinct device keys behind the transitions (the RSSI-independent stable key). Unit: identifiers.
`lag_p50_minutes`	integer	Exact median minutes between the two sightings. Unit: minutes.
`transition_distance_p50_m`	double	Exact median great-circle distance between the two screens' stored coordinates. The transition writer stores 0 where either screen has no coordinate, so 0 means co-located or not located. Unit: meters.

`audience_metrics`

One row per event, from February 2026

Every 10-second camera window: faces, dwell, attention, mood, crowd density and movement pace, with the screen's location.

Grain: One row per metric_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 3.

`attention_ledger`

One row per event, from February 2026

Seconds of gaze read by the camera, by screen and time.

Grain: One row per id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 2.

`inference_outcomes`

One row per event, from February 2026

Every on-device model inference: model version, confidence and what it detected.

Grain: One row per inference_snapshot_id.
Delivery: One file per day, as Parquet or JSONL.
Basis: Measured.
Definition: Version 3.

`curated_inventory_packages`

Snapshot, from Current state

The audience packages we sell: segment, geography, screens and partner venues, share with its interval, index against national, peak daypart and floor CPM.

Grain: One row per package_key.
Delivery: One file per delivery, as JSONL.
Basis: Modelled.
Definition: Version 1.

Column	Type	What it holds
`package_key`	string	The package's id: one behavioural segment in one geography.
`segment_key`	string	The merchandised segment the package sells.
`geo_scope`	string	Geography of the package: national or a DMA.
`dma_code`	string	Nielsen DMA code of a DMA package. Null: A national package.
`dma_name`	string	DMA name. Null: A national package.
`display_name`	string	Buyer-facing package name.
`headline`	string	Buyer-facing headline. Null: No headline written.
`audience_line`	string	Buyer-facing audience description. Null: No audience line written.
`delivery_line`	string	Buyer-facing delivery description. Null: No delivery line written.
`venue_types`	array	Venue types the package selects. Null: Not restricted by venue type.
`dma_codes`	array	DMAs the package selects. Null: Not restricted by DMA.
`countries`	array	Countries the package selects. Null: Not restricted by country.
`owned_screens`	integer	Our screens in the package's footprint. Unit: screens.
`partner_venues`	integer	Partner venues in the footprint (supply we buy). Unit: venues.
`share`	double	Fitted share of the audience in the segment, 0-1. Unit: share. Null: The segment was not observed in the package's cells (a venue-led package).
`share_ci_low`	double	Lower bound of the share's Wilson interval. Unit: share. Null: No share.
`share_ci_high`	double	Upper bound of the share's Wilson interval. Unit: share. Null: No share.
`index_vs_national`	double	The share over the national share (1.0 = national). Unit: index. Null: No share.
`projection_tier`	string	Which panel rung the share was projected from. Null: No share.
`panel_support_n`	integer	Panel observations the share is computed over. Unit: observations. Null: No share.
`panel_match_n`	integer	Of those, the observations in the segment. Unit: observations. Null: No share.
`panel_screens`	integer	Panel screens behind the share. Unit: screens. Null: No share.
`peak_daypart`	string	Daypart where the segment concentrates most. Null: The package does not lead on a daypart.
`peak_daypart_index`	double	The peak daypart's share over the package's share. Unit: index. Null: No peak daypart.
`floor_cpm`	double	Derived floor price of the package, USD CPM. Unit: usd_per_1000_plays. Null: No floor derived.
`window_start`	date	First day of the panel window the share was fitted on.
`window_end`	date	Last day of that window.
`computed_at`	timestamp	When the package was computed, UTC (rebuilt whole each run).
`evidence_tier`	string	Strength of the residence evidence for the areas sold. Null: Not measured.
`evidence_home_block_groups`	integer	Census block groups with a home-location leg behind the areas sold. Unit: block_groups. Null: Not measured.
`evidence_three_leg_block_groups`	integer	Block groups with all three legs (camera, device, MAID). Unit: block_groups. Null: Not measured.
`evidence_three_leg_share`	double	Share of the block groups with all three legs. Unit: share. Null: Not measured.
`evidence_device_equivalents`	double	Device keys behind the residence evidence. Unit: identifiers. Null: Not measured.
`evidence_window_start`	date	First day of the evidence window. Null: Not measured.
`evidence_window_end`	date	Last day of the evidence window. Null: Not measured.

`panel_segment_cells`

Snapshot, from Current state

The fitted share of each audience segment by venue type, daypart and market, with its interval and support.

Grain: One row per segment_key, venue_type, daypart and market.
Delivery: One file per delivery, as JSONL.
Basis: Modelled.
Definition: Version 1.

Column	Type	What it holds
`segment_key`	string	The merchandised segment (visits-* segments come from the store-visit panel, the rest from the camera panel).
`venue_type`	string	Venue context of the cell (__all__ = every venue).
`daypart`	string	Daypart of the cell (__all__ = every daypart).
`market`	string	Market of the cell (__all__ = every market).
`share`	double	Fitted share of the cell in the segment, 0-1. Segments overlap, so shares do not sum to 1. A segment never observed in a cell has no row. Unit: share.
`ci_low`	double	Lower bound of the share's Wilson interval. Unit: share.
`ci_high`	double	Upper bound of the share's Wilson interval. Unit: share.
`match_n`	integer	Evaluable observations in the segment. Unit: observations.
`support_n`	integer	Evaluable observations of the cell: camera observations carrying a dimension the segment reads, or store-visit exposure edges by resolvable devices on the visits leg. Unit: observations.
`panel_screens`	integer	Panel screens behind the cell. Unit: screens.
`window_start`	date	First day of the panel window.
`window_end`	date	Last day of the panel window.
`computed_at`	timestamp	When the cell was computed, UTC.

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: Frame corpus for AI training