On a broadcaster’s streaming service, audience data usually lives in three places.
- The player telemetry knows, within a minute or two, how many people started an episode and where they stopped watching.
- The ad server knows which breaks it stitched in, how long they ran, which audience segments they went to, and what they earned.
- The content systems (the CMS or asset management system, the schedule, and the as-run log) know what the programme was: its genre, its cast, its break structure, and the promo that sent people to it.
Each system passes the others what it needs for its own job. What usually doesn’t exist is a joined record, at the level of the moment in the programme, that anyone outside those teams can query. So the minute-level signal that could explain a result sits in a store that the people making content decisions can’t reach.
On the linear side, this is old news. Audience research teams have read minute-by-minute overnights against the as-run log for decades, and Barb publishes those overnights the next morning, followed by consolidated figures that fold in seven days of playback, eight days after transmission (Barb Audiences Ltd, n.d.). In the US, ratings can be read at live plus three, seven, or even thirty-five days (Nielsen, 2025). That consolidated figure is the number the industry trades on, and it should stay. The question is what sits beside it, because on streaming and connected TV, where a growing share of viewing now happens, the platform’s own telemetry is faster and far more granular, and it tends to be split across systems that were never built to meet.
The wider series argued that operational telemetry is turning into one of the most valuable assets a business owns (see Telemetry Is Becoming the Business). Media is where that bites hardest, because for a streaming platform the telemetry is the closest thing it has to watching its audience directly.
The number of record
Here’s a composite example of how the gap plays out. A broadcaster launches a new drama on its streaming service, which carries ads. The first episode holds its audience well. The second shows a sharp drop at about the twentieth minute, mostly on connected TVs. The following week, “soft after episode two” goes into the notes, and the marketing push planned for the back half of the run gets trimmed.
What the headline number can’t show is why. The drop sat right on a mid-roll break, and the break itself sat at a natural pause, where the editors had put the cue point. What had changed was its length. Ad operations had lengthened connected-TV pods to lift yield, so on those screens the break ran nearly three minutes, against about ninety seconds everywhere else. The audience was leaving the break, and the show was fine.
A linear audience team would have spotted the equivalent on a broadcast channel the next morning. On streaming, the minute-level viewing data sat with video engineering, the pod lengths sat in the ad server, and the break map sat in the asset management system. Nobody was hiding anything. The three were simply never joined, so nobody could see it.
Many newsrooms have run live editorial dashboards tied to article data for years, and games live-ops teams have worked this way for about as long. Long-form video is where that joined, moment-level view is still rarer, and that join is what separates the three postures.
Telemetry as back-office reporting
Audience telemetry tends to sit in one of three postures, and most platforms run all three in some form. The useful question is which one carries the weight.
| Back-office reporting | Ad-tech input | First-class data product | |
|---|---|---|---|
| Owner | Video engineering, operations | Ad operations, ad-tech vendors | A named data product owner |
| Latency | Batch, daily or weekly | Real time | Tiered, seconds to next day |
| Grain | Sessions and devices, often sampled | Impressions and segments | Events tied to the moment in the programme |
| Content context | Episode ID, little else | Contextual signals for targeting | Scene, break, and promo |
| Consent | Rarely carried | Scoped to advertising | Carried on every record, by purpose |
| Who can use it | Operations | Ad sales and ad decisioning | Editorial, commercial, and product |
The oldest posture treats telemetry as an operational signal. Player and streaming analytics were first collected to keep the service healthy: start failures, buffering, bitrate shifts, and error rates by device and region. Video engineering and operations owned the data, it flowed in batches into a reporting warehouse, and it surfaced as dashboards and monthly reports. It was usually aggregated to sessions and devices, often sampled, and kept only as long as operations needed it.
For its purpose, that was a good design. It answered whether the service was working, and it answered well. It usually knows which episode was playing and where the playhead was. What it doesn’t know is what was happening at that point: the scene, the break, or the promo that ran before it. Most platforms still have this layer, and it still earns its keep. It just gives slow, blurred answers when the business starts asking it strategic questions.
Telemetry as ad-tech input
The second posture is where a lot of platforms sit today. Here the telemetry feeds ad decisioning and measurement: impressions, completion rates, frequency, and the audience segments that inventory is priced and sold against. It runs on real-time audience data and tends to be well funded, because it’s attached directly to revenue.
Its shape is set by the ad stack. Identity is keyed to advertising identifiers, segments, or, increasingly on broadcaster streaming services, registered users. The event schema is built around impressions, and a good deal of the ad-tech data sits inside vendor platforms under their retention and access rules. That’s the right design for selling and delivering advertising, and it does that job well. It was never meant to explain a programme to an editor, and in the drama example, a sensible yield decision cost the programme viewers without anyone on the content side seeing it happen.
This is also where privacy law does the most work, on both sides of the Atlantic. Under GDPR, personal data collected for one purpose can’t simply be reused for an incompatible one (European Parliament and Council, 2016, Article 5(1)(b)). In California, consumers can opt out of the sale or sharing of their personal information, and businesses have to honour opt-out preference signals such as Global Privacy Control (California Privacy Protection Agency, 2026). Ad-tech is actually good at carrying consent with each request. The catch is that the consent is scoped to advertising, and it’s often lost when the data lands in a warehouse for some other team to use.
Telemetry as a first-class data product
The third posture treats audience telemetry as a product in its own right: one governed audience data product that editorial, commercial, and product teams all work from.
In practice, viewing events are streamed in once and enriched at ingest with content metadata: the programme, the segment, the break structure, and the promo. They’re joined to ad delivery within minutes, then reconciled when the billing-grade figures land. Consent and permitted purpose travel on every record, so each team gets only what it’s allowed to use, and the privacy rules aren’t re-implemented in every downstream tool. Latency is tiered on purpose, with seconds for real-time personalisation, minutes for editorial dashboards, and next day for commissioning analysis. The consolidated currency and other broadcast data are loaded alongside the platform’s own figures, with definitions reconciled (people against devices, average audience against starts) so the two views can be compared instead of argued over.
Building this is a media data architecture problem, and it’s bigger than a dashboard. At the ingestion layer it’s a related shape of problem to the one Orbx, a simulation and games studio, had when its multi-cloud estate made a unified data platform impossible. Its fix was a pipeline that now streams live positional data into Bigtable at one-second precision (Orbx case study). The audience-specific parts, consent on every record and the join to content and ad delivery, get built on top of that kind of foundation, and that’s the kind of work Sakura’s Data & AI practice does with media businesses.
The ad stack doesn’t go away in this model. It keeps making ad decisions, and it becomes one more consumer of the same data product.
When the third posture carries the weight
Once the data product exists, content analytics gets a lot more specific. Editorial can see completion by episode and by segment, the minute where viewers leave, and which promos brought people in, all tied to what was on screen. Commissioning still reads the consolidated currency for the market view, and now it has the platform’s own data to help explain it.
The commercial conversation changes too. When ad sales and editorial look at the same numbers, the argument moves off whose figures are right and onto the real trade-off, such as whether a longer connected-TV pod is worth the viewers it loses. Product teams personalise from the same source, so the recommendation a viewer gets and the commissioning analysis agree about what that viewer watched.
One caution. Newsrooms learned years ago that a live chart can pull editors toward whatever is spiking, and a streaming service can make the same mistake with completion curves. The drama that lost viewers at minute twenty needed a shorter break. Better media analytics makes that kind of thing visible, and it still takes an editor to decide what to do about it.
Across this series, the same pattern kept turning up. Peak traffic that nobody scheduled (Launch Day Is Every Day), provenance a studio couldn’t produce when it was asked (Protecting the IP in a Generative World), and here, audience data a business collects but can’t put to work. In each case the fix was in how the systems underneath were built.
Running that audience data product day to day, and keeping its pipelines, consent rules, and latency tiers healthy as the platform changes, is the work Sakura’s Managed Services team does for media businesses once it’s built.
References
Barb Audiences Ltd, n.d. Frequently asked questions. Barb. Available at: https://www.barb.co.uk/frequently-asked-questions/ [Accessed 28 September 2026].
California Privacy Protection Agency, 2026. California Consumer Privacy Act Regulations, Cal. Code Regs. tit. 11, §7000 et seq. (implementing Cal. Civ. Code §1798.120), effective 1 January 2026. Available at: https://cppa.ca.gov/regulations/pdf/ccpa_statute_eff_20260101.pdf [Accessed 28 September 2026].
European Parliament and Council, 2016. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data (General Data Protection Regulation). Official Journal of the European Union, L 119, 4 May, pp. 1-88. Available at: https://eur-lex.europa.eu/eli/reg/2016/679/oj [Accessed 28 September 2026].
Nielsen, 2025. Need to Know: What are Nielsen ratings? Nielsen, November. Available at: https://www.nielsen.com/insights/2025/what-are-nielsen-ratings/ [Accessed 28 September 2026].