Most audience measurement systems reliant on ACR (automated content recognition) and many others return one-dimensional, point-in-time data (timestamp, panelist_identifier, content_or_channel). These must be processed into two-dimensional respondent-level records (timestamp_start, timestamp_end, panelist_identifier, content_or_channel). This is far from a straightforward process:

Timeshifting is the viewing of video programming or listening to audio programming after it has aired. It can be done through on-demand services, catch-up services integrated into a pay-TV EPG (electronic programme guide), a timeshifted channel, or a physical or cloud DVR. In most cases, the content is streamed to the TV set or other device through which it is viewed or listened to.

VOD (video on demand) or AOD (audio on demand) streaming has become the main method of timeshifting. A very large part of all programme content is available only through streaming after it airs on television, and many programmes never air on television but are released exclusively on streaming. Streaming services also carry many theatrical films and exclusive programming that they produce. Streaming services charge a subscription fee, are funded by advertising (which usually cannot be skipped), or both. Some streaming services have multiple tiers from mostly or entirely ad-funded ones to ad-free subscription-based ones.

Most television audience measurement services do not capture VOD use either through their panels (which are large enough to capture viewing of at least the most popular programming) or by integration with the VOD service operators (many of which resist independent measurement and refuse to share their server-side data). In cases in which the measurement company does have access to VOD service server-side data, fusing it to panel data to impute coviewing (viewers per viewing household, VPVH), personas and demographics presents a substantial challenge to which there are no broadly accepted solutions.

DVR (digital video recorder, sometimes termed PVR, personal video recorder) is a time-delay recording technology that records viewing, either by explicit decision of the viewer or in accordance with a set of rules (a single broadcast, a “season ticket” to record an entire series for “binge watching” at season’s end, explicit genre rules, or in some systems, individual “if you liked this, you’ll probably like that” suggestions). Largely replaced by on-demand viewing, DVRs (today usually implemented in the provider’s cloud) are still relevant because they add trickplay capability (pause, rewind, fast-forward, speed multiples) to both realtime and timeshifted viewing (they buffer 90&endash;120 minutes of the currently tuned channel to enable these functions in realtime). They permit catching up to a live broadcast without ads or boring segments, and fast-forwarding through, or on some models entirely skipping, advertising breaks.

DVR viewing is not present in RPD data unless the data collection system is deliberately designed to include it.

OTT (over-the-top) refers to content viewed bypassing conventional television, especially on devices other than big-screen TV sets (broadband-connected computers and smartphones (from which it can optionally be streamed to TV monitors using products like Google Chromecast, Amazon Fire TV or Roku), on smart TVs or on broadband-connected gaming platforms). Because streaming content can now be viewed regardless of device, the term “OTT” is falling out of use, subsumed by “VOD” or “streaming”.

Pay-TV operators fear being reduced to eventual irrelevance if exclusive sports content becomes unmoored from the need to subscribe to their own services, and although this threshold is far from having been universally reached, the trend is definitely in this direction.

At the turn of the millenium, advertisers often could reach all or most of their target audiences with a simple conventional TV (terrestrial and pay-TV) buy. This is no longer true. In most Western countries, people under age 50 still watch conventional TV but in much smaller numbers, having moved to on-demand viewing. This means advertisers must buy both conventional TV and streaming/VOD. They want to know how much a specific streaming/VOD buy extends their reach, and how it affects the count of impressions per viewer, which may be so high as to cause annoyance or so low as to be insufficient for the message to be effective. The advertisers can cap impressions per platform in programmatic buys, but they need this information to know how to use this capability. Thus "Deduplicated reach" (and incremental impressions) has become the holy grail of advertisers; in most Western countries, it remains elusive.

Conventional audience measurement providers struggle to measure streaming/VOD platforms: their technical equipment often cannot measure them, and even if it did, their panel sizes would be able to detect only the most popular programming on these platforms. Major streaming/VOD platforms such as Netflix and Amazon Prime Video have long tried to avoid independent measurement; they refused to sell access to their own and forced connected-TV manufacturers to agree not to measure them as the cost of being permitted to preinstall the platforms’ must-have apps. To some extent, these platforms still behave this way, but they are gradually realising that independent measurement is good for their ad sales.

But the use of more than one data source, as in any audience measurement system that integrates more than one census-level source (not just VOD but also connected TVs and pay-TV systems), or a census-level source and a panel, or two panels (as in The Netherlands), this raises another problem: how to remove duplicates from one of the two datasets (the panel and the server-side, census-level data from the platform).

There are multiple solutions to this problem, varying in complexity and both capital and operational cost. A simple match on the viewer’s current IP address is key, but in the samples of the large global TAM providers there will often be router-meters, which require professional installation and are labour-intensive to keep correctly configured, both on the LAN and the WAN sides.

The intab (sometimes termed the net daily reporting sample or usable sample) is the daily cooperating part of the sample, on which the ratings published for that day are based. This requires detecting measurement devices with data loss on that day and excluding them from the intab. If this is impossible (as it is when mobile devices are used for measurement, which are often unpowered part of the day), special algorithms must be used to ensure accurate reporting.

Cooperation, in the TAM meter and RPD contexts, means that:

Census-level data are viewing and other related measurement data collected by software on streaming services, connected TVs, pay-TV STBs (set-top boxes), digital terrestrial multiplex receivers/decoders (used in many countries outside the U.S.), or DVRs and sent by Internet to the operator. Usually, unlike in TAM, no opt-in or sample recruitment is needed as the data collection is authorised in the user’s or subscriber’s user agreement. Typical sample sizes are far larger than in classical TAM, often reaching six digits rather than four or five for TAM. The costs per device are much lower than in TAM and often trivial. Costs of processing the data can also be controlled much better than in a dedicated operation, although a more complex technological environment would result in higher costs.

However, census-level data are of devices or households rather than individual viewers; a viewing event does not disclose how many people viewed or indeed whether any at all viewed the TV set (or, usually, even whether the TV set was on). Nor are there demographics other than the socioeconomic class derived from the service address. To personalise the data, it must be fused with panel data.

Sample management is the selection and tracking of households, devices and individuals in the sample—the group of households or devices capable of returning measurement data or, less commonly, a selected subgroup of these.

This area is one in which classical TAM and census-level data providers’ (VOD/AOD services, pay-TV operators, connected-TV manufacturers, streaming dongle makers, DVR companies) practices differ greatly. TAM providers can afford only a small sample because of the high cost of recruiting and metering each household. Furthermore, because they quite consciously do not control their samples by income, education or occupation or even geography at a scale granular enough not to be overly diverse, but only by large geographic area, age, and gender (a practice that leaves a Boston corporate lawyer and a Chelsea, Massachusetts bus driver who never finished high school looking exactly identical), they overly rely on the randomness of the sample household selection as a panacea for those failings—which it is not, at those predetermined low sample sizes. The TAM provider will make several efforts to recruit the randomly selected household, and if it fails, will turn their attention to its nearest neighbour. Copious empirical evidence has shown that a different sample recruited in this manner to the same composition by the few tracked attributes, after a year to shake out early participants (which have an increased tendency to leave), can return completely different results, favouring different channels, than the preexisting sample, and that this process can be repeated endlessly with the same outcome.

The use of census-level data, on the other hand, is founded on the strategic advantage of much larger samples due to the default availability of measurement collection software on devices. Thus, census-level data sources usually accept all the sample they can get.

Like any audience measurement provider, we balance our panel sample using a scientifically accepted multivariate algorithm that avoids the need for the sample distribution to be exactly that of the subscriber universe as well as the operational difficulty of populating numerous microcells of various attributes (used by some very simple radio ratings services, for example, but not practical for multichannel television). Our implementation of such an algorithm runs daily. To include census-level data in balancing, it must be personalised and equipped with demographics through fusion.

Channels must be described adequately and their viewing schedules identified. In some pay-TV systems, this task is extremely complex.

Some TAM providers obligate channel operators to encode their channels and programmes especially for measurement, a time-consuming and costly process with which many don’t comply, becoming unmeasurable. Others use alternative means of automated identification, also with poor reliability.

RPD systems use the pay-TV operator’s channel identification, which can range from simple and reliable (when the channel lineup rarely changes and a stable, standard channel identifier such as the DVB triad is used) to incomplete and maddeningly complex (entirely lacking a stable identifier and other essential metadata, such as the channel and feed name). Furthermore, multiple instances of the same content channel appearing on the same or multiple included sources (such as different pay-TV systems or a given operator’s CATV and IPTV systems may vary.

We successfully provide channel management even in these most complex environments.

Schedule management is the processing of detailed descriptions of broadcasts at particular times in the schedules of channels, and their linking to the channels themselves. Although this task sounds simple, in practice it often isn’t. First, the schedules must be related to the channel and to its viewing (sometimes on different identifiers), so all the difficulties of channel management are present.

A second problem is identification of the series or special and the episode within a series; these are often inconsistent or incomplete, while programme and episode titles are even less standardised. Whether a sports game is live, a recorded première or a repeat dramatically affects its ratings but such designation is rarely in the schedule and then has to be supplied in processing. Genre data, and for movies, directors’ and actors’ names, all useful in searches and filtered reports, are often also incomplete, inconsistent or absent.

Overruns of live events, such as sports games, are usually noted for the benefit of DVR users, but not always, and deviations from schedules due to breaking news are unexpected and are almost never corrected. Because there is no standard practice for dealing with an overrun or deviation (the scheduled program might be joined in progress, the entire announced schedule might be postponed, or individual broadcasts deleted and rescheduled), such problems might not be capable of easy correction, to the consternation of clients. Indeed, because the ratings service is usually the only part of a pay-TV operation that cares about the schedule after a broadcast ends, there might not even be reliable access to such postbroadcast corrections as are made.

The scheduling of TV programming of any large broadcast company must cope with different time zones. For example, the U.S. and Canada have six time zones, Mexico has four, and South America has three, with further variation in the observance (or not) of daylight savings time and its start and end. Broadcast networks and pay-TV channels use multiple feeds, identical but offset from each other by several hours, to show programming at the correct time in each time zone. This is further complicated by network affiliates’ and pay-TV operators’ decisions (which may affect some or all of them in the market) in some areas to deviate from the schedules intended for them by using different feeds, and by the tendency of pay-TV operators to offer multiple feeds of the same channel, especially for premium channels. Furthermore, live entertainment, breaking news and some scheduled news events, and much sports programming are carried live regardless of timezone, disrupting the rest of the same feeds’ schedules.

The need to manage the resulting complexity varies by pay-TV and channel operator but generally tends to increase along with distribution capacity, which creates an incentive for pay-TV operators to add features such as regional broadcasting and for channel operators to add more geographically targeted feeds (whereupon offering them in all timezones is a free way of increasing customer satisfaction, which almost all channel operators take up).

We have successfully dealt with this daunting issue in some of the most challenging regions of the world, including, in several settings, in the U.S., in which single national ratings are expected for each network broadcast, which may be aired by its hundreds of affiliated local stations at different times—and occasionally not at all.

This problem is of much greater concern in RPD measurement because data on the monitor’s state (on/off/standby) are not available. Various methods are used to estimate the probability that part of a long period of viewing of the same channel is tuning-without-viewing. Unfortunately with some exceptions in individual countries, none of the large global TAM providers that also undertake census-level data processing publish their algorithms in even general terms, instead treating them as industrial secrets and making comparison between them almost impossible.

Controversially, some JICs (joint industry committees, which in most countries award the currency TAM contracts) have imposed strict capping requirements on all data sources when the viewer does not reconfirm his or her presence in response to a people-meter prompt, or a mobile phone used for measurement is immobile more than briefly, for both conventional TV and streaming. Such practices have the potential to alter the measurable audience dramatically, and thus to wreak havoc in the advertising market, or benefit some stakeholders (unsurprisingly, those with greater representation on the JIC) at the expense of others. This is especially inexcusable in light of the copious evidence of people-meter household members’ undetectable, unenforceable partial noncompliance with confirmation prompts, and separately research that identifies the narrow percentage of viewing sessions when the viewer was not colocated with his or her smartphone, permitting tuning-without-viewing detected by smartphone-based apps such as ours to be modelled.

The forthcoming WiFi standard 802.11bf (backwards compatible with WiFi 6 and later) offers the possibility of WLAN sensing of a person in the room. This, in combination with mobile device-based measurement, may provide the necessary evidence for the market to be confident that most tuning-without-viewing is thus eliminated.

VOD (video-on-demand) is a service that permits playback of programs entirely outside the context of broadcast schedules, although much VOD content does come from that which was broadcast. VOD can also include programs that no-one is planning to broadcast, greatly expanding the variety of available content.

Most VOD viewing is to standalone services such as Netflix or Amazon Prime Video, but pay-TV services also have their own VOD. In both cases, a broadband connection is typically required (some older systems do not require Internet connectivity). While easy to provide in a cable system, it requires separate provision of connectivity, often by the subscriber, in a DBS (direct-broadcast satellite) setup. With broadband, a larger selection can be offered. A DVR is not required, but if it exists, the delay before playback start can be shortened to trivial using file-based push technology, the quality of the playback improved, and the IP network decongested by offloading the playback stream.

There are several types of VOD:

VOD services have measurement on the server side, but like any other census-level data source, are unable to tell whether anyone is watching or how many people, nor do they have demographics beyond an often vague geographic indication that is relatable to socioeconomic class (SVOD and pay-TV system operators have the service address, but AVOD systems have only the IP address, which correlates to a vague and ever-shifting area which is almost impossible to track with precision and, even if tracked, may cover multiple socioeconomic classes). This problem can be solved by fusion with panel data.

VOD can be integrated into an RPD audience measurement system. For this, the content (programme) identifier has to match that used in realtime and DVR viewing. This is not true in some pay-TV systems, but can be made true with some effort. The advantage of measuring VOD and realtime viewing in combination, also including the provider’s OTT service, is to inform the pay-TV operator of its long-tail cost, revenue and ROI (return on investment).

Contrary to an understandable superficial impression, one cannot simply route raw audience-measurement data from the data collection system to the delivery system. The result would be utter rubbish:

These are only the most serious consequences of routing data collection directly to the delivery platform. The results would look completely outlandish.

Therefore, an intermediate processing system must exist in any conventional-TAM or RPD service, even if it is never acknowledged as a separate stage in the pipeline from collection to publication.

In RPD, the intermediate processing system is typically developed by the delivery-platform provider and contains components that are handled improperly (e.g., intab definition), opaquely (capping), or not at all (channel management, weighting). The delivery-platform providers, most of which are conventional-TAM operators, generally do not understand RPD-specific challenges, keep development to a minimum, and hide weaknesses behind “proprietary” algorithms. A pay-TV operator who permits this to happen will be unable to respond to clients’ complaints, because it does not understand how data are processed and partly is kept in the dark as to proprietary components. More importantly, the pay-TV operator would find it very difficult to change delivery-platform providers, as the published data would change radically and for reasons that it could not adequately explain. The delivery-platform provider would naturally leverage its power over the pay-TV operator for ever greater financial demands.

As part of the development, we: