National television audience figures are estimates produced from a sample, not a census of who watched. Understanding how the sample is built explains why ratings move in ways that seem strange and why they are still trusted by advertisers.
A panel stands in for the country
Measurement firms recruit households that agree to have their viewing metered, and those households are chosen to mirror the national population across age, income, region, household size and ethnicity.
The panel is a small fraction of American homes. Its value comes from being representative rather than from being large, in the same way a well-drawn poll can describe an electorate.
Recruitment is the hard part. Households that refuse participation are not randomly distributed, so firms spend heavily on persuading reluctant groups to join and stay enrolled.
Weighting scales the sample up
Each panel household represents many real households, and the multiplier differs by demographic group depending on how the panel over- or under-represents that group.
If younger urban households are harder to recruit, the ones enrolled carry more weight per person. That weighting is what converts a modest panel into a national projection.
Heavy weighting introduces volatility. When a small number of households represent a large slice of the population, one household's unusual week visibly moves the published number.
Ratings and share measure different things
A rating expresses viewers as a portion of all households with a television, whether or not the set is on. Share expresses them as a portion of sets actually in use.
Late-night programming can post a weak rating and a strong share because most sets are off. Both numbers are correct and they answer different questions.
Advertisers usually care about neither in isolation. They buy against demographic subgroups, so a program's value depends on which audience it delivers rather than its raw total.
Delayed viewing complicates the count
Recorded and on-demand viewing is credited in windows, commonly counted for a set number of days after broadcast. A program has several audience figures depending on the window used.
Commercials complicate this further, because a viewer who skips advertising has watched the program without delivering the thing being sold. Measurement systems account for commercial viewing separately.
Streaming pushed measurement further still, since platforms hold their own server-side data and define a view by their own criteria. Independent panels remain the only common currency across services.
Why error bars are rarely published
Every panel estimate carries statistical uncertainty, and for small cable programs that uncertainty can be a large portion of the reported figure.
Published charts almost never show it, because the industry uses ratings as a transactional currency and a currency with visible ranges is difficult to trade against.
The practical consequence is that small week-to-week movements in low-rated programs often mean nothing at all, while sustained multi-week trends in large ones are real.