· SourceSignal
Views, duration and concentration — a historical research note
Historical examples show how duration weighting changes a YouTube ranking. The resulting proxy is modelled, not observed watch time or market size.
Originally published 16 July 2026. Measurement language reviewed 10 October 2026.
Public view counts and video durations answer different questions. Views record a platform-defined count; duration records how long an upload runs. Neither tells us how many minutes viewers actually watched. Multiplying them creates a model, not a measurement of watch time.
This note retains selected tables from the original analysis as historical illustrations. Their values have not been recomputed or independently verified in this review. The original article described a 15-category corpus, but did not provide a reproducible query, exact collection window, inclusion rules or source export. These examples should not be used as current category estimates or buying recommendations.
What the duration model assumes
The original analysis used this illustrative weighting:
duration-weighted proxy = views × duration × assumed_retention(duration)
assumed_retention(L_minutes) = clamp(0.79 − 0.124 × ln(L), 0.15, 0.85)
View count and duration are public input fields. Retention is an assumption. The formula assigns roughly 65% to three minutes, 50% to ten minutes and 28% to an hour. Those percentages are outputs of the formula, not observed completion rates.
The earlier version attributed the constants to published studies without providing enough source detail to establish that calibration. This review has not verified those studies or a corpus-wide retention benchmark. We make no claim that the curve predicts actual watching behaviour. Channels can have different retention at the same video length.
When duration is in hours, the output can be expressed as modelled viewer-hours. It must remain labelled as a proxy: it is neither actual watched hours nor a measure of commercial market size.
Duration weighting changes the distribution
The historical table compares reported shares of recorded views with shares of the duration-weighted proxy. Percentages are rounded; the original denominators and exact snapshot date were not retained with the article.
| Video duration | Reported share of videos (%) | Reported share of views (%) | Share of assumed-retention proxy (%) |
|---|---|---|---|
| <1 minute | 14.1 | 34.0 | 2.2 |
| 1–5 minutes | 37.2 | 19.2 | 7.2 |
| 5–10 minutes | 23.5 | 10.6 | 8.7 |
| 10–20 minutes | 18.0 | 17.9 | 23.7 |
| 20–40 minutes | 5.5 | 11.2 | 23.5 |
| 40 minutes–2 hours | 1.4 | 5.1 | 17.5 |
| >2 hours | 0.3 | 2.0 | 17.2 |
Longer videos receive more weight by construction. The table illustrates how a measurement choice changes a ranking. It does not show that viewers spent these shares of their time on each format, or that one format is a better marketing investment. Duration alone also does not establish whether an upload is a YouTube Short, so the shortest band is labelled by length.
Concentration depends on grouping and weighting
The original Coffee & Espresso example reported these effective channel counts under the duration-weighted model. An effective count is the inverse of the sum of squared channel shares: it describes concentration rather than counting distinct channels.
| Duration group | Historical modelled effective channel count |
|---|---|
| <1 minute | 6.0 |
| 1–10 minutes | 28.7 |
| 10–20 minutes | 8.0 |
| 20–40 minutes | 3.9 |
| >40 minutes | 3.1 |
A pooled concentration figure can hide differences between duration groups. This example does not establish separate commercial markets, an entry opportunity or the ability to displace a leading channel. Such conclusions require evidence about audience overlap, relevance, production costs and performance that this model does not observe.
Similar durations may reduce the influence of the weighting assumption on shares, but do not establish robustness. The original sensitivity and correlation claims require a reproducible rerun before being cited.
Query wording and detected links measure different things
The original analysis compared a keyword-wording classification with detected commerce links. These are dataset attributes, not observations of purchasing or conversion.
| Category | Historical keyword-wording proxy (%) | Historical detected-link proxy (%) |
|---|---|---|
| Running | 80 | 26 |
| Golf | 38 | 6 |
| E-Bikes | 41 | 16 |
| Coffee | 38 | 67 |
| Smart Home | 31 | 61 |
| Power Tools | 18 | 53 |
| Gaming | 46 | 48 |
The article did not document exact denominator, attribution and weighting rules for these percentages. They illustrate two classifications; they do not estimate the proportion of viewers ready to buy or the proportion of attention monetised.
Words such as “best”, “review” and “vs” can mark comparison-oriented phrasing. They do not prove a viewer's intent. A detected commerce link records a URL in a description; it does not prove affiliate status, a click, a purchase or revenue. Missing links can reflect detection gaps or other monetisation methods. Agreement between proxies does not validate them, and disagreement does not identify which is wrong.
How to use this note
For product marketing and competitive research, the useful question is which field or modelling choice produced a conclusion. Public views, durations, query results and links can support descriptive comparisons when dates and denominators are clear. Predictions about audience attention, sales or creator performance need additional evidence.
This is exploratory Research, separate from SourceSignal's product-review evidence reports. The duration model is not a claim about measurements included in those reports.
For the available matched-pair, market-relative time series, see category momentum and creator momentum. Research methodology explains the current public measurements and their limitations.