How Affiliate Links Become a Measurable Compliance Problem
The YouTube affiliate-disclosure study does not ask only whether creators use affiliate links. It asks how researchers can identify those links at scale and then determine whether the associated text explains compensation and relationship clearly enough to match the FTC-aligned standard. This chapter explains the study’s population, measurement pipeline, and headline rates so that the numbers can be read with their denominators and limits rather than mistaken for a platform-wide legal verdict.
What the sample actually contains
The study combines four collection sources: YouTube links found in Reddit posts, a random sample of video identifiers, videos promoted through YouTube’s Trending surfaces, and videos selected through YouTube’s Shopping surface. Together, these sources contain 2.00 million publicly listed videos uploaded between January 2015 and December 2024, spanning 539.9 thousand channels and 16 video categories. The researchers extract hyperlinks from video descriptions and, when present, the tagged-products section associated with YouTube Shopping. That produces 4.13 million unique hyperlinks for analysis.
The sampling design is useful precisely because the sources expose different parts of the platform. Reddit-linked videos reflect one route by which audiences and communities surface content. Random identifiers provide a less curated comparison. Trending videos represent algorithmically promoted material, while Shopping videos reflect a platform surface organized around product discovery. None of these sources is identical to the experience of a typical YouTube viewer, and their combination does not remove selection effects. It does, however, let the study compare affiliate behavior across contexts instead of treating the platform as a single undifferentiated feed.
The sources also create different interpretive baselines. A video found through a community post is not selected for the same reason as a video placed in a Shopping surface. A Trending sample says something about content receiving algorithmic exposure, while a random sample supplies a different reference point. A high or low rate in one source should therefore be read as a property of that sampled context, not as a ranking of all creators or all YouTube viewers. The study’s comparisons are valuable because they expose contextual variation, but the variation is part of the evidence boundary rather than a nuisance that can simply be averaged away.
That distinction should govern how the results are stated. “In the study’s dataset” is not a timid substitute for “on YouTube.” It identifies the population to which the measurement directly applies. The dataset is large enough to reveal patterns that a handful of case studies could miss, but size alone does not make it a census. The paper notes that Gaming, Music, and Entertainment are over-represented, partly because Reddit disproportionately surfaced Gaming content. Shopping data covers only 2024, and Trending data begins in 2020, which also limits some comparisons over time.
The unit of analysis matters as well. The study reports both videos and channels because a video-level percentage answers a different question from a channel-level percentage. A channel may publish many videos, and repeated links from one creator can influence a video count without representing the behavior of the same number of distinct creators. Keeping the denominators visible prevents a large number of videos from being silently translated into a claim about an equally large number of independent voices.
The study has a second boundary: it examines links and disclosures in specific textual surfaces. It analyzes English-language text in video descriptions and the Shopping tab, not every spoken statement or visual notice inside a video. That choice follows the links being measured, but it means the findings cannot answer whether a creator made an adequate disclosure in audio or on-screen graphics. A creator could satisfy an obligation through a modality outside the study’s measurement, although the study cannot verify that possibility. The reported rates therefore concern a defined text-and-link problem.
Identifying the links is itself nontrivial. A simple keyword search can recognize a URL that openly contains an affiliate identifier, but it can miss shortened links or links whose commercial behavior appears only after several redirects. The study therefore combines known URL patterns with browser interaction graphs that follow each relevant link to a final landing page and record the structure and redirect behavior along the way. This approach is meant to observe how the link behaves, not merely how it looks when copied from a description.
The distinction matters for prevalence. A visible domain name is not the same thing as the economic path that a browser follows. A shortened URL may hide the eventual merchant, and a dual-purpose domain may serve both ordinary and affiliate traffic. Redirect-aware measurement can therefore identify relationships that a surface-level text rule would miss. It also introduces its own assumptions: the crawler’s behavior, the selected domains, the ground-truth labels, and the definition of an affiliate link all affect the final estimate. A method that follows redirects is more informative than a keyword-only rule for this problem, but it is still a measurement of observed link behavior, not direct access to a creator’s private contract.
For disclosure quality, the researchers annotate 2,000 sentence-level segments and train classifiers to identify disclosure text and assess two dimensions: clarity of compensation and clarity of relationship. The reported holdout F1 scores are 96.8% for disclosure detection, 98.6% for compensation clarity, and 97.0% for relationship clarity. The link classifier reports F1 scores of 92.8% to 96.1% on two holdout sets, compared with 18.5% to 23.1% for earlier regular-expression approaches described in the paper.
Those performance figures support confidence that the study’s method can reproduce its labels on its evaluation sets. They do not mean that the classifier has direct access to legal truth. A model can be accurate against an annotation scheme while the scheme remains a practical approximation of a fact-specific regulatory judgment. The relevant question for reading the results is therefore not whether automation has eliminated uncertainty. It is whether the method is sufficiently explicit about what it counted and whether the resulting patterns survive the stated limitations.
The measurement pipeline can be summarized as follows:
The diagram separates the stages that are easy to conflate. The study first identifies a population of videos and links. It then identifies which links appear to be affiliate links. Only after that does it evaluate the text associated with those links. A low disclosure rate is therefore not a count of all descriptions that lack the word “affiliate”; it is a result conditional on the study’s affiliate-link detection and its textual disclosure labels.
What the measurements say—and what they do not
Across the full dataset, the researchers identify 352.7 thousand unique affiliate links in 146.8 thousand videos. The 146.8 thousand videos represent 7.35% of all videos in the dataset. The links appear across 36.8 thousand channels, or 6.81% of all channels, and represent 8.55% of the dataset’s hyperlinks. Among videos that contain affiliate links, the average is 5.28 unique affiliate links, and those links account for 38.02% of the links in their descriptions. These denominators matter: 7.35% is a share of sampled videos, while 8.55% is a share of hyperlinks, and neither is a direct estimate of the share of viewers exposed to affiliate commerce.
The figures nevertheless establish that affiliate use is not an isolated practice in the measured ecosystem. A channel can place several affiliate links beside a recommendation, and a large number of ordinary-looking URLs can make it difficult for a viewer to identify which routes create compensation. The more links a description contains, the more disclosure quality depends on attribution and organization, not only on whether a general notice appears somewhere above the list.
The transition from the full dataset to the disclosure denominator is another important step. Affiliate detection identifies a subset of videos; the disclosure analysis then focuses on 139.9 thousand English-language affiliate videos. That subset is not interchangeable with the original 2.00 million videos. A reader who carries the 12.20% compliance rate back to all videos would be changing both the population and the question. The rate asks how the detected affiliate videos in the measured English-text scope are classified, not how many of all YouTube videos satisfy a general advertising rule.
Of those 139.9 thousand videos, 45.52% contain some form of disclosure. The study classifies 12.20% as clearly compliant, 18.61% as possibly compliant, and 69.19% as non-compliant. Within the non-compliant group, 54.19% contain no disclosure at all. The remaining non-compliant cases are not necessarily silent; they fail because compensation is not explained or because the relationship cannot be attributed clearly enough to the relevant links.
The three categories are built from two dimensions. Compensation is clear when the language explicitly states a compensatory relationship, such as the creator receiving a commission. It is ambiguous when the language implies support without stating the financial arrangement, and it is absent for the study’s purposes when a phrase names an affiliate link without explaining what compensation means. Relationship clarity concerns the link or product to which the notice applies. An explicit notice identifies each relevant link; a grouped notice can cover a defined list; a mixed-group notice leaves the viewer uncertain about which links are affiliate links and which are not.
This construction explains why “45.52% disclosed” and “12.20% clearly complied” are not contradictory. The first number asks whether any disclosure-like text appears. The second asks whether the text communicates compensation and relationship in the way the study operationalizes the FTC standard. A creator can make a good-faith attempt and still fall into the possible-compliance group if the viewer must guess which links are covered or what benefit the creator receives.
The rates should also be read as estimates with two kinds of uncertainty. There is statistical and classification uncertainty inside the sampled data: a model may make an error, and a different annotation boundary could change some labels. There is external-validity uncertainty outside the sample: different languages, categories, platforms, video surfaces, or unmeasured audio and visual notices may behave differently. The authors repeat their RQ1 analyses on random half-samples within subgroups and report that the trends remain consistent, with coefficients of variation below 10%. That robustness check makes it less likely that the main directional pattern is driven by a few outliers, but it does not turn the sample into a census or remove modality bias.
The result is also different from saying that 69.19% of all YouTube videos violate the law. The denominator is the study’s English-language affiliate-video subset, the labels are produced by classifiers trained on human annotations, and the analysis excludes audio and on-screen visual disclosures. “Clearly compliant” and “non-compliant” are research categories aligned to the paper’s reading of the FTC guidance, not adjudications of individual videos. The study’s automated collection also expanded descriptions by selecting “more,” an action some viewers may not take. Because that procedure makes more text available to the classifier than a viewer may initially see, the paper describes its visibility measurement as an upper bound.
The right conclusion is therefore directional and bounded. In this large, multi-source sample, affiliate links are common enough to create a platform-scale transparency problem, while clear textual disclosures are much less common than affiliate use. The evidence is strong enough to justify investigating language, placement, interface, and stakeholder responsibility. It is not sufficient to establish the cause of each missing disclosure, quantify all consumer harm, or determine whether every unmeasured audio or visual notice would change the result.
This distinction between signal and scope is important for the next step. Aggregate percentages show that the gap exists, but they do not tell us how it is produced. To understand why a disclosure is classified as possible or non-compliant, we need to look at the mechanisms beneath the labels: vague compensation language, mixed link groups, hidden placement, and the different contexts in which creators use affiliate links.