Millions, Hundreds, Dozens

Figure 2 of Deep Neural Networks for YouTube Recommendations, the paper Covington, Adams and Sargin presented at RecSys '16 in Boston between 15 and 19 September 2016, is a diagram with three widths printed down its side: millions, hundreds, dozens. Millions is the corpus the system starts from. Hundreds is what the first stage hands to the second. Dozens is what one viewer is shown. Almost everything in this chapter falls out of that one figure, and the figure is ten years old, because no Google publication since 2019 has re-described the architecture end to end.

The reason the first stage exists is published on a different page. YouTube's How search works page, retrieved 5 September 2026, states that 500 hours of video are uploaded to the platform every minute. Nothing scores a corpus of that size against one viewer inside the time a page takes to load. So the system does the job twice, cheaply and then expensively, and the two halves fail in ways that look identical on a dashboard and call for opposite responses.

Candidate generation is the cheap half. In the words of Covington et al., it "takes events from the user's YouTube activity history as input and retrieves a small subset (hundreds) of videos from a large corpus." It runs on "coarse features such as IDs of video watches, search query tokens and demographics," and the 2016 paper is explicit about what those buy: "broad personalization via collaborative filtering." Broad is the load-bearing word. At this stage the system is not reading a title or looking at a thumbnail. It is looking for videos that people with a watch history resembling this viewer's went on to watch.

Ranking is the expensive half. The 2016 paper says the second stage assigns "a score to each video according to a desired objective function using a rich set of features describing the video and the user," and section 4 of that paper gives the order of magnitude: "We typically use hundreds of features in our ranking model." Hundreds of features, applied to hundreds of candidates, for one viewer, on one page load.

Two stages, two ways to lose. A video that candidate generation never nominated was not scored badly. It was not scored. A video that ranking scored badly was nominated, weighed against a few hundred others for that particular viewer, and beaten. Those are different failures with different causes, and the response to one is useless against the other. The conflation is the single most expensive habit in this subject, because a creator who has been passed over at the first stage will spend a weekend redesigning a thumbnail nobody was ever shown.

One sentence in the 2016 paper does more damage to the definite article than any myth-busting article has managed. The two-stage design, Covington et al. write, "enables blending candidates generated by other sources." There is not one retrieval system. There is a committee of them, and Zhao, Hong, Wei and colleagues confirm it in Recommending What Video to Watch Next: A Multitask Ranking System, presented at RecSys '19 in Copenhagen between 16 and 20 September 2019: "Our video recommendation system uses multiple candidate generation algorithms, each of which captures one aspect of similarity between query video and candidate video."

So a creator asking what the algorithm did is asking about several nominating systems, one scoring system and one viewer, all at once, and expecting a single answer. Nothing in this book will restore the definite article, because the published record does not support it.

Three figures from the 2016 paper are worth carrying, not as a description of the system running in 2026 but as a sense of the object being argued with. Covington et al. report that "our models learn approximately one billion parameters and are trained on hundreds of billions of examples." Section 3.5 of the same 2016 paper describes what goes in: "a vocabulary of 1M videos and 1M search tokens were embedded with 256 floats each in a maximum bag size of 50 recent watches and 50 recent searches." Section 3.1 states the budget the whole thing runs inside: "a strict serving latency of tens of milliseconds."

Fifty recent watches and fifty recent searches. That was the window the first stage was given in the 2016 paper, and it is the window a video is being matched against. A title is not in that list. A thumbnail is not in that list. What is in it is the record of what one person has already watched, which is not a thing any creator can edit.

The last of those three figures shapes the others. A latency budget of tens of milliseconds does not buy an exhaustive comparison, and the 2016 paper does not claim one: serving at the first stage is described there as an approximate nearest-neighbour search in dot-product space. Covington et al. then add a line that deserves to be read slowly by anybody who has paid for an explanation of the mechanism: "A/B results were not particularly sensitive to the choice of nearest neighbor search algorithm." The method by which a video is fished out of the corpus was, in 2016, interchangeable enough that swapping it barely moved the live numbers. The nomination step is an approximation chosen for speed, and its exact form did not greatly matter to the people who built it.

The most useful sentence in the 2016 paper for anyone who has watched a graph flatten sits in section 4.1, and it is about impressions. "Features describing the frequency of past video impressions are also critical for introducing 'churn' in recommendations (successive requests do not return identical lists)," the paper says. "If a user was recently recommended a video but did not watch it then the model will naturally demote this impression on the next page load."

Read the second sentence twice. It describes something that happens to an impression, on one page load, for one viewer, because that one viewer declined the video. It is per-viewer, not per-video. There is no counter on the video being decremented. There is a person who was shown a thumbnail, did not click it, and will be shown it a little less on their next page load. There are a great many of those people, each demoting independently, and what they add up to is a curve that climbs and then flattens without anybody deciding anything.

There is no lever. Nobody did anything to you.

That is a mechanism and not an exoneration. The 2016 paper describing the churn feature is ten years old; the settings the live system runs on have never appeared in any Google publication; and as of September 2026 no external mechanism existed by which anyone outside the company could inspect them. A system can be innocent of the specific thing it is accused of and still be unauditable. Both of those sentences are true at once, and a book that prints only the first one has been written by the defendant.

The same section 4.1 of the 2016 paper puts the per-viewer point in a form that reads strangely against the way channels talk about their audiences. "The most important signals are those that describe a user's previous interaction with the item itself and other similar items," Covington et al. write, and the worked example they give is a pair of questions the model asks about the viewer: "how many videos has the user watched from this channel? When was the last time the user watched a video on this topic?" That is one viewer's history with a channel. It is not a subscriber count, and it is not a property of the channel at all.

The platform's own documentation does not leave the matter entirely at the level of the individual viewer, and the honest version of this section prints both readings rather than the convenient one. The How YouTube Works recommendations page, read on 5 September 2026, describes the system in terms of a viewer's own behaviour and feedback, and then names something that is not per-viewer at all: "the reputation and the quality of a channel." Those words sit on a Google-published page, and they sit awkwardly beside a 2016 paper whose channel signal is a question about one person's history. Neither document explains the other. The paper is dated to September 2016 and the page carries no publication date at all, which is one reason this book quotes both and reconciles neither.

That page is worth a note on its address as well as its contents. The equivalent page about search on the same site now redirects into YouTube's Help Centre, while the recommendations page still resolved when it was read on 5 September 2026. The site YouTube built to explain itself is being folded into the general help system a page at a time, and a citation to it is a citation to a document in the middle of a move.

The training design points back toward the individual. Covington et al. describe generating "a fixed number of training examples per user, effectively weighting our users equally in the loss function," which, they write, "prevented a small cohort of highly active users from dominating the loss." The system was built, in 2016, to refuse to let its heaviest viewers decide what everybody else is shown. That is a deliberate flattening at the level of people, and it is the opposite of the picture in which a video is pushed at a crowd. The unit being modelled is a person, one at a time, and the paper keeps saying so in different ways.

No single seller wrote that sentence. It is a composite, assembled from wording that circulates across the search results for the term this chapter is about, and it is printed as a composite rather than attributed to anybody; if an archived seller page carrying the wording verbatim is obtained before this book is set, it will be substituted and named. Every clause in it inverts the architecture in the 2016 paper, and they are worth separating, because they fail for different reasons.

The first clause has the direction of travel backwards. Retrieval is triggered by a viewer opening YouTube, not by a creator pressing publish: the first stage "takes events from the user's YouTube activity history as input," in the 2016 paper's words, and a video is one of the millions of things that history is matched against. Todd Beaupré, then leading YouTube's growth and discovery team, told Search Engine Journal on 4 March 2024: "A lot of creators think of YouTube as pushing videos out to a bunch of people, but it's actually more the reverse." He said it about the system as it stood in 2024, and it agrees with a paper published in 2016.

The second clause invents a batch. No Google publication and no YouTube help page read for this book on 5 September 2026 describes a test audience of a defined size that a new upload is issued to. What is documented is a per-request retrieval running for individual viewers, and, in the 2016 paper's section 4.1, a per-viewer demotion applied to people who declined. As of September 2026, no primary source for a fixed test batch had been located, and a claim that cannot be found in a document is reported here as absent rather than as false.

The third clause, that a video gets killed, describes an outcome that is real and a mechanism that is not. Impressions do stop. Beaupré, in an interview recorded 1 September 2026 and reported by Luis Rijo for PPC Land on 5 September 2026, described a cold-start channel where the system matches "topical preferences against what a video is understood to be about," and said that where no interested audience emerges, "impressions stop." That is a video failing to find anybody, which is the first stage of the funnel doing nothing, rather than a second stage passing sentence.

Beaupré's account carries a label, and the label comes from the outlet that reported it. PPC Land, in that report of 5 September 2026, notes that the interview published no methodology and that the interviewer was a YouTube employee who put no challenges to him. This book adopts that characterisation. What he says about cold-start retrieval is authoritative, in the sense that he is among the people who would know, and unaudited, in the sense that nobody outside the company can check it. Both halves travel with every employee statement quoted in these pages.

The fourth clause is the only one with a recurring price attached, and the price is paid in the reader's own hours. A first-hour ritual is not a one-off cost. It is a block of attention spent on every upload, indefinitely, and what it is spent watching is the churn feature from section 4.1 of the 2016 paper doing its work: individual viewers being served a thumbnail, declining it, and demoting it for themselves on their own next page load. Nothing in that description is a decision about the video, taken at a moment, by anything. There is no meeting to be present for. The instruction in the last clause rests on the push picture in the first, and it falls with it.

Search is a different system from recommendations, and tags live in search. This is the one place in the subject where two official statements are read as a contradiction and are not one.

YouTube's How search works page, retrieved 5 September 2026, names three ranking factors, and the first of them is where tags appear: relevance, defined on that page as "how well the title, tags, description, and video content match your search query." The other two factors named on that page, read the same day, are engagement and quality. The page carries no changelog and YouTube edits these pages silently, so anyone acting on it in a later month should open it first and read it again.

The same page, read on 5 September 2026, also closes off the question that arrives next in every comment section: "YouTube doesn't accept payment for better placement within organic search results, nor do we treat content that is Google-owned more favorably than any other creator." That is a statement a party has published about itself, and it is printed here for the same reason every other quotation in this chapter is printed. It is what the document said on the day it was read.

YouTube's tags help page, retrieved 5 September 2026, carries the other half. "Tags can be useful if the content of your video is commonly misspelled," it says. "Otherwise, tags play a minimal role in your video's discovery." The same page, read the same day, ranks what matters instead: "Your video's title, thumbnail, and description are more important pieces of metadata for your video's discovery."

Both statements are official, both were read on 5 September 2026, and neither corrects the other, because they describe different surfaces. The formulation this book uses is the narrow one that survives both pages: tags are a minor search-relevance input and near-irrelevant to recommendations. Tag research sold as a discovery method is work on the smaller of two surfaces, priced as though it were work on the larger. The use the tags page itself puts forward, read on 5 September 2026, is narrow enough to reprint as the whole of the advice: a video whose subject is commonly misspelled can carry the misspelling. That is the case the page makes, and it is the case this book repeats.

The only worked example in this chapter is a failure, and the failure belongs to the system rather than to a creator. Section 3.4 of the 2016 paper describes an experiment its authors ran on themselves: they gave the model the viewer's most recent search query, and the model learned to predict that the viewer would watch whatever was sitting on that search results page. "Unsurprisingly, reproducing the user's last search page as homepage recommendations performs very poorly," Covington et al. report. The fix was subtraction. They threw the sequence away and represented the query as an unordered bag of tokens, deliberately blinding the model to something it knew.

The same part of the 2016 paper carries a second design choice with a consequence a creator can actually use. The authors found "much better performance predicting the user's next watch, rather than predicting a randomly held-out watch," and they give the reason in one sentence: "Episodic series are usually watched sequentially and users often discover artists in a genre beginning with the most broadly popular before focusing on smaller niches." A model trained to predict what comes next, rather than to fill a hole in the middle of a history, learns an order.

That order runs one way. Where viewers arrive at a subject through one video and then move outward into it, the relationship between the two videos is directional, and being the video people arrive through is a structurally different position from being the video they arrive at afterwards. The 2016 paper puts no value on that position, and no figure for it has been published since. What the paper does say is that the asymmetry was pronounced enough to change how the model was trained.

They generalise the lesson in one sentence, and it is the best short account anyone at Google has published of why this system resists reverse-engineering: "Somewhat counter-intuitively, great care must be taken to withhold information from the classifier in order to prevent the model from exploiting the structure of the site and overfitting the surrogate problem." The engineers who built it withheld information from it on purpose, to stop it learning the wrong lesson from the shape of their own product. A system assembled that way has no front door, and nobody is going to derive its behaviour from a weekend of dashboard-watching, including the people selling that derivation.

One more line from the 2016 paper corrects an inversion worth naming, because the correction runs against the reader's interest as often as with it. Training examples are generated from "all YouTube watches (even those embedded on other sites) rather than just watches on the recommendations we produce," because otherwise, Covington et al. write, "it would be very difficult for new content to surface." That describes what the nominating stage is trained on. It is not a statement about impressions, and it does not mean that off-platform traffic registers as one.

Two admissions close this chapter, and the first one is about the evidence it just used. The architecture described here was published in 2016, and its ranking half was described again in 2019. Whether the system serving a viewer in 2026 still has that shape is not something anybody outside Google can establish, which is why this chapter asserts the shape of the funnel and asserts nothing whatever about its settings.

The second is that nobody in this chapter is a success story, and the failure data underneath it is missing in a specific, nameable way. No published figure exists for what share of uploads are never nominated by candidate generation for anybody at all. YouTube has never published its traffic-source distribution by category, and no creator cohort has published one either, so the base rate of the failure this chapter describes — the video that was not considered — is unknown. As of September 2026, this was unresolved. Two things would close it, and neither of them is YouTube: a cohort of creators publishing their own Studio traffic-source exports as impression counts, with format and length attached and the channels that stopped uploading still in the sample; and an academic with partner-level access to the Analytics API who is willing to print the sampling frame beside the result.

For the last video that did not work, write down one number before touching anything else. Open Studio, open Traffic sources, and take impressions from Browse features. Not the percentage. The count. A percentage of a small number is a small number, and the two widths printed on that ten-year-old figure fail differently enough that the count is the first thing that tells them apart.