The Objective They Threw Out

Section 4 of Deep Neural Networks for YouTube Recommendations opens on the sentence the creator economy has argued about for ten years and rarely cites. Covington, Adams and Sargin presented the paper at RecSys '16 in Boston between 15 and 19 September 2016, and the section begins: "Our final ranking objective is constantly being tuned based on live A/B testing results but is generally a simple function of expected watch time per impression. Ranking by click-through rate often promotes deceptive videos that the user does not complete ('clickbait') whereas watch time better captures engagement."

That is the entire anti-clickbait argument, and none of it is a moral position. It is a note on why one objective was discarded in favour of another, written by the three engineers who discarded it, in a paragraph about loss functions. The thing a creator has heard repeated as a feeling — that a misleading thumbnail wins the click and costs the channel — is a design decision from 2016 with an ablation table attached, and the table is the part nobody quotes.

Read the objective as it is written: expected watch time per impression. Not total watch time, which a long video accumulates partly by being long, and not views. A rate. The unit being scored is one opportunity to show one video to one person, and the quantity being predicted is how much of that person's time the opportunity is likely to be worth. Advice to make videos longer and advice to make them shorter are both arguing about a numerator without naming the denominator.

The first clause of the same sentence dates everything built on top of it. The objective is "constantly being tuned based on live A/B testing results." On the day the paper was published in 2016, its authors were telling the reader that the thing they had just described was already moving. An objective described as under continuous adjustment in 2016 is not a specification anybody can quote as current in 2026, and the two papers are the strongest evidence in this book partly because they say that about themselves.

The mechanism under the sentence is weighted logistic regression, and the weighting is where the objective actually lives. In training, an impression that was clicked is weighted by the watch time observed after the click, and an impression that was not clicked carries unit weight. The odds the model learns therefore approximate expected watch time rather than probability of a click, per Covington et al. A click with nothing behind it enters the arithmetic as a small number. That is not a punishment for anything. It is a weight.

Depth was chosen by measurement, and the measurement was printed. Covington et al. report weighted per-user loss on a next-day holdout, lower being better, across the seven network configurations they tried for the 2016 paper:

They shipped the deepest configuration, and the published reason is a budget rather than an insight: it "gave us the best results while enabling us to stay within our serving CPU budget." The same section records what the objective itself was worth. Weighting clicked and unclicked impressions equally, instead of weighting by watch time, "increased the watch time-weighted loss by a dramatic 4.1%," which is the 2016 paper's own adjective.

Two cautions travel with that table. Weighted per-user loss is an internal training metric and converts into nothing a creator can open in Studio; it is reproduced here to show that the objective was chosen by measurement rather than by taste. And the comparison is dated 2016, run on a next-day holdout, and nothing in the paper gives a reason to think the same seven rows would come back the same way now.

The turn that decision belongs to had already been announced in public four years earlier. YouTube's blog post YouTube Now: Why We Focus on Watch Time, dated 10 August 2012 and bylined Eric Meyerson, then Head of Content Creator Communications, puts it in two sentences: "Our video discovery features were previously designed to drive views. This rewarded videos that were successful at attracting clicks, rather than the videos that actually kept viewers engaged." The post is often attributed to Cristos Goodrow, the engineer usually named as the internal champion of the switch. The byline on the page dated 10 August 2012 is Meyerson's, and the difference matters to anyone who intends to quote it in front of somebody who will check.

Watch time stopped being the objective by 2019 at the latest, and the document that settles it is the second paper Google published about this system. Zhao, Hong, Wei and colleagues presented Recommending What Video to Watch Next: A Multitask Ranking System at RecSys '19 in Copenhagen between 16 and 20 September 2019. It splits what the ranker optimises into two families, verbatim: "1) engagement objectives, such as user clicks, and degree of engagement with recommended videos;

1. satisfaction objectives, such as user liking a video on YouTube, and leaving a rating on the

recommendation."

The same paper of 2019 states the feedback it learns from in the same two-part shape: "Our ranking system learns from two types of user feedback: 1) engagement behaviors, such as clicks and watches;

1. satisfaction behaviors, such as likes and dismissals." A dismissal is a documented input. So is

a like. So is a click. The list is not ranked and the paper does not rank it.

Two things that sound identical are being held apart here, and the difference decides what a reader carries away from the whole argument. Covington et al. rejected ranking by click-through rate in

1. Zhao and colleagues list user clicks among the engagement objectives in 2019. The click was

never removed from the system. It was demoted from being the objective to being one input among several, for the reason the 2016 paper gives in the same breath: a rate that can be won by deception is a poor proxy for the thing the ranker is trying to buy. Anyone who reads "clickbait does not work" as "the click does not matter" has taken the wrong half of a published sentence.

Those families do not always agree, and Google is the party saying so. The Multi-gate Mixture-of-Experts architecture the 2019 paper introduces was adopted because "hard-parameter sharing techniques sometimes harm the learning of multiple objectives when correlation between tasks is low." That is a sentence about conflict. A video that wins the click and loses the viewer pulls two of the system's own objectives in opposite directions, and the 2019 architecture exists to arbitrate that pull rather than to abolish it.

Then, in §4.2, the sentence this chapter exists to deliver. Describing how the separate predictions are combined into a single score, the 2019 paper says: "we take the input of these multiple predictions, and output a combined score using a combination function in the form of weighted multiplication. The weights are manually tuned to achieve best performance on both user engagements and user satisfactions."

Manually tuned. There is a combination function, it has weights, and human beings set them. That is the one place in the published record where a hand is documented as turning something, and it is on the far side of the glass from the reader. The settings have never appeared in a Google publication, no court and no regulator has ever compelled their disclosure, and as of September 2026 no external mechanism existed to inspect them. Every percentage weight for one signal against another that a reader has been sold was invented by the seller.

Two things follow from that sentence and only two. The first is that a number exists which nobody outside Google has seen. The second is that the number is a property of the system rather than of any channel. A combination weight sits inside the scoring function for everybody at once, and nothing in either paper describes a per-channel adjustment, nor does any employee statement in this book's source list. The dial is real, and it is not pointed at anybody.

Two figures from the same paper fix the scale of what a win looks like inside this system. Section 5.2.2 of the 2019 paper defines the two metrics it reports, and they are narrower than the objective families above them: the engagement metric captures user time spent watching recommended videos, and the satisfaction metric captures user survey responses with rating scores. In Table 1 of that 2019 paper, the mixture with eight experts, at 6.1 million multiplications, returned +0.45% on the time-spent metric and +3.07% on the survey-rating metric in a live YouTube A/B test against the shared-bottom baseline. The neighbouring rows of that 2019 table give the shape of the gain: a shared-bottom model at the same 6.1 million multiplications returned +0.1% and +1.89% on the two metrics, and a mixture with four experts at 3.7 million multiplications returned +0.20% and +1.22%.

Table 2 of the same 2019 paper prices a smaller problem, position bias, on the engagement metric: position supplied as an input feature returned −0.07%, an adversarial loss returned +0.01%, and the shallow tower the paper adopts returned +0.24%. Read the sign on the first of those. One of the three approaches the team tried made its own metric worse, and the paper printed the number anyway, which is more failure data than the course economy has ever published about itself.

That table exists because the 2019 system models the effect of screen position and then subtracts it. A separate shallow tower learns how much of a click is explained by where the video sat on the page, and its output is added to the main model's score so that the prediction factorises into user utility on one side and an estimated propensity on the other, per Zhao et al. The reading for a creator is narrow and worth having anyway: the system already knows that a good slot produces clicks by itself, and it is built to discount them.

Half a percent on a time-spent metric is a headline result at this scale, published in a peer-reviewed venue in 2019 by the team that operates the system. Any pitch describing a multiple of a channel's views is describing a magnitude the engineers do not claim about their own work.

The 2019 paper concedes one more thing that the sales material never does. Models are trained on data the system itself produced, which is a record of what the system already chose to show. In the paper's words, such models "will be biased towards the current system, causing a feedback loop effect," and: "How to effectively and efficiently learn to reduce such biases for ranking systems is an open question." An open question, in a 2019 paper, written by the people running it, about whether the system can see past its own history.

One structural claim about this evidence has to be corrected before the evidence is used, because it is the claim this book very nearly made. Google's published record on YouTube ranking is routinely described as stopping in 2019. It thins sharply and it does not stop. No paper since 2019 re-describes the architecture end to end, so the shape of the system as published is still Covington 2016 and Zhao 2019. The component-level work on the live model continued.

Tang, Drori, Chang and colleagues presented Improving Training Stability for Multitask Ranking Models in Recommender Systems at KDD '23 in Long Beach between 6 and 10 August 2023, reporting on "a real-world multitask ranking model for YouTube recommendations" in work "Conducted on a YouTube production dataset." Singh, Vu, Mehta and colleagues, in Better Generalization with Semantic IDs, last revised 30 May 2024, write that "we evaluate our approaches in a real-world ranking model for YouTube recommendations." Earlier than either, Chen, Beutel, Covington and colleagues described "a production top-K recommender system at Youtube" in Top-K Off-Policy Correction for a REINFORCE Recommender System, presented at WSDM '19.

Four names recur across that run. Xinyang Yi, Lichan Hong, Li Wei and Ed H. Chi appear on the 2019 paper, on the 2023 paper and on the 2024 one, which is the signature of a team still publishing rather than of a record that stopped. The honest position is three sentences long. The shape of the system was last published end to end in 2019. Component work on the live ranking model was still being published in 2024. The current weights have never been published at all.

The nearest thing to an official list of signals is not a paper. Cristos Goodrow, VP of Engineering, published On YouTube's recommendation system on blog.youtube on 15 September 2021 and named them: clicks, watchtime, survey responses, sharing, likes and dislikes, with the surveys measured on a 1–5 star scale sitting behind the post's own phrase, valued watchtime.

Click-through rate is not on that list. It is not on the recommendation-signal list YouTube's help documentation carried when it was read on 5 September 2026 either, where the signals named are watch history, search history, channel subscriptions, likes, dislikes, "Not interested," "Don't recommend channel," and satisfaction surveys. Impressions click-through rate is a metric the platform reports to a creator about their own videos. It has never been published as an input to recommendations.

The word satisfaction in these documents is not a mood, and it is not a guess made from behaviour. It is an answer to a question. Goodrow's post of 15 September 2021 describes surveys that ask viewers to rate videos on a 1–5 star scale, and §5.2.2 of the 2019 paper reports its satisfaction metric as user survey responses with rating scores. That is a data source a creator never sees. It does not surface in any Studio panel, it is collected from viewers the platform selects, and no published route exists for a creator to read even the responses their own videos received. Anybody selling a method for raising satisfaction in 2026 is selling a method aimed at a survey whose results the buyer will never be shown.

Two sentences from the post of 15 September 2021 belong together, because they point opposite ways. "Recommendations drive a significant amount of the overall viewership on YouTube, even more than channel subscriptions or search." And: "our recommendation system doesn't operate off of a 'recipe book' of what to do." The first says the system decides most of what gets watched. The second says there is no list to hand over. Both were published by the same executive on the same day.

A vacancy that size has a market in it, and the market is served.

The direction of the first claim is right and the year attached to it is invented. The satisfaction architecture was published at RecSys '19 in September 2019, and the satisfaction surveys were public in Goodrow's post of 15 September 2021 at the latest. The dating fits in four rows, and every one of them is a document carrying a publication date on its face.

Nothing in that column is dated 2026. A piece that dates the satisfaction shift to 2026 is redating a change documented in 2019 and made public in 2021, and the redating is the product.

The other two claims are worth walking to their ends, because both ends are on the vendor's own page. The weight chart carrying the first-thirty-seconds figure is labelled, by its own author, "illustrative, not to exact scale" — a number presented as a ranking weight and disclaimed as decorative by the person who drew it. The multiple for shares against likes is sourced, on that page of 24 April 2026, to the vendor's own internal data, which is a citation to itself. The words "YouTube confirmed" are hyperlinked on the same page to a blog post published by a different vendor. No YouTube document is cited anywhere in the piece.

YouTube has never published a signal weight. What is on the record is the opposite: Zhao and colleagues wrote in 2019 that the weights exist and that people set them by hand, and that sentence is the whole of the public knowledge about them.

The one employee statement on the record about how those weights behave says they do not hold still. Beaupré, in the interview written up by Ad Outreach on 14 March 2025, said the system had been enabled to learn that different factors carry different importance in different contexts, and gave the example himself: "Watch time may be more important in television versus mobile, or it may be more important in certain types of content like podcasts as opposed to music." A fixed percentage attached to any one signal is therefore not merely unpublished. It is the wrong shape for the thing being described, and a chart that assigns one is describing a system nobody has claimed exists.

The absence runs both ways, and this chapter would be dishonest if it reported only the half that suits it. Nobody has published outcome data on the creator side of this either. There is no cohort study of channels rebuilt around engagement signals, and no survey with its sample disclosed of channels rebuilt around satisfaction signals. As of September 2026 the published record on objectives ran in one direction only: Google, describing its own metrics moving by fractions of a percent, in venues Google chose.

Which leaves the sentence a YouTube director now offers in place of all of it. Todd Beaupré, Senior Director of Growth and Discovery, recorded on 1 September 2026 in the Creator Insider interview reported by Luis Rijo for PPC Land on 5 September 2026, describes the system this way: "Most of the recommendations are more pulled for each viewer." "Each viewer gets a different ranking." "No metric on its own is a good indicator of value." The formulation that has grown up around those quotes, inside the company and outside it, is that there is no algorithm, there are audiences.

As physics, it is accurate. Ranking is computed per viewer and per request, and both published papers describe it that way. There is no single ordered list of videos held anywhere in the system, and no score attached to a video that exists independently of a particular person being served it.

As accountability, it is evasive. A system described as an aggregate of audiences has no object inside it to audit. The framing turns an unauditable system into an unfalsifiable one, steps around the manually tuned weights the 2019 paper says are there, and relocates every flop to the creator who made it. PPC Land noted on 5 September 2026 that the interview published no methodology and that the interviewer was a YouTube employee who posed no challenges, and this book adopts that characterisation.

Both halves are true at the same time, which is why both are printed. As of September 2026, no external audit route to those combination weights existed, and the only public statement about them remained one line in a conference paper from 2019 saying that somebody sets them by hand.