What is in the score? A guess. For each video the app guesses: how likely is this person to watch it to the end, to like it, to share it, to comment. Each guess is a number between 0 and 1, like the chance of rain.
A like1 weight
A comment5 weight
A share10 weight
Made-up weights.
Then each guess is multiplied by a weight, and the results are added up. The weights say what the app cares about. A share might be worth ten times a like, because a share brings another person in. The weights are chosen by people at the company. Nobody else sees them.
Like: 0.5 times 10.5 points
Share: 0.1 times 101 point
A video: chance you like it 0.5, weight 1. Chance you share it 0.1, weight 10. What is its score?
points
Right. Half a point from the like and one point from the share. The share guess was small, but the weight made it most of the score. That is what a weight does.
A million people like you
↓→
What they did with videos like this
↓→
The guess for you
Where do the guesses come from? From everyone who came before you. Millions of people saw videos like this one, and the app knows what they did. Then it narrows: people who act like you, on videos like this. That pattern is the guess. The app has never met you. It has met a million people who look like you.
Set how likely you are to finish a video, and how much the company weights finishing. Likes add a fixed 0.8 in every frame.
From watching to the end0.2 points
From likes0.8 points
0.2 · 1Chance 0.2, weight 1: 0.2 from watching plus 0.8 from likes is 1.0 points.
From watching to the end1 point
From likes0.8 points
0.2 · 5Chance 0.2, weight 5: 1.0 from watching plus 0.8 from likes is 1.8 points.
From watching to the end4 points
From likes0.8 points
0.2 · 20Chance 0.2, weight 20: 4.0 from watching plus 0.8 from likes is 4.8 points.
From watching to the end0.5 points
From likes0.8 points
0.5 · 1Chance 0.5, weight 1: 0.5 from watching plus 0.8 from likes is 1.3 points.
From watching to the end2.5 points
From likes0.8 points
0.5 · 5Chance 0.5, weight 5: 2.5 from watching plus 0.8 from likes is 3.3 points.
From watching to the end10 points
From likes0.8 points
0.5 · 20Chance 0.5, weight 20: 10.0 from watching plus 0.8 from likes is 10.8 points.
From watching to the end0.9 points
From likes0.8 points
0.9 · 1Chance 0.9, weight 1: 0.9 from watching plus 0.8 from likes is 1.7 points.
From watching to the end4.5 points
From likes0.8 points
0.9 · 5Chance 0.9, weight 5: 4.5 from watching plus 0.8 from likes is 5.3 points.
From watching to the end18 points
From likes0.8 points
0.9 · 20Chance 0.9, weight 20: 18.0 from watching plus 0.8 from likes is 18.8 points.
The company raises the weight on watching to the end from 1 to 20. What kind of video wins now?
Ones people finish: short ones, or ones with a hook that keeps you.
Right. At weight 20 a 0.9 chance of finishing is 18 points against 0.8 from likes. The like part barely matters; the video that keeps you to the end takes the top.
Ones people like the most.
Not yet. Likes are still worth 0.8 in every frame. At weight 20, finishing dwarfs that, and the best-liked video loses to the best-finished one.
Nothing changes. A weight is just a setting.
Not yet. Set the chance to 0.9 and move the weight from 1 to 20: the score goes from 1.7 to 18.8. A setting that changes a score ten times over changes which video is first.
Move both controls.
A like1 point
A reaction, such as angry or love5 points
Facebook's weights from 2018, reported in 2021.
Real weights, once. In 2018 Facebook changed its weights so that a reaction, such as an angry face or a heart, counted five points and a plain like counted one. A comment counted more still. Reporters learned this in 2021 from papers a former employee took out of the company.
Watched to the endmeasured
vs
Glad you didnot measured
The app can measure whether you watched to the end. It cannot measure whether you were glad you did. What follows for the score?
It guesses the thing it can check, watching, and cannot guess the thing it cannot, gladness.
Right. A guess gets better by being checked against what happened. Watching leaves a record to check against. Gladness leaves none, so no guess about it can ever improve.
It guesses gladness from watching. If you watched, you were glad.
Not yet. That is an assumption, not a measurement. People finish videos they hate, out of habit or shock. The score cannot tell those from the ones they loved.
It leaves watching out, since it might mislead.
Not yet. Watching is the app's clearest signal and its best-paid one. It is in every score. What is left out is the thing with no signal, not the thing with a noisy one.
Each of these is a GUESS the app makes about you, or a WEIGHT the company sets. Which?
One card at a time. Tap the pile it belongs to.
Card 1 of 6
Chance you will share this video: 0.05.
A share counts ten times a like.
Chance you will watch to the end: 0.7.
Comments from friends count double.
Chance you will hide this post: 0.02.
Videos under a minute get a bonus.
Chance you will share this video: 0.05. → A guess about you
Not yet. A guess. It has a chance in it, and it is about one person's likely action on one video.
A share counts ten times a like. → A weight for everyone
Not yet. A weight. It has no chance in it, and it is the same for every person and every video on the app.
Chance you will watch to the end: 0.7. → A guess about you
Not yet. A guess. Seven in ten is a prediction about you, built from people who acted like you on videos like this.
Comments from friends count double. → A weight for everyone
Not yet. A weight. Doubling is a rule the company wrote; no prediction about anyone is inside it.
Chance you will hide this post: 0.02. → A guess about you
Not yet. A guess. Hiding is an action, and two in a hundred is the app's bet on whether you take it.
Videos under a minute get a bonus. → A weight for everyone
Not yet. A weight. A bonus for short videos applies to every short video for everyone, whatever any one person does.
Guessesabout you
vsvs
Weightschosen for everyone
Right. A guess has a chance in it and is about one person. A weight has no chance in it and applies to everyone. The score is guesses times weights, and only one of the two is about you.
A story openedcounted
vs
A story finishedcounted too
A news app wants people to read whole stories, not just headlines. Using this lesson, what would it change?
Raise the weight on reading to the end, so stories people finish rise above stories people only open.
Right. Weights are the lever the company holds. Put more points on finishing and the score starts to favour the stories people finish.
Raise the guess for every story.
Not yet. A guess is a prediction about a reader and cannot be raised by decree. Raising all of them equally would change nothing about the order.
Show fewer stories.
Not yet. Showing fewer changes where the cut falls, not which stories are above it. The order is set by the score, and the score by the weights.
Lesson complete
A post's score is the app's guess at what you will do with it, times a weight the company chose.