How Facebook Decides Which Reels to Show.

Meta publishes how Facebook Reels ranking works, and the published version is more useful than most of what gets said about it.

The Transparency Centre system card for Facebook Reels describes an AI system that gathers candidate reels, runs a set of machine learning models over each one, and then “calculates a relevance score for each reel and puts them in order by this score”.

Ten predictions sit behind that score. Exactly one of them is about a like. None of them is about a share or a comment. Four are framed around the viewer getting out: skipping past it, hiding it, asking to see less, leaving right away.

That is not the shape of a popularity contest. It is the shape of a system looking for a reason to stop showing your video.

The ten predictions Facebook makes about every reel

Meta describes the setup plainly: “Within one AI system, multiple machine learning models work together to deliver your experience.” The system first collects the candidates, including reels from accounts you follow and reels similar to ones you have interacted with. Then it predicts, for each reel:

1. How long you are likely to watch it
2. How likely you are to keep watching reels after opening one in full screen
3. How likely you are to click a reel in your Feed to view it in full screen
4. How likely you are to like it
5. How likely you are to watch it rather than skip past it
6. How likely you are to watch it to completion
7. How likely you are to ask to see less content like this
8. How likely you are to hide it
9. How likely you are to start a viewing session after clicking a reel, rather than leaving right away
10. How likely you are to move on to the next reel after watching one

Those ten feed the relevance score that orders your Reels. Meta does not publish how they are weighted against each other.

Only one of the ten is a like

Nine of the ten predictions are about watching behaviour or exit behaviour. One is about approval.

Read the list against the brief a typical brand writes for a Reel and the mismatch is obvious. The brief asks for something people will like, share and comment on. Two of those do not appear as predicted outcomes at all, and the third carries a tenth of the published list.

Shares and comments are not absent from the system. They appear on the other side of it, as signals describing what a viewer has already interacted with. Being an input that profiles the audience is a different job from being an outcome your video is judged on.

This is the practical correction to how most teams grade short form video. A reel that earns 40 likes and holds attention for eight seconds is doing better work than a reel that earns 200 likes from existing followers and gets skipped by everyone else. If your short form video strategy is built around approval metrics, it is optimising for the smallest item on Facebook’s list.

Four of the ten are framed around you leaving

Look at what the system watches for on the downside. It predicts how likely you are to skip past a reel, to ask to see less content like it, to hide it, and to click into a reel and then leave right away rather than start a session.

Two of those are deliberate. Hiding a reel and asking to see less are controls Meta offers the viewer, alongside saving, sharing and reporting. Somebody has to be annoyed enough to use them.

The other two are passive, and they are the ones that quietly do the damage. Nobody hides your reel. They simply do not stop scrolling, and they do not come back for another one.

That reframes the production problem. The question is not what will make someone react. It is what would make someone leave, and whether you have removed it.

What Facebook says feeds those predictions

The system card names the inputs behind each prediction. They fall into a handful of families:

• What you have interacted with recently, including reels you have liked, watched, shared or commented on
• How other people behaved with the same reel, including how often it was watched to completion and how many times it was viewed
• How recently and how much you have been watching reels generally
• Whether the reel’s topic matches content you have recently watched or clicked
• The author of the reel
• The length of the reel

The second family is the one brands forget. Facebook uses how other viewers treated your reel as an input to whether it shows it to the next person. Early watch behaviour is not only a result you read afterwards, it is an ingredient in what happens next.

Two signals most Reels briefs ignore

The author of the reel is named as an input right across the list, not on one or two predictions. Whatever the system has learned about your page is sitting inside the judgements it makes about every video you post.

That is an argument for topical consistency at page level, not just quality at video level. A page that posts a product clip, then a staff birthday, then a meme, then a customer story is asking the system to keep starting over on what it is. Running a Facebook and Instagram presence so the page reads as one thing is ongoing management work, not a campaign.

The length of the reel is named as an input on two predictions: whether you start a viewing session, and whether you move on to the next reel. Length is a packaging decision that belongs in the brief, not whatever survives the edit.

What the system card does not tell you

Three things are genuinely unknown, and pretending otherwise is how bad advice gets written.

Meta publishes the predictions and their inputs, but not the weights. There is no published answer to whether avoiding a hide is worth more than earning a completion, so anyone quoting you a ranked list of factors is guessing.

The card also says these models and their input signals are dynamic and change frequently as the system learns. Treat the list as the current published shape, not a permanent contract.

And this describes Facebook Reels. Instagram runs its own recommendation systems, documented separately, and the instinct to treat Meta as one algorithm is wrong often enough to cost you.

Brief against the skip, not for the like

Here is the change worth making. Add one line to every Reels brief before anything is shot.

Name the moment a viewer would skip, and say what removes it.

That single question does more than a page of creative direction, because it forces the team to defend the opening against a real behaviour rather than an abstract standard. Most Reels that underperform do not fail on quality. They fail because the first second gives no reason to grant the second one.

The rest of the brief can stay as it is. What changes is what the brief is arguing against.

What changes in the edit

The opening carries four predictions at once: watch duration, skip avoidance, session start and completion. That concentrates where effort should go.

• Open on the subject, not the set-up. A logo sting or a slow establishing shot spends the exact frames the skip prediction is reading
• Show the thing being discussed in the first frame, rather than after a sentence explaining it
• State the stake early. A viewer needs a reason to stay before they get the payoff
• Cut the tail. A completion prediction is easier to satisfy on a tighter cut, and length is a named input elsewhere on the list
• Remove the middle lull. Plenty of reels lose people in the middle third, not the opening
• Keep it legible with the sound off, because a muted viewer who cannot follow it skips it

What changes in the packaging

One prediction is specifically about clicking a reel in the Feed to view it full screen. That is a separate job from holding attention once someone is already scrolling reels, and the cover frame decides it.

Cover frames get treated as an afterthought on Facebook far more than on other platforms. If the text on yours is unreadable at Feed size, it is not doing the job the system is measuring.

Topic legibility matters here too, since topic match is a named signal. A reel that is visibly about a subject is easier to match than one that could be about anything.

What good looks like

A team getting this right can answer four questions before a reel is shot.

• What would make someone skip this in the first second, and what have we done about it
• What is this reel visibly about, in one phrase, from the first frame
• How long is it, and why that length
• Does it belong to the same page identity as the last ten

They also grade differently after publishing. Watch-through and skip behaviour lead the review, likes are noted rather than celebrated, and a reel that held attention without much reaction counts as a win rather than a disappointment.

How NBK thinks about Reels ranking

Published ranking systems change, which is why chasing them post by post does not work. What survives a model update is a production standard that already optimises for attention rather than approval, applied consistently enough that the page itself becomes a signal worth having.

That is a systems problem more than a creative one. It lives in the brief template, the review standard and the way results get read, which is the same argument for running social as a system rather than a calendar of posts.

NBK’s team has shipped short form video at publisher scale, and the pattern holds across brands: the pages that grow are the ones where every reel defends the first second rather than decorating it.

Next step

If your Reels are well made and still not travelling, the constraint is usually the brief rather than the craft. NBK can help find where the system is losing attention and rebuild the standard behind it.