How the Facebook Reels algorithm works: Meta's own words and our data
What Meta itself says about how Facebook Reels are picked and ranked, read November 22, 2026, and what 5,919 Facebook posts show about how uneven it is.
By Abondy ResearchPublished 12 min read
In short: Meta says Facebook Reels are picked for each person by an AI system that predicts what they'll want to watch. Its Facebook Reels system card, updated June 19, 2026, describes four steps: gather candidate reels, narrow them to "approximately 10-100 of the most relevant reels" with a lightweight model, predict things like how long you'll watch and whether you'll like or hide each one, then rank them by a relevance score. Many of those reels come "from creators you may not follow." Meta publishes the predictions and signals it uses, not their weights. In our data that per-person ranking showed up as wide spread: versions of the same video on different accounts landed anywhere from 0 to several hundred views.
Most pages about the Facebook algorithm list ranking factors nobody can check. This one sticks to what Meta has published itself, on its Transparency Center, its creator blog and its earnings calls, quoted and dated, all read November 22, 2026. Then it sets that against what 5,919 Facebook posts made through Abondy actually got. For the plainer question of whether Facebook Reels are worth posting at all, see are Facebook Reels worth it. The same approach for YouTube is in how the YouTube Shorts algorithm works.
How does the Facebook Reels algorithm work?
Meta's system card says "The content you see on your Facebook Reels is selected, ranked and delivered to you by an artificial intelligence (AI) system," in which "multiple machine learning models work together." It adds that "These models and their input signals are dynamic and they change frequently." The card lays out four steps:
| Step | What Meta says happens |
|---|---|
| Gather inventory | "the system collects all potential reels you may be interested in," including reels "from people or accounts you follow, or reels that are similar to those you've recently interacted with," and from accounts "similar to those you've followed or interacted with" |
| Read signals | It looks at signals such as "the length of the reel, similarity to other reels and how much the reel matches the content you tend to interact with." Then "A lightweight model is run to select approximately 10-100 of the most relevant reels for you," and "certain integrity processes" are applied |
| Make predictions | Models "make predictions about content you'll find most relevant and valuable" |
| Rank reels by score | "the system calculates a relevance score for each reel and puts them in order by this score" |
Two points from the card matter most for a brand. First, Reels is a recommendation feed: Meta describes it as "content you may be interested in from creators you may not follow," and says it "might also be cross-app recommended content from Instagram." Second, every step is about the viewer. The system predicts what this person will do with this reel, so the same video can be shown widely to some people and not at all to others.
What signals does Facebook use to rank Reels?
Meta lists "some of the significant predictions" its Reels system makes, and the signals behind each, split across three recommenders on the card. It doesn't say how they're weighted. Here are some of them, as the card words them:
| Meta predicts | Signals Meta lists for it, in part |
|---|---|
| How long you're likely to watch a reel | "How many times the reel has been watched to completion by others"; "How many times the reel has been viewed by others"; "The author of the reel"; "How often the reel has been skipped by others" |
| How likely you are to watch a reel to completion | "How many times the reel has been watched to completion by others"; "The length of the reel" |
| How likely you are to watch a reel rather than skip past it | "How long you've watched reels recently"; "The authors whose reels you've watched" |
| How likely you are to like a reel | "The reels you've recently liked, watched, shared or commented on"; "Whether the reel is original content" |
| How likely you are to share a reel to WhatsApp, or via Messenger | "How many times the reel has been watched to completion by others"; "The authors whose reels you've interacted with" |
| How likely you are to ask to see less content like this, or to hide a reel | "How much the reel's topic matches topics you tend to watch" |
| How likely you are to start a viewing session after clicking a reel | "How long others have typically watched the reel"; "The length of the reel" |
Read across the list and four kinds of input repeat: how other people watched the reel (views, completions, skips), what this viewer has watched and engaged with lately, who made the reel, and the reel itself (its length, its topic, and "Whether the reel is original content"). Watching to the end and not skipping come up again and again. Likes and shares are predictions too, but they sit among many others.
Meta's page on Feed ranking, updated June 11, 2025, covers posts from people and Pages you're connected to rather than Reels, but it explains the approach behind both: "We use thousands of different signals," "The Feed ranking system has over 100 different prediction models," and "no one prediction is a perfect gauge of whether a post is valuable to you." Anyone who tells you the single factor that decides Facebook reach is going beyond what Meta has published.
Can viewers change what Facebook Reels shows them?
Yes, and that feeds back into ranking. The system card lists controls: hiding a reel "helps the system show you fewer similar reels," while saving or sharing "can help the system show you similar content." For results that aren't personalized, Meta points people to search: "your search results will be based on the search term you enter." On its July 29, 2026 earnings call, Meta said "on Facebook, we launched 'Shape Your Feed'" as a way of "giving people more direct control of the content they see."
What did Meta change in the Reels algorithm in 2026?
Meta's earnings calls describe its ranking work in some detail. From its third-quarter call on October 21, 2026:
- New posts get found sooner. Meta said it "completed the global launch of a new generative retrieval architecture" for Facebook Feed and Reels, "helping our systems surface relevant new posts before they've accumulated significant engagement." It's "now rolled out to roughly one-quarter of unconnected Feed views and half of Reels views and increased Reels watch time by 1%."
- More of each viewer's history is used. "On Facebook Reels, we doubled the amount of history our ranking model can consider, driving a 1% increase in Reels watch time in the US."
- Reels watching is growing. "On Facebook, Reels time spent reached all-time highs both globally and in the US, with both growing double-digits year-over-year."
On the second-quarter call, Meta said Facebook video time spent "increased 9% globally year over year and over 10% within the US & Canada, where it was driven by ranking improvements." These are Meta's own descriptions of its results; we can't check them.
The first change is the one to note if you post new content often. Meta says it's trying to surface relevant new posts before they've gathered engagement, rather than waiting for likes and shares to pile up.
What does Facebook push down?
Meta's Types of Content We Demote page, updated July 2, 2025, says Facebook reduces the distribution in Feed of content it sees as "problematic or low quality," including clickbait links and engagement bait. It says reduced distribution "may vary depending on the number of times the poster or commenter has violated our rules previously."
On reused content, Meta's creator post Combating Unoriginal Content (July 14, 2025) says accounts that "improperly reuse someone else's videos, photos or text posts repeatedly" will "receive reduced distribution on everything they share," and that "if our systems detect duplicate videos on Facebook, we will reduce the distribution of the copies." The same post gives advice that bears on Reels:
- "Pages and profiles receive the greatest distribution when they consist primarily of original content that you filmed or created."
- "avoid super short videos that offer viewers little value."
- "Avoid using visible third party watermarks and content that is visibly recycled from other apps or sources."
- "captions with no links, minimal use of capital letters, and no more than five hashtags perform best."
How these rules apply when one video goes out on several apps or accounts is covered in does cross-posting hurt reach.
What do our Facebook posts show about the algorithm?
Meta ranks each reel for each viewer. Our data can't see inside that, but it can show what it looks like from the outside. We looked at 5,919 Facebook posts made through Abondy from August 29 to November 12, 2026 whose views our reader got back. The median post had 24 views and the 90th percentile 845; 9.0% passed 1,000. For 901 more counted Facebook posts (13.2%) our reader never got a number back, often because of a login wall, so they're left out.
The same video landed very differently on different accounts
Abondy's video model makes many versions of one upload, and each goes out on a different poster's account. For the 85 uploads with at least 20 read Facebook posts (2,549 posts in all, a median of 25 different accounts per upload):
| Inside one upload | Median across the 85 uploads |
|---|---|
| 10th-percentile post | under 1 view |
| Median post | 12 views |
| 90th-percentile post | 368 views |
| Best post | 863 views |
In 46 of the 85 uploads the 10th-percentile post had under 1 view, and in all 39 others the 90th-percentile post had at least ten times its views. Measured on a log scale, only 12.4% of the spread in views lay between uploads; the rest was between posts of the same upload. So which upload a post came from accounted for little of the spread. That fits Meta's account of ranking each reel for each viewer, though our data can't show why one version took off and another didn't.
An account's past posts said little about its next one
For the 170 Facebook accounts with at least 10 read posts (4,603 posts), we split each account's posts into odd and even by time and compared the two halves. The rank correlation between an account's two medians was 0.30, which is weak. Only 11.8% of the spread in (log) views lay between accounts. Account medians ranged widely, from 4 views at the 10th percentile to 288 at the 90th, but a single account's posts varied far more than that.
The author still showed up in the numbers. The 16 posts that passed 10,000 views came from 10 accounts. Those accounts' other 445 posts had a median of 244 views and 20.4% passed 1,000, against 24 and 9.0% for all our Facebook posts. Meta lists "The author of the reel" among its signals; our figures are consistent with that but don't prove it, since those accounts may differ in other ways.
Results moved a lot over time
| Facebook posts made in (New York time) | Posts read | Median views | Reached 100+ | Passed 1,000 |
|---|---|---|---|---|
| September 2026 | 931 | 274 | 54.4% | 24.0% |
| October 2026 | 4,043 | 52 | 47.9% | 7.6% |
| November 1 to 12, 2026 | 940 | 6 | 18.7% | 0.4% |
The five posts from August 29 to 31 are left out of the table. The fall held inside the largest campaign too: its September posts had a median of 394 and its October posts 27. We don't know why. It isn't post age: November posts were last read a median of 17.8 days after they went up, and at the same age of seven days September posts had a median of 248 views against 5 for November's. It isn't the posts we couldn't read, since a smaller share went unread in November (11.0%) than in September (15.4%). It isn't new posters, because posters in their first two weeks and returning posters both fell. And the same accounts fell: posts from the 45 accounts with at least five read posts in both October and November went from a median of 108 views to 5. Campaigns and videos changed over those weeks, and Meta reported ranking changes in the same quarter, so we can't pin it on any one thing. What the table does show is that last month's Facebook results are a weak guide to this month's.
How should a brand work with the Facebook Reels algorithm?
From Meta's own pages and the patterns above:
- Judge Facebook on many posts, not one. With this much spread inside one video, a single post tells you little. Medians and shares past a threshold across dozens of posts tell you more.
- Make new versions instead of re-uploading one file. Meta says it reduces the distribution of duplicate videos, and its Reels card lists "Whether the reel is original content" as a signal. Changes need to be real; Meta says "simply stitching together clips or adding your watermark does not qualify." YouTube made a similar change for Shorts on October 1, 2026, covered in YouTube's Shorts reused content policy.
- Give people a reason to keep watching. Most of the predictions Meta lists are about watch time, completion and skipping.
- Keep captions plain. Meta's own advice is no links, few capitals and no more than five hashtags.
- Don't chase engagement with bait. Meta lists engagement bait among the content it demotes.
- Re-check results each month. Our medians moved from 274 to 6 between September and early November.
Where Abondy fits
Abondy gets a brand's video posted by many real people on their own TikTok, YouTube Shorts, Instagram Reels and Facebook accounts. Its video model, Flagship 1.2, makes a separate version for each post rather than the same file again (how it works), and the brand sees every post's views. That's a way to get the volume of posts that Facebook's spread calls for. A post costs 12 credits, about 16¢ to 20¢ on a plan, and posts under 100 views 24 hours after they're handed in aren't charged (pricing). We make Abondy, so weigh this section with that in mind.
How we counted
- Meta's statements: quoted from the Transparency Center, Meta for Creators and Meta's earnings call transcripts linked above, read November 22, 2026. Meta updates its system cards often; the Reels card we read was dated June 19, 2026.
- Sample: 6,820 counted Facebook posts made from August 29 to November 12, 2026 (New York time), pulled November 22, 2026. Every figure uses the 5,919 posts our reader got a number back for, by 573 posters on 550 accounts across 20 campaigns. Mostly gaming clips, largely Roblox, on small accounts. One brand's campaign accounts for 43.2% of these posts, and Abondy's own campaigns for 23.2%.
- Versions of one upload: grouped by the upload each version was made from; uploads with at least 20 read Facebook posts. "Spread between uploads" and "between accounts" are the share of the variance of log(1 + views) explained by the group means.
- Views: each post's count when we last read it. Flagged, rejected or deleted posts, Abondy's test accounts and posts with views confirmed as bought are left out.
- Associations, not proof: posts weren't randomly assigned to accounts, months or videos.
Questions people ask
How does the Facebook Reels algorithm work?
Meta says an AI system gathers reels you might like, narrows them to about 10 to 100 with a lightweight model, predicts how you'd respond to each (watching, finishing, liking, hiding) and ranks them by a relevance score. It's personalized to each viewer.
Do you need followers to get views on Facebook Reels?
Not necessarily. Meta describes Reels as "content you may be interested in from creators you may not follow," so reels are recommended beyond a Page's followers. Who made the reel is one of the signals Meta lists, though.
What does Facebook look at most when ranking Reels?
Meta doesn't publish weights. The predictions it lists lean heavily on watching: how long people watch, whether they finish, and whether they skip. It also lists likes, hiding and "see less" requests.
Does Facebook show reposted Reels to fewer people?
Meta says it reduces the distribution of duplicate videos and of accounts that repeatedly reuse other creators' content without permission or meaningful changes. Your own content can be shared across your own Pages and profiles, Meta says.
Why do some of my Facebook Reels get thousands of views and others almost none?
Because ranking is done per viewer and per reel. In our data even versions of the same video on different accounts ranged from 0 to several hundred views, and an account's earlier posts said little about its next one.
Read next
- Short-video data · 17 Nov 2026 · 9 min read
Are Facebook Reels worth posting? What 5,806 posts got
Our Facebook posts had a median of 24 views, yet all 4 posts past 100,000 in our data were on Facebook. What 5,806 posts say about Facebook Reels.
- Short-video data · 22 Nov 2026 · 11 min read
Does cross-posting hurt reach? One video on TikTok, Reels and Shorts
We tracked 12,736 video files posted to two or more apps: reach on the second app, and what TikTok, Instagram, YouTube and Meta say about reposts.
- Short-video data · 22 Nov 2026 · 10 min read
How does the YouTube Shorts algorithm work? YouTube's words, our data
What YouTube itself says about how Shorts are recommended, quoted and linked, and what 15,007 of our Shorts show about seed audiences and the first day.
