How the Instagram algorithm works: one app, many rankings

Instagram explains its ranking in plain words, but never the full recipe

Almost everything known about Instagram's ranking comes from Instagram and Meta themselves. None of it gives the weights or a full technical description of today's system. Here's what exists:

• Instagram's explainer: "Instagram Ranking Explained" by Adam Mosseri, head of Instagram, May 2023, naming the main signals and predictions for Feed, Stories, Explore and Reels. confirmed · 1
• Meta's system cards: one for Instagram Feed dated February 2022, then eight Instagram cards in June 2023, for Feed, Feed Recommendations, Stories, Explore, Reels Chaining, Search, Suggested Accounts and Notifications. confirmed · 2, 3
• Meta's engineers: how Explore's production system narrows and ranks posts (August 2023), and the more than 1,000 models behind Instagram's recommendations (May 2025). confirmed · 5, 6
• Instagram for Creators: short official posts on small-audience testing (April 2024), trial reels (December 2024), the 3-minute limit (January 2025) and originality for photos and carousels (April 2026), plus an undated FAQ. confirmed · 7, 8, 9, 10, 11
• Research: Facebook's 2019 paper on DLRM, an open-source model that shows the general type of ranking model, not what Instagram runs today. confirmed · 12

Dates matter: the most detailed explainer is from 2023, and Instagram says it adds and removes signals and predictions over time. confirmed · 1 Read the lists below as dated snapshots.

Each part of the app runs the same kind of funnel, with different inputs

Meta's engineers say each part of the app has its own funnel: gather candidates, rank them quickly, then rank the best more carefully (in Explore, the top 100). confirmed · 5, 6 Meta also describes a step before, understanding the content, and one after, learning from feedback. confirmed · 4

Step 1: understanding the post. For Reels, Instagram uses signals about what's inside the video, such as the audio track and visuals, plus popularity. confirmed · 1 Meta says its models try to understand a post's meaning across image, text, audio and video together. confirmed · 4 Instagram spots trends from audio and hashtags. confirmed · 11 When it finds identical copies, it recommends only the original. confirmed · 7 No published paper or official page lists what the models take from each video. I watch videos all day, every second, and I still can't see that. Nobody outside Instagram can.

Step 2: finding candidates. Each part of the app starts from a different pool. confirmed · 1

• Feed: recent posts from accounts you follow, plus suggested posts based on what and whom you've followed, liked or engaged with recently. confirmed · 1
• Stories: only accounts you follow, plus ads, minus stories that break the Community Standards. confirmed · 1
• Explore: mostly accounts you don't follow, found through posts you've liked, saved, shared and commented on. confirmed · 1
• Reels: mostly accounts you don't follow; Instagram gathers reels it thinks you might like, then orders them. confirmed · 1

Explore's engineers describe several candidate "sources" picking from billions of posts, from simple rules, such as trending posts, to machine-learning models, some updated in real time from your latest interactions. confirmed · 5

Step 3: ranking by predicted actions. For each candidate, Instagram estimates how likely you are to do several things with it. confirmed · 1 In Mosseri's words: "The more likely you are to take an action, and the more heavily we weigh that action, the higher up in Feed you'll see the post." confirmed · 1 The named actions differ by part of the app:

• Feed: roughly a dozen predictions, five watched most closely: spending a few seconds on the post, commenting, liking, sharing and tapping the profile photo. confirmed · 1 The top signals, in rough order: your activity; the post itself, including how fast it's gathering likes, comments, shares and saves; the account; and your history with it. confirmed · 1
• Stories: tapping into a story, replying, or moving on to the next. confirmed · 1 Signals: how often you view and engage with that account's stories, likes and DMs included, and how close you are to the author. confirmed · 1
• Explore: likes, saves and shares. confirmed · 1 Popularity signals, meaning how many people engage and how quickly, "matter much more in Explore than they do in Feed or in Stories." confirmed · 1
• Reels: resharing, watching all the way through, liking, and opening the audio page, a stand-in for whether you might make your own reel. confirmed · 1 Instagram also surveys people on whether a reel was worth their time or fun, and learns from the answers. confirmed · 1

How do several predictions become one order? Meta's 2022 Feed card says the system "combines these likelihoods into a single numerical score for each post" and weighs them "in the same way for everyone." confirmed · 2 The 2023 Explore post adds detail: each prediction gets a weight, and predicted unwanted actions, such as "see less", are subtracted. confirmed · 5 Raise the weight on likes, and the order will "pay more attention" to likes. confirmed · 5 Engineers tune those weights in experiments, trading one engagement goal against another. confirmed · 5, 6

What sits inside such a model? Facebook's 2019 DLRM paper shows the general recipe: turn categories, such as which person and which item, into learned lists of numbers, mix in other numeric inputs, and output the probability of one action, such as a click. confirmed · 12 It was tested on public ad-click data, not Instagram's. confirmed · 12 Instagram's Explore post describes models built the same way, turning users and posts into learned lists of numbers. confirmed · 5 That suggests DLRM is a fair picture of the model family, not a blueprint. inferred

Step 4: final rules reshape the scored list. confirmed · 1, 2, 5

• Feed avoids too many posts from one account in a row, or too many suggested posts back to back. confirmed · 1 In 2022, the cap was three in a row from one account. confirmed · 2
• Explore's last pass can drop or push down items on safety scores, and shuffle for variety, for example so one author doesn't appear twice in a row. confirmed · 5
• Under its Recommendation Guidelines, Instagram aims not to recommend some posts it allows, such as people fighting, or the use of tobacco, vaping products or pharmaceutical drugs. confirmed · 1 Repeat offenders can lose recommendation eligibility for a while, shown in Account Status. confirmed · 1
• Instagram aims to make reels "less visible" when they're low-resolution, watermarked, muted, bordered, mostly text or already posted on Instagram. confirmed · 1 Reels over 3 minutes aren't recommended to non-followers. confirmed · 11
• Accounts that repeatedly repost others' content without material changes (10 or more times in 30 days, under the 2024 rule) are left out of recommendations. confirmed · 7 Since April 2026, this also covers photos and carousels, without changing what followers see. confirmed · 10

Step 5: learning from what happens next. In April 2024, Instagram said every eligible post would first go to a small audience it thinks will enjoy it, followers or not. confirmed · 7 As that audience engages, the top-performing reels go to a slightly wider audience, then a wider one, and so on. confirmed · 7 Watching a whole video or liking a post updates how Meta's systems model your taste. confirmed · 4 "Not Interested" taps, reports and "is this post worth your time?" surveys also feed back into what gets recommended. confirmed · 1

What this means for your posts: earn the right actions from the right people

These are my deductions, tagged inferred. The facts under them keep their own tags.

• Give each post one job. Stories are shown only to people who follow you, while the Reels tab mostly shows accounts the viewer doesn't follow. confirmed · 1 So a story's job is conversation, and a reel is your main way to be found by new people. inferred In Feed and Stories, your history with each person (comments, DMs) is a named signal confirmed · 1, so replying is part of your reach there. inferred
• Pass the eligibility checks first. Instagram's small-audience test is for eligible content: original, no visible watermark, within its guidelines. confirmed · 7 A strong reel that fails one check may never reach anyone new, so keep reels to 3 minutes or less, sharp, with sound, without borders, not mostly text, and your own. inferred
• Make reels people finish and send. Two of the four Reels predictions Instagram names are a full watch and a reshare. confirmed · 1 An ending worth waiting for serves the first; a subject someone would send to one particular friend serves the second. inferred
• Judge a post per viewer, not in totals. Instagram's creator FAQ says what matters most for distribution is engagement per viewer (watch time, sharing, following, liking, commenting, profile visits) and how likely each viewer is to watch and engage. confirmed · 11 By that measure, a reel seen by 800 people and shared 40 times is doing better than one seen by 80,000 and shared 60 times. inferred
• Aim at the right people, not more people. Instagram warns that irrelevant hashtags can hurt engagement and then distribution, and lists clickbait, engagement bait and contest or giveaway promotions as limits on distribution. confirmed · 11 Anything that pulls in the wrong viewers lowers the per-viewer rate that decides whether a reel reaches a wider group. inferred
• Test on new people first. Trial reels go to non-followers first and show views, likes, comments and shares after about 24 hours. confirmed · 8 You can let Instagram share one with your followers automatically if it does well on views within 72 hours. confirmed · 8 That tests an idea on strangers before your followers see it. inferred Photos and carousels with added audio may also appear in the Reels tab. confirmed · 9

What nobody outside Instagram knows, and four common beliefs I checked

The public record stops at the recipe. Feed makes roughly a dozen predictions, and Instagram names five. confirmed · 1 No published paper or official page gives the weights for any part of the app, the size of the first test audience, what counts as "top performing", or how fast a post's score updates. Nobody outside Instagram knows these, me included.

• "There is one Instagram algorithm." Contradicted. Instagram says each part of the app uses its own confirmed · 1, and Meta's engineers describe more than 1,000 models. confirmed · 6
• "Instagram hides posts so you'll buy ads." Denied, but only on Instagram's word: it says it doesn't suppress content to encourage people to buy ads. confirmed · 1 What's confirmed is narrower: a post can be allowed on Instagram but not recommended, and Account Status shows whether yours is eligible. confirmed · 1
• "Sends (shares by DM) matter most." Partly supported. Sharing is among the top predictions for Feed, Explore and Reels confirmed · 1, and part of the engagement per viewer that Instagram says matters most for distribution. confirmed · 11 No published paper or official page I could check ranks shares above watch time or likes, or says they count extra for reaching non-followers.
• "Big accounts always win on Reels." Changing, by Instagram's own account. Follower count was a named Reels signal in 2023. confirmed · 1 In 2024, Instagram said recommended reels had been ranked mainly on how an account's followers engaged, so the largest accounts often got the most reach. confirmed · 7 It also added a ranking input to give smaller creators more distribution. confirmed · 7 Its 2025 claim that creators of all sizes now have an equal chance confirmed · 9 can't be tested from outside.