How the Instagram algorithm works in 2026, according to Meta's own system cards

There is no Instagram algorithm. Meta documents nine separate ranking systems for Instagram, each with its own models, and it publishes what those models are trying to predict. Six of those nine documents were updated in 2026. The blog post that almost every guide to this subject quotes was published on 31 May 2023 and has not been rewritten since.

There is no Instagram algorithm. Meta documents nine.

Instagram has said for years that different parts of the app rank differently. Its 2023 explainer puts it plainly: "Each part of the app, Feed, Stories, Explore, Reels, Search and more, uses its own algorithm tailored to how people use it." What is less well known is that Meta publishes a separate technical system card for each one, in the Transparency Centre rather than on the Instagram blog, and keeps them updated.

That gap matters more than it sounds. The 2023 post describes signals in plain language and ranks them loosely by importance, which is why it is quotable and why it has been quoted into the ground. The system cards do something different: they name the individual predictions each model makes and list the inputs feeding them. They are duller, longer, and current. We did the same exercise for the LinkedIn algorithm, where the primary material is thinner and the guesswork in circulation is correspondingly worse.

What Meta says it actually predicts

The Feed card does not say Instagram measures engagement. It lists specific questions the models try to answer about you and a given post, each with its own inputs. Here are the named predictions on the Feed card, in Meta's words.

• "How likely you are to swipe through the whole carousel", informed by how many cards you usually view, whether the carousel was reordered in your Feed, and how many cards it has.
• "How likely you are to enjoy this post", informed by whether you take a positive action such as liking, commenting or resharing, and how much time you spend on it.
• "How likely you are to skip this post", informed by whether it is a photo, video or carousel, and how likely you are to skip it compared to other posts shown alongside it.
• "How much time you're likely to spend on the author's profile after seeing this post", informed by whether you visit the profile at all after seeing the post.
• "How likely you are to share a post with someone in a direct message", which has the most interesting signal list on the page and gets its own section below.

Read that list as a description of what the system is optimising and two things stand out. The first is that a skip is a modelled outcome in its own right, not merely the absence of a like. The second is that your profile is part of the ranking of your posts: how long somebody spends there after seeing a post feeds back into how that post is ranked. Most advice about Instagram treats the profile as a conversion page. Meta's documentation treats it as a ranking input. If that is right, your bio and your highlight covers are doing work in the feed, not only after it.

The LLM that reads your post

On the Feed recommendations card, which governs what non followers are shown, one of the named predictions is "How informative a post is". It has exactly one listed signal, and it reads: "This signal is purely generated by LLM to assess the post's content quality".

A language model reads your post and scores it, and that score is an input to whether people who do not follow you ever see it.

It is worth being careful about what that does and does not establish. Meta does not say what the model reads, whether it is the caption, the image, the audio, or all three. It does not say how heavily the score counts. It does not say what informative means to the model. What it does establish is that a subjective judgement about content quality is being made by a machine that reads content, rather than inferred purely from how people behaved, and that this happens on the surface that decides your reach beyond your own followers.

The practical reading, and this is our inference rather than Meta's statement, is that captions have a job again. A post whose caption is three emoji and a hashtag block gives a language model nothing to assess. Whether that shows up as a lower score is not something we can prove, and it is the direction the documentation points. It also sits awkwardly with the decade of advice to keep captions short, which was written for a system that did not have a model reading them. The related question of whether hashtags still do anything is one we went through separately in do hashtags still work, and the answer there has moved for similar reasons.

What makes a post get sent in a DM

Sends are the signal every creator on Instagram has been told to chase, usually on the authority of a Mosseri video. The Feed system card is more specific than any of that advice, because it lists what the model uses to predict a send.

• "Whether the post is informative content (news, product reviews, tutorials, how-tos)". Meta names four categories in a parenthesis. That is unusually concrete for a document of this kind, and it is the same word, informative, that the LLM prediction above is scoring.
• "How recently the post was created (within 90 days)". A ninety day window appears explicitly, which puts a number on how long a post stays a candidate.
• "Whether the post is a repost or organic repost". Reposting is a modelled property of the post, not a rumour about a penalty.
• "Whether you open the post's share sheet" and what you do next, which is the behavioural half of the same prediction.

Two of those four are about what the post is rather than how anybody behaved. That is the finding worth taking away from this article, because it cuts against the standard advice that the algorithm only sees engagement. On the evidence of Meta's own documentation, the system forms a view about your content's type and quality before very many people have reacted to it. What you make is a signal, not only how it performs.

Reels: the prediction is that you will leave

The Reels chaining card, which decides what plays next, names "How likely you are to watch less than 3 seconds of a reel" as one of its predictions. Its inputs include how many times viewers watched at least three seconds of that reel anywhere on Instagram, how many times it was dismissed, "How many times the post has been skipped within 2 seconds of being opened", and how often it has been viewed with the sound on.

Framing matters here. A system that predicts watch time is asking how good your reel is. A system that separately predicts a sub three second bail is asking how bad your opening is, and treating that as its own thing. The first two seconds are not a growth hack in this reading, they are a documented model input. Sound is in there too, which is a quiet argument against the habit of publishing reels that make no sense muted and no difference unmuted. If Reels are your main output, what goes in the first two seconds is worth more attention than the schedule, and we have a piece on writing them.

The Explore card runs the same logic from the other end, naming "How likely you are to watch more than 95% of a video" as a prediction, alongside a signal counting how many other people did. Explore also documents four separate recommenders working together, labelled Recommender One through Four, which is a reminder that even one surface is not one model.

The signals nobody mentions

Reading nine cards end to end surfaces a set of inputs that appear in none of the guides currently ranking for this term. None of these is a growth tip. They are simply documented, and their absence from the discourse tells you how much of it is recycled.

• "How much time spent a viewer on an author for the past 84 days". An eighty four day window, on Feed recommendations. Twelve weeks of accumulated attention to you, as a number.
• "The number of times other viewers went to the post author's profile page after viewing the post". Other people's profile visits, feeding the ranking of your post to a new viewer.
• Content age buckets. Feed recommendations counts how many impressions you spend on media aged one to three days and separately fourteen to twenty one days, which implies the system models your appetite for fresh versus older content.
• "How many times you've seen a short, squared reel in Explore". Aspect ratio appears as a signal by name, which is a small argument for getting the post size right rather than letting the crop happen.
• "How likely it is that you and the story's author are family or good friends", on the Stories card. A modelled relationship category, not a generic closeness score.
• Sound on views. Counted separately on both the Reels and Feed recommendation cards.

What none of this tells you

It would be easy to read the last two thousand words as having cracked something. It is not that, and the limits are worth stating as plainly as the findings.

• No weights, anywhere. Meta lists which signals feed which prediction and never how much any of them counts. Every article you have read assigning percentages to Instagram ranking factors made those numbers up.
• Not the full list. The cards say these are some of the significant predictions. There is no claim of completeness and no reason to assume one.
• It changes. Meta states the models and their input signals are dynamic and change frequently. A card updated in June 2026 describes June 2026.
• Predictions are not instructions. Knowing that a model predicts DM shares does not tell you how to earn them, and the four content types in the parenthesis are what the model associates with sends, not a content strategy.
• None of it is account specific. These systems are personalised per viewer. Nothing in the documentation describes your account, and your own Insights remain a better guide to your own audience than any of it.

What to actually change on Monday

Three things follow from the documentation with reasonable confidence, and they are unglamorous.

Write the caption for a reader and a model. A language model is scoring how informative your post is, and the four content types Meta names for DM shares are news, product reviews, tutorials and how-tos. If your caption gives neither a person nor a model anything to work with, you are opting out of a signal that costs nothing to supply. This is the same shift that social media SEO describes for search, arriving inside the feed.

Earn the first two seconds separately from the rest. The sub three second bail is its own prediction with its own inputs, including skips inside two seconds. Whatever else a reel does, it has to survive that window, and a slow establishing shot spends the one thing the model is watching for.

Treat your profile as part of your reach. Time on your profile after somebody sees a post is a documented Feed prediction, and other viewers' profile visits feed recommendations to new people. That makes the profile a ranking surface rather than a landing page.

And keep the whole subject in proportion, because this is the part the algorithm content industry never says. Nothing in nine system cards changes the fact that the account publishing three considered posts a week beats the account publishing one a month with perfect signal hygiene. How often to post and what to post about remain the two questions with more leverage than any of this, when to post matters less than either, and if the honest problem is that nothing gets published at all then post ideas will move your reach further than a reading of Meta's documentation. Including this one.