A phone lights up with a wall of unfamiliar faces and places, dozens of tiny squares packed edge to edge.
None of the accounts are ones you follow.
Somehow the mix still feels oddly tuned to you, a recipe here, a woodworking clip there.
That grid is called Explore, and it runs on math most users never see.
Behind every square is a ruthless elimination process most scrollers would never guess.
How the hidden sorting actually happens before you ever see a single post
Every time the Explore tab opens, Instagram’s system does not search the entire app for something to show you. Instead it runs a narrowing funnel, and Meta’s own transparency page lays out exactly how tight that funnel is, stage by stage, with real numbers attached to each cut.
The retrieval stage is responsible for selecting a set of candidate items that are relevant to the user’s interests, and at the end of the retrieval stage, up to 1500 media are fetched. Meta describes this using collaborative filtering, personalized PageRank, and two-tower sparse networks.
So the app is not choosing from everything you have ever liked. It is choosing from a pre filtered batch already trimmed before serious ranking begins.
Then comes a second, harsher cut. The early stage ranking stage is responsible for narrowing down the set of candidate items selected in the retrieval stage to a smaller set of the most promising candidates, and at the end of this stage, the top 100 media are passed to the next stage. Ninety three percent of the original shortlist never makes it past that gate.
The scene inside your own feed once the funnel finishes its work
Picture those surviving 100 posts lined up like finalists in a contest nobody entered. A heavier, more expensive model studies each one, weighing dozens of separate predictions at once rather than a single popularity score.
Those predictions include likelihood to linger, save or reshare a post, not just whether you glance at it. Meta’s Explore card lists predictions such as following you, viewing for more than 5 seconds, watching more than 95% of a video, saving, and resharing.
That is a different bar than counting likes. A cooking video earning a slow full watch and a save can beat a flashier clip that gets a quick tap, even with fewer raw views.
Meanwhile the same basic skeleton, gather candidates, strip rule breakers, predict behavior, then sort, runs across Feed, Stories and Reels too. Meta’s own Transparency Center documents at least five separate AI ranking systems for the app: Feed, Feed Recommendations, Stories, Explore and Reels, each with different input signals.
Where this comes from and how recently the company opened the books
None of this is guesswork or a leaked internal memo. The company publishes these mechanics directly on its transparency center page for Instagram Explore, part of a broader push toward public system cards that began years ago and keeps getting updated. The company has also opened the reverse side of this process. Users can now see and edit the topic list the ranking model builds about them, through a feature called Your Algorithm.
Instagram launched Your Algorithm on December 10, 2025, letting users see and adjust the topics influencing their Reels recommendations, and the company expanded the tool to Explore in April 2026. It was then extended to the main feed as of June 2026, bringing the controls to the surface where most users spend the bulk of their time in the app, with edits in one place rippling everywhere a person scrolls.
It is a rare case of a tech company showing its homework, similar in spirit to recent moves around AI tools reaching schools with published guidance rather than a black box.
Where the funnel breaks down and why perfect tuning never quite arrives
Even the company’s own documentation admits the limits here. Its cards repeatedly stress that multiple machine learning models work together within one system, and these models and their input signals are dynamic, changing as the system learns over time.
So the funnel described today is not fixed. It shifts as the models get retrained, meaning the exact numbers could move without any announcement.
There is also a cold start problem baked into the design. Brand new accounts have no engagement history yet, so retrieval has to guess relevance from almost nothing. That tension shows up constantly in creator complaints about unpredictable reach, echoing a pattern seen whenever platforms automate judgment calls once handled by a person watching closely, the kind of shift also visible in automated workflow tools now making decisions on their own.
What the numbers mean for anyone who scrolls Explore without thinking twice
None of this changes what the tab looks like on a Tuesday night, a dense grid of strangers’ posts that somehow still feels personal. But knowing a shortlist of 1,500 gets cut to 100 survivors before any final ranking helps explain why Explore can feel so oddly specific.
It also reframes what a single tap of Not Interested actually does. That tap does not just hide one post, it adjusts future weighting that decides who gets pulled into next time’s shortlist.
For ordinary users the practical upside is the newer editing tools, a simple topic list anyone can open, trim or expand without touching a settings menu. And for a company that spent years defending opaque feeds, publishing the actual funnel numbers is a modest but real shift toward treating curious users like they deserve an answer.