GuideDistribution36 min read

You are a cold item: what Instagram's own documents say about a new post

Everyone repeats that Instagram tests a new post on 500 people. That number is in no Meta document we could find. The 500 Meta does publish counts posts, not people.

Aubrium ResearchEditorial · 13 August 2026

You posted at 6pm because someone said 6pm. You used the audio that was on everything last week. Four hours later the post has a couple of hundred views. The one before it did thousands. You look at the two side by side. You cannot see what is different.

So you go looking for a reason. One is waiting. Instagram showed the post to a test batch of about 500 people. They did not bite. The post got throttled. Everyone knows this. It is in hundreds of videos. It turns up every time someone asks why their reach fell off a cliff.

We went looking for where that claim came from. Meta has not written it anywhere we could find.

Here is what we read. Meta puts out a page for each part of Instagram. Each page explains how that part decides what to show you. Meta calls them system cards. We read four of them. We also read old copies of one card. The Internet Archive saves web pages on a date, so anyone can go back and read what a page used to say. Its copies of this card go back to 2023. And we read what Adam Mosseri wrote about all of this. He is the head of Instagram.

The number 500 is in Meta's writing. It is there once, in the five documents we read. And there it counts something else. It is how many posts get scored for one person's feed. It is not how many people see yours.

That is the shape of the whole problem. Creators worry about this more than anything. And the most repeated explanation for it is a real number attached to the wrong thing. Almost no one has gone to check.

The strange part: no one needed to invent one. Meta publishes a lot about how this works. It printed a scoring formula with a minus sign in it. The writing is dry. It sits in engineering blogs and transparency pages no one reads. It answers the question better than folklore does.

It also answers it in a way that is harder to be angry about. Instagram does take posts down for breaking its rules. It says it shows some kinds of reel less, too. Set those aside. An ordinary post is not being held back. It arrived at a stage built for sorting things that already have a history. It had none. The people who build these systems call every post an item. An item with no history yet has a name. It is cold.

What follows is a reading of what Meta and Instagram publish. It covers Instagram. It stops where those pages stop. Every source below is linked, dated and quoted.

A thousand models, not one

Start with the word. "The algorithm" is one noun. So you picture one thing. One thing with one opinion about you. Something you could please, or offend. That word does more damage than any false claim here.

The people who build it do not describe it that way. Three of their words turn up in the next sentence they wrote, so take those first. A model is a piece of software that scores things. Ranking is what happens to the scores: they put posts in an order, for one person. A classifier is software that sorts a post into a category, such as news or a tutorial. The flattest version of all this comes from the man in charge:

Instagram doesn't have a singular algorithm that oversees what people do and don't see on the app. We use a variety of algorithms, classifiers, and processes, each with its own purpose.Adam Mosseri, Instagram Ranking Explained, 31 May 2023

That is from Instagram Ranking Explained, published 31 May 2023.

Two years later Meta's engineers put a number on "a variety". Their post of 21 May 2025 is called Journey to 1000 models. ML in it is short for machine learning, which means software that learns from examples instead of being told the rules. The post says Instagram has scaled to "over 1000 ML models" without losing quality or reliability.

A thousand models is not a fact you can use on its own. How they are arranged is. The same post says how. First it gathers posts it might show you. Those are its candidates. Posts it is thinking about, before it decides. Then a cheap scoring pass runs over them. Then an expensive one. Meta's engineers call that shape a funnel. Wide at the top, narrow at the bottom. What puts the passes in that order is what each one costs to run:

We operate on fewer candidates as we progress through the funnel, as the underlying operations grow more expensive.Engineering at Meta, Journey to 1000 models, 21 May 2025

Two things fall out of that sentence. The rest of this post leans on both.

The first is that the sorting happens separately in each part of the app. Your feed is one part. The Explore tab is another. So are the Reels tab and Search. The people who build this call each one a surface. Meta publishes these pages on a website of its own, the Transparency Center. Its index of ranking explanations lists them. Ten sit under Instagram. Nine are parts of the app. The tenth is Threads, which is not even Instagram. Feed has its own card. So do Stories, Explore, Search and Reels Chaining. Chaining is Meta's word for the run of reels that keeps arriving once you open one. Each card has its own steps. Each has its own list of guesses about you.

The second is that the sorting happens in layers, and the layers do not cost the same. Cheap models look at many posts. Expensive models look at few. Cost is why the list narrows at all. Cost sets the widths in the next section.

So there is no single thing to please. There is no single place where a post wins or loses. A reel that never showed up in the Reels tab was scored by one model. A different model decided whether your own followers saw it, against a different list of guesses. When someone tells you what "the algorithm" wants, ask which of the nine they mean.

Why the funnel narrows

Every step here costs money to run. That one constraint gives the funnel its shape. Meta describes that shape twice, three years apart. The two descriptions fit together. The engineering post gives the general case. The system cards give numbers. One part of the app at a time.

Take the general case first. Meta's engineers wrote about scaling Explore on 9 August 2023. A system this size starts with thousands of candidates, they say, and narrows to hundreds. They also say what it is narrowing from. The pool is "billions of items". Billions down to hundreds. For one person. Every time that tab loads.

Explore is the part Meta documents best. If a test-audience number existed anywhere, we expected it there. Its system card, stamped 22 June 2026, names three steps. It publishes a count on two of them.

Step one is the cheap sweep. The card calls it retrieval. Its job is to pull together posts worth a look. Fast and rough. It names three ways of doing that, and all three are ordinary for software like this. The first takes a post you liked and pulls the other posts that the same people liked. The second starts at the accounts and posts you have already touched, then steps outward. From you, to a post you saved, to the account that made it, to that account's other posts. The third writes you a short description in numbers, writes every post one too, and then only has to check which pair of descriptions match. Comparing two short descriptions is cheap. Reading every post from scratch is not.

What matters is where the sweep stops. The card says "up to 1500 media are fetched". Media is the card's word for posts. That is a ceiling, not a promise.

Then a cheap ranker scores whatever came back. A ranker is the software that scores posts and puts them in order. It passes the best hundred forward. Only after that does the expensive model run. Its description on the card gives it away. It is one model that guesses many different things at once. You can afford that on a hundred posts. Not on fifteen hundred.

Up to fifteen hundred in. One hundred out. Whatever else the sweep found, the expensive model never sees it.

Feed narrows in a different way, because its stock is different. The Feed card is about posts from accounts you already follow. It describes four steps. The system gathers candidates. It drops anything that breaks the Community Guidelines. A cheap model picks roughly 700 posts. Guessing models run over those. Then comes the step to hold on to:

Finally, the system calculates a relevance score for about 500 posts and puts them in order by this score.Meta Transparency Center, Instagram Feed AI system card

What that sentence counts is the whole point, and a later section turns on it. The 500 are posts. They are the posts being sorted for one person. Someone who just opened the app. They are not people. They are not an audience.

Reels Chaining is the narrowest of the three. It is also the only one whose counts do not narrow. A cheap model picks about 100 reels. About 100 reels get scored and ordered. That is what you would expect where the screen shows one reel at a time.

Every one of those numbers has one thing in common, and it is the thing the rest of this post turns on. They all count posts. And they count them for one person at a time. Feed sorts about 500 posts for you when you open the app. Then it starts again, from nothing, for the next person. So your post is never in one contest. It is a candidate in a separate sorting for every person who might see it. One person opens the app, and it competes against that person's pile. The next person opens the app, and it competes against a different pile. There is no single verdict on your post anywhere. There is no single place where one gets made.

One warning goes with all six numbers. It goes with every quote here too. Meta rewrites these pages and never says when. The 2023 engineering post and today's Explore card describe the same tab. They do not even agree on how many steps it has. You only find that by reading both. The numbers above are the numbers of their moment, and the moment is 17 August 2026.

The two different 500s

Here is the claim in the form it usually arrives. Instagram shows a new post to a test batch of about 500 people. What that batch does decides whether the post goes any further.

We read five documents against it. Four system cards and Mosseri's ranking post. The cards are for Feed, Explore, Reels Chaining and Feed Recommendations. Your feed carries two kinds of post. Ones from accounts you follow, and ones from accounts you do not. The Feed card covers the first kind. The Feed Recommendations card covers the second. Not one of the five gives a number of people a new post is shown to first. Not one gives any other test-audience size either. Not on any surface. Not anywhere in the five.

Mosseri's post is the one usually produced as the proof, so be exact about what is in it. It walks through Feed, Stories, Explore and Reels. It says Feed makes roughly a dozen guesses. It lists the most important Feed signals, in rough order of importance. A signal, here, is one fact about a post or a viewer that the software is allowed to read. Then it takes the complaints head on. That includes the one that costs money to answer: "we don't suppress content to encourage people to buy ads." It describes no test group. Not of any size. Not anywhere.

The number 500 does turn up in those five documents. Once. It is the Feed card sentence quoted above, about posts being scored for one viewer.

Building the dated record turned up something we have not seen written down anywhere. It makes the mix-up much easier to picture.

Go back to the oldest saved copy of the Feed card. It is dated 29 June 2023. It already carries today's sentence about 500 posts scored for one viewer. And it carries a second 500. The cheap model at the top of the funnel now picks roughly 700 posts. Back then it picked "approximately 500 of the most relevant posts".

So the page said 500 twice. Both times it meant posts.

The copy saved on 21 May 2024 still read 500 at that first step. Today's card reads 700 there. The scoring step has said 500 the whole time. One of the two numbers moved. The other never has.

What the record cannot do is settle it. This is where we stop. An oldest saved copy is a floor. It is never a publication date. It shows the wording existed by 29 June 2023. It says nothing about when the folklore started. We claim the "500 people" number has no source. We do not claim to have shown how it got there.

The score has a minus sign

"Engagement" is one word for a lot of different guesses. A formula shows the difference. Meta printed one.

The 2023 engineering post on Explore says what its models hand back. Guessed odds. "P(click), P(like), P(see less), etc." The P there stands for probability. The chance of a click. The chance of a like. The chance that a viewer asks to see less. Each one gets a weight. All of them get combined into one number. Meta calls it a value model. Then the post prints the formula. The click term is added. The like term is added. The see-less term is taken away.

That minus sign is the most useful thing in this post. There is a guessed chance that a viewer wants less of you. It carries its own weight. It comes off the top of the score. Whatever "engagement" means, you cannot only add to it.

Meta names the result in its own words. It is "our approximation of the value that each media brings to a user". It says the weights are tuned. One metric gets traded off against another. It does not print the weights. Not for Explore. Not for any other surface. Not in any document we read. So every claim about which signal matters most is unproven. That includes every reading in this post.

Now you know a negative term is in there. The lists read differently. Each card lists what its models try to guess about you. Three of those guesses point the wrong way. They are things no one would choose.

Feed guesses how likely you are to skip the post.
Explore guesses how likely you are to tap "Not Interested" on it.
Reels Chaining guesses how likely you are to watch less than three seconds of a reel.

Feeding that last one is a blunt input. In the card's own words: "How many times the post has been skipped within two seconds of being opened."

Two of the positive guesses are stranger still. Feed guesses "How much time you're likely to spend on the author's profile after seeing this post". So your post is scored on what it makes a viewer go and do. Somewhere else in the app. Feed also guesses about the first post you see. Will it keep you in your feed for the rest of the session? That judges a post by its effect on the rest of the feed.

So there is no single goal called engagement. There is a bundle of guessed behaviours. No two parts of the app carry the same bundle. Four different bundles is the argument against treating this as one system.

The numbers, and the AI

Almost nothing in the cards has a number. Two entries break that, and they break it in opposite directions.

The Feed card lists "How recently the post was created (within 90 days)" as an input signal. Ninety days is the only window Meta puts on a post's own age. That holds across all five documents we read. Hold on to that. Someone will quote you a first-hour rule.

The same card lists a second signal by its categories. Whether the post is "informative content (news, product reviews, tutorials, how-tos)". That one is an input, not a guess. It feeds a guess about the viewer. How likely they are to send the post to somebody in a message.

The other entry is on the Feed Recommendations card, stamped 29 June 2026. It is not a measurement at all. That card publishes a guess called "How informative a post is". Its whole list of inputs is one sentence:

This signal is purely generated by LLM to assess the post's content qualityMeta Transparency Center, Instagram Feed Recommendations AI system card

LLM is short for large language model. It is software trained to read and write text. Every other guess across the four cards is a guess about a viewer. Will they click, skip, share, stay. This one is a score on the post. And the thing doing the scoring is an AI reading it.

The same card counts one thing that is easy to read backwards. It counts how many times a viewer gets shown posts of a given age. Over a period of seven days. One to three days old. Then eight to fourteen. Then fourteen to twenty-one. Those are the card's own ranges, gaps and all. It names nothing between four and seven days. One showing of one post to one person is an impression. Impressions are what is counted here. The subject of that sentence is the person, not the post. It measures an appetite for fresh things. It says nothing about your post's shelf life.

The cold-start problem

Every guess above runs on facts about the post. The strongest facts are records. What people already did with the post. Meta says so itself. It says so while explaining something else. Why the cheap sweep is not allowed to use them:

the model can't consume user-item interaction features (which are usually the most powerful) because by consuming them it will lose the ability to provide cacheable user/item embeddings.Engineering at Meta, Scaling the Instagram Explore recommendations system, 9 August 2023

Unpack that, both halves. "User-item interaction features" means facts about what a person did with a post. Watched it. Liked it. Skipped it. Meta's own bracket calls those the most powerful facts it has.

The second half says why the cheap sweep cannot have them. A "user/item embedding" is that short description in numbers from the sweep's third trick. One for you, one for each post. "Cacheable" means it can be worked out once, stored, and reused for everybody. But a fact about you and this post exists only for that pair. It cannot be worked out in advance and it cannot be reused. So the sweep would have to do the work again. For every post, every time. That is the cost it refuses. It gives up the most powerful facts it has. On purpose. To stay cheap.

The expensive model further down has no such limit. The same post says it "can also consume the most powerful user-item interaction features". So the last stage leans hardest on history.

A post published a minute ago has none of that. No one has watched it. No one has skipped it, saved it, sent it on, or clicked away two seconds in. Look at the Explore card's list of signals. It is full of counts that begin "How many people have". How many have seen the post on Explore. How many watched more than 95% of the video. How many clicked on it. Every one is zero for a new post. And the model is not reading a low score. It is reading a blank.

The named signals split in two. On one side, what a model can read at once. On the other, what arrives later, only after someone has seen the post. For a new post the second side is empty. That is what cold start means. Scoring something brand new, when no one has done anything with it yet.

Five of them are readable at once. They are facts about the post and author. Photo, video or carousel. How many cards are in the carousel. How much text sits on the photo. Whether it counts as informative. How long the reel runs. The other five are counts of other people. There are no other people yet.

Nobody has done anything to your post yet. Some facts tell a strong post from a weak one. They get read at the expensive end of the funnel. For a post a minute old, they are not there at all.

The people who build these systems call that a fault. Not a verdict on the post. They are blunter about it than folklore is. Six researchers wrote Item-centric Exploration for Cold Start Problem. They presented it in Prague in September 2025. The venue is a yearly conference on recommender systems. It is run by the ACM, a computing society. A recommender is software that picks what to show you. It runs when you did not ask for anything in particular. Their opening sentence describes the same spot. From the platform's side of it:

Recommender systems face a critical challenge in the item cold-start problem, which limits content diversity and exacerbates popularity bias by struggling to recommend new items.Wang et al., Item-centric Exploration for Cold Start Problem, RecSys 2025

Their diagnosis is the line to carry away. Hunting for the best post for a person "can inadvertently obscure the ideal audience for nascent content". Nascent means brand new. The software asks which post suits this person. Turn the question around and it becomes: which people suit this post. A new post has no one asking that on its behalf.

Nobody has settled what to do about that. Two papers show how unsettled it is. The first found something backwards. To push a new item, you would push it at your busiest users. Send it to less active users and you "will statistically yield better performance" instead. That is Promoting Cold-Start Items in Recommender Systems. It ran in PLOS ONE on 5 December 2014. It is a simulation. It runs over the real purchase records of two Chinese retail sites. Not over a live system. Then the authors knock their own finding down. Switch the recommender to the other standard kind. That one matches people to people, not items to items. The advice flips.

The second is a 2019 survey of what deep learning has tried. It sorts the attempts into two families. Give the model more about the item itself, its audio, images and text. Or put people and items on one map, worked out from numbers. Then drop the new item onto it. Both are ways to score an item that has no history. Which is what has to happen to your post in its first minutes.

Instagram has said this bias existed on its own app. From its creators blog, 21 January 2025:

Content ranking on Instagram historically favored creators with larger followings. However, last year we announced updates to our recommendations of reels. Through this process, creators of all sizes have an equal chance of breaking through and reaching new audiences.Instagram for Creators, Finding success on reels in 2025, 21 January 2025

No measurement is attached to that. So we report it as a claim, not as a finding.

What the random tests showed

Real experiments have been run on this. They ask whether early success causes later success. Early success handed out for no reason at all. The evidence is unusually strong, and the reason is the method. The researchers picked who got the boost at random. Like a coin flip. That is what settles cause. You do not have to argue it.

Three studies get cited, usually second-hand. None of them makes the claim it gets used to support.

Salganik, Dodds and Watts built a fake music market. They published it in Science in 2006. Their 14,341 participants chose among 48 songs by unknown bands. Some were told nothing about anyone else's choices. The rest were dropped into one of eight parallel worlds. Each world counted its own downloads, from zero. Same 48 songs. Eight separate histories. The histories did not agree with each other.

They ran that whole design twice. The second run made each world's own download counts more prominent. So the two runs differ in one thing. How much of its own past a world could see. That difference is the result:

Increasing the strength of social influence increased both inequality and unpredictability of success.Salganik, Dodds and Watts, Science, 2006

The more each world could see of its own past, the more unequal it got. And the harder it became to pick winners in advance.

One sentence in that paper deserves more circulation than it gets. It is about quality. "The best songs rarely did poorly, and the worst rarely did well, but any other result was possible." Quality sets a floor and a ceiling. It does not set a rank.

The authors are just as flat about their own limits. Their experiment "is clearly unlike real cultural markets in a number of respects". They also expect real-world social influence to be stronger than the version they ran.

Muchnik, Aral and Taylor ran a random test on a live social news website (Science, 2013). They tampered with the vote counts on comments as those comments appeared. The lopsidedness is the finding. A fake up-vote stuck, and later readers piled more on top of it. A fake down-vote did not stick. The next readers corrected it. Their own summary of the size of it:

positive social influence increased the likelihood of positive ratings by 32% and created accumulating positive herding that increased final ratings by 25% on averageMuchnik, Aral and Taylor, Science, 2013

We could not read the paper itself. The published version sits behind a paywall. The abstract is free. That is the summary printed at the top of a paper. It gives no sample size. So we cannot give you the sample size behind the 32% and the 25%. One of the authors hosts a pre-publication supplement openly. It does carry the counts. It is also stamped "not for citation". So we do not report its numbers as the paper's. The abstract adds two limits. The effect depended on the topic. And on whether a rater was looking at the opinions of friends or of enemies.

Van de Rijt, Kang, Restivo and Patil did the version that is hardest to argue with (PNAS, 2014). They handed out success at random. To real people, on four live websites. Then they watched. Unfunded Kickstarter projects got a donation. Wikipedia editors were given an award. Young change.org campaigns got a dozen signatures. Epinions is a site where readers rate how useful a review was. There, unrated reviews were given the site's top rating: "very helpful". In the first round the four pools ran from 200 to 521 cases each. Recipients were drawn at random. Everyone else was left alone.

Every site moved the same way. Take Kickstarter, the cleanest of them. Of the projects left alone, 39% went on to attract more funding. Of the ones given a single donation by the researchers, 70% did. The other three sites moved in the same direction. All four gaps were statistically significant. A fluke that big is unlikely. The test statistics are in the Sources section rather than here.

Then comes the sentence this section exists for. Their finding, in their words: "success exhibited decreasing marginal returns, with larger initial advantages failing to produce much further differentiation." More is not proportionally better.

The authors put the shape of it in one line. It is the least convenient line in this literature for anyone selling a growth shortcut:

The per-donor effect of a single donation by a single donor on fundraising success was greater than that of four donations by four separate donors.van de Rijt, Kang, Restivo and Patil, PNAS, 2014

Why none of it is Instagram

None of the three experiments ran on Instagram. None ran on software that recommends things at all. They measure song downloads, ratings on a news site, funding of projects and endorsements of reviews. Each one is a web market. The next person to arrive sees a visible tally. So the cause runs person to person. Through people who can read the count. Instagram is not that. You cannot see how many people watched a post before you did. The count goes to the models instead. It reaches them as the signals the cards name. Weighted in ways no document we read publishes.

No published experiment we could find shows that early engagement causes more reach inside Instagram's own ranking software. Not a weak one. None.

Two things follow, and they point opposite ways. That is the honest state of the question. On one side, the mechanism is plausible. The expensive model can read those facts about what people did. A post with some beats a post with none. Meta itself calls those features "usually the most powerful". But that last step is our reading of Meta's documents. It is not a published finding. On the other side, the size of any such effect is unpublished. Its shape over time is unpublished. And the one random result that speaks to size at all found the returns flattening, not compounding.

The direction has a documented reason. The size has nothing behind it. So someone quotes you a multiplier for early engagement. They are quoting an experiment that does not exist. Ask them for it.

Trial reels: the one real test

Everything above says the same thing. You cannot see the score on a cold post. Mostly you cannot see what the software sees. There is one exception. It is a product. Instagram announced trial reels on 10 December 2024. It described them in one sentence: "Trial reels will be shown to non-followers first."

Start with who is allowed to run one. The Help Center page sets out the conditions. The account has to be public. A professional account needs at least 200 followers. A personal account needs at least 1,000. And a personal account is eligible at all. The page says so in as many words: "If you have a personal account, you must have at least 1,000 followers."

The clocks are Instagram's own. Your numbers arrive about 24 hours after you share. You read them where you watch your own reels. Say you switched the auto-share setting on when you made the reel. Instagram may then push the reel to your followers as well. That is its call, not yours. The setting carries forward to your future trial reels. Until you change it.

The 72 hours everyone quotes comes from one page. The other page has dropped it. Instagram's announcement of 10 December 2024 set a clock on it. Its condition: "if we determine it's performing well based on the views it receives within the first 72 hours". The Help Center describes the same auto-share today and gives no clock. There the reel goes out "if it's performing well soon after you publish it, based on signals like views and engagement". Two Meta pages, two wordings. Only the older one carries a number. And on the newer one, views are no longer the only thing named. We report both and do not average them.

The split from your followers is not total. The same December 2024 announcement says so, one sentence before the one about metrics:

Some followers may still see a trial reel in other places. For example, someone might share your reel with them in a direct message or on a page that shows reels with the same audio, location, or filter.Instagram for Creators, Trial reels, 10 December 2024

So a trial reel goes to people who do not follow you. That is what normally happens, not a wall. Instagram shows the reel to non-followers first. It does not fence your followers out. Which is the difference between a controlled showing and a controlled experiment.

Running one is cheap in the only currency Instagram talks about. The Help Center says trial reels "are evaluated independently in Instagram's ranking and don't impact the performance or ranking of your standard reels content". That is a statement about ranking. It is not a promise about the rest. The same page names one cost. A trial reel may take longer to get views. It is being shown to people who do not follow you.

There is one penalty. It is stated outright. A trial reel "may get limited reach if Instagram detects that you've previously shared the same content". Mosseri's 2023 post names reels already posted on Instagram. They are among the things the company aims to show less of. So testing something you published before means testing a handicapped copy of it.

Treat trial reels as a measuring tool. Not a growth tactic. What you get is one post's showing against a mostly unconnected audience. On a clock Instagram published. Without spending your own followers to find out. What you do not get is the score. Or the weights behind it. Or any comparison with someone else's post.

The limits

Here is what would prove this post wrong. Say Meta publishes a page describing a fixed test audience. Then everything above about the "500 people" claim is wrong. The correction goes in this post, with the date we caught it. Or say Meta publishes the value model's weights. Then everything here about which guesses matter stops being a list. It becomes checkable.

The second limit is harder to sit with. You now know the widths, the guesses and the cold-start problem. That tells you why a new post starts out invisible. It does not tell you which post stops being invisible. Salganik's eight worlds are the reason to expect that nothing ever will. The best songs rarely did badly. The worst rarely did well. Every other outcome stayed on the table. Explaining why outcomes spread out is one job. Picking which one wins is a different job. The two get confused constantly in this subject.

The question worth someone's time is narrower. It can be answered. What does an exposure-controlled test like a trial reel really measure at 24 hours? And how much of that survives to 72?

Which leaves the post you opened this on. Still sitting at a couple of hundred views. If it breaks no rule, nothing was done to it. It went into a funnel that narrows billions of items down to hundreds. One person at a time. It did that separately for every person who might have seen it. It arrived at the stage that sorts by history, carrying none. That is the ordinary condition of every post ever published, for the first few minutes of its life. The 6pm had nothing to do with it. Neither did the audio.

What you have instead of a tactic is smaller. It also lasts longer. Someone will explain your reach to you with a number. Ask which document it is in. Then ask what that number counts.

What we could not find out

Five things, named so no one assumes we checked and stayed quiet.

Where the "500 people" claim started. We can show the number appears in no Meta document we read. We can show what the Feed card's oldest saved copy says. It is dated 29 June 2023. It already used 500 twice. Both times as a count of posts. What we could not do is date the claim's first appearance to a primary source. The pages carrying it earliest in a search are commercial marketing posts. We do not cite those. And we found no dated non-commercial source carrying it. So the story above is one-sided. We know how far back the public record of the document runs. We do not know how far back the folklore runs. So we say nothing about which came first.

Any weight in any ranking model. Meta publishes the guesses and the signals. One list per part of the app. It publishes no weight for any of them. Not on any surface. Not in any document we read. Every claim about which signal dominates is unproven. Ours included.

Any decay curve. The Feed card's "within 90 days" is the only published window on a post's own age. No published curve. No half-life. No first-hour rule. Not in the five ranking documents we read.

The sample size in Muchnik 2013. Science keeps the full text behind a paywall. It refuses automated retrieval too. The abstract is free and gives no sample size. One of the authors posts a pre-publication supplement openly. It does carry counts. It is also stamped "not for citation". So we do not report its numbers as the paper's. Salganik's abstract states 14,341 participants. Van de Rijt's full text is open at PubMed Central. So those two are reported with their samples. Muchnik is not. And no detail of its design beyond the abstract appears above.

Whether the Explore funnel has a fourth step. The 2023 engineering post describes a final reshuffle, after the expensive model. Results may be filtered for integrity, which is Meta's word for keeping rule-breaking and harmful material out. Or shuffled, so that several items by the same author do not run in a row. Today's card names three steps. It describes nothing after the expensive model at all. We are not claiming the step was removed. We are recording two things. Two Meta documents about the same tab count its steps differently. And only the older one says what happens to the list after the score is worked out.

How we did this

Scope: this post reads published documents. No data collection. No experiment of our own.

How we read them: each source was opened directly on 17 August 2026. Every quoted sentence was copied off the page as a reader sees it. Not off a summary. Six of the sources send an empty page first and then build the words with code, so downloading the address on its own returns nothing to read. Those six are the five Meta Transparency Center pages and the Instagram Help Center page, and all six were read in a browser. Five more arrive as finished pages from their own servers. Those are the two Engineering at Meta posts, the two Instagram for Creators posts and Mosseri's ranking post. The research papers were read as the Sources list describes. Archive dates were checked the same day, against the Internet Archive's own index of saved copies. One typographic change is applied to every quotation. Curly apostrophes and quote marks in the sources are rendered as straight ones here, for consistency across the post. No other character is altered.

Limits: eleven of eighteen sources are first-party statements. All from the company we are writing about. A system card is a description Meta chose to publish. Not an audited account. Meta rewrites these cards without saying when. So every quotation here carries the card's "updated" stamp. The one it showed on the day. If we get something wrong, we fix it and date the correction.

Sources

Instagram Explore AI system card, stamped UPDATED JUN 22, 2026. Meta Transparency Center. Read in a browser. Retrieved 17 August 2026.
Instagram Feed AI system card, stamped UPDATED JUN 29, 2026. Meta Transparency Center. The card's exact wording at the two stages this post rounds: "approximately 700 of the most relevant posts" and "about 500 posts". The informative-content entry is listed as an input signal, not as a prediction, under the prediction "How likely you are to share a post with someone in a direct message"; the card does not use the word classifier anywhere. Read in a browser. Retrieved 17 August 2026.
Instagram Reels Chaining AI system card, stamped UPDATED NOV 11, 2025. Its two counts read "approximately 100 of the most relevant reels" and "about 100 reels". Read in a browser. Retrieved 17 August 2026.
Instagram Feed Recommendations AI system card, stamped UPDATED JUN 29, 2026. Meta Transparency Center. Source of the LLM-generated informativeness guess, quoted above as the page renders it, with no full stop. Also the source of the impression-age counts, whose full wording is "How many impressions a user spends viewing media in a specific age range (1-3 days old) over a period of 7 days". Read in a browser. Retrieved 17 August 2026.
Our approach to explaining ranking, stamped UPDATED DEC 31, 2023. Meta Transparency Center. The index the Instagram card list was counted from: ten entries, of which Threads Feed is the tenth. Read in a browser. Retrieved 17 August 2026.
Adam Mosseri, Instagram Ranking Explained, 31 May 2023. About Instagram. Its signal list is introduced as "The most important signals across Feed, roughly in order of importance". Retrieved 17 August 2026.
Vladislav Vorotilov and Ilnur Shugaepov, Scaling the Instagram Explore recommendations system, 9 August 2023. Engineering at Meta. Source of the four-stage description, the value-model formula and the retrieval-model constraint. Retrieved 17 August 2026.
Luke Levis, Sing Sing Ma and Eduardo Nava, Journey to 1000 models: Scaling Instagram's recommendation system, 21 May 2025. Engineering at Meta. Source of the quoted sentence about operating on fewer candidates as the funnel progresses. The funnel layers it names are sourcing (retrieval), early-stage ranking and late-stage ranking. Retrieved 17 August 2026.
Trial reels: Try content with non-followers first to see what performs best, 10 December 2024. Instagram for Creators. Source of three quotations above: the "shown to non-followers first" sentence, the note that some followers may still see a trial reel elsewhere, and the 72-hour auto-share condition. The Help Center's page carries a differently worded sentence about followers still seeing the reel, and describes the same auto-share with no clock, so both quotations are cited to this announcement rather than to it. Retrieved 17 August 2026.
Finding success on reels in 2025, 21 January 2025. Instagram for Creators. Retrieved 17 August 2026.
About Trial Reels on Instagram. Instagram Help Center, undated by the publisher. Client-side rendered, so it returns an empty shell to a fetcher and was read in a browser. It is the source of the public-account requirement, the 200-follower and 1,000-follower thresholds, the duplicate-content note, the independent-ranking statement, the note that a trial reel may take longer to get views, and the clockless auto-share wording quoted above. Retrieved 17 August 2026.
Dong Wang, Junyi Jiao, Arnab Bhadury, Yaping Zhang, Mingyan Gao and Onkar Dalal, Item-centric Exploration for Cold Start Problem. Proceedings of the Nineteenth ACM Conference on Recommender Systems (RecSys '25), Prague, 22 to 26 September 2025, doi:10.1145/3705328.3748113. Full text read at arXiv:2507.09423, 12 July 2025. The arXiv record does not state the authors' employer, so none is claimed here. Method: Bayesian item-quality control in a production exploration system; one online experiment; relative changes on unnamed proprietary metrics; no sample size, no confidence intervals, and the deployed product is not named. No number from this paper is used above, for those reasons. Retrieved 17 August 2026.
Jin-Hu Liu, Tao Zhou, Zi-Ke Zhang, Zimo Yang, Chuang Liu and Wei-Min Li, Promoting Cold-Start Items in Recommender Systems. PLOS ONE 9(12):e0113457, 5 December 2014, doi:10.1371/journal.pone.0113457; preprint arXiv:1404.4936. Method: bipartite-network simulation on the purchase records of two Chinese e-commerce sites, Tmall.com and Coo8.com, under item-based collaborative filtering. Authors' stated limit, verbatim: "We have tested the user-based collaborative filtering, under which the MaxD is usually better than MinD" — that is, the result inverts under user-based collaborative filtering, which is the second standard kind named above. Full text read at PLOS. Retrieved 17 August 2026.
Kiran Rama, Pradeep Kumar and Bharat Bhasker, Deep Learning to Address Candidate Generation and Cold Start Challenges in Recommender Systems: A Research Survey, 17 July 2019. A survey, not a primary result. The two families it names are "additional features (for audio, images, text)" and "learning hidden user and item representations". Abstract read at arXiv. Retrieved 17 August 2026.
Matthew J. Salganik, Peter S. Dodds and Duncan J. Watts, Experimental study of inequality and unpredictability in an artificial cultural market. Science 311(5762):854 to 856, 2006, doi:10.1126/science.1121066. Randomised experiment, 14,341 participants, 48 songs, two conditions and eight independent social-influence worlds. The paper reports two experiments on that design; the second made a world's own download counts more salient than the first, and the contrast between them is the finding quoted above. Inequality measured as the Gini coefficient of market share across songs. Full text and supporting material read at the author-hosted copy; abstract cross-checked at PubMed. Retrieved 17 August 2026.
Lev Muchnik, Sinan Aral and Sean J. Taylor, Social influence bias: a randomized experiment. Science 341(6146):647 to 651, 2013, doi:10.1126/science.1240466. Randomised experiment on a social news aggregation website. The published full text is paywalled and the abstract states no sample size. A pre-publication copy of the supplementary information is downloadable from corresponding author Sinan Aral's own site and does state the counts, but it is stamped "UNDER EMBARGO — NOT FOR CITATION", so no number from it is reported above and the sample size is booked as undetermined. Abstract read at PubMed. Retrieved 17 August 2026.
Arnout van de Rijt, Soong Moon Kang, Michael Restivo and Akshay Patil, Field experiments of success-breeds-success dynamics. PNAS 111(19):6934 to 6939, 2014, doi:10.1073/pnas.1316836111. Randomised field experiments across four live platforms. The samples quoted above are round one: 200 Kickstarter projects, 305 Epinions reviews, 521 Wikipedia editors and 200 change.org campaigns. A second round was run on Kickstarter and Epinions only, so the paper's totals for those two are larger (293 and 481); Wikipedia and change.org have one round each. The Epinions treatment was the site's "very helpful" rating, which is a distinct category above "helpful" on its scale. Round-one results, control against treated: Kickstarter 39% to 70% (χ² = 19.4, P = 0.000), Epinions 77% to 90% (χ² = 9.54, P = 0.002), Wikipedia 31% to 40% (χ² = 4.72, P = 0.030), change.org 52% to 66% (χ² = 4.05, P = 0.044). Full text read at PubMed Central, PMC4024896. Retrieved 17 August 2026.
Internet Archive CDX index, queried 17 August 2026 for the Instagram Feed system-card URL at both hosts. Earliest captures returned and read: 29 June 2023, at transparency.fb.com and 21 May 2024, at transparency.meta.com. A wildcard query over the whole explaining-ranking path at transparency.fb.com returned nothing captured before 29 June 2023.

instagramrankingcold-startrecommender-systemsexploredistribution