How your feeds decide what you see

Following is reverse chronological. The people you chose, newest first, and nothing else. Friends has no ranking of any kind, no counts, and nothing to optimize. Neither one sorts, scores, or decides anything on your behalf.

That leaves two feeds where ranking happens, and this page is about those. One of them, Trending, ships today. The other is For You, which is designed but not built, so the beta runs a simpler ranked feed in its place. What follows is the arithmetic, what we got wrong, and what we have not settled.

In 2018, Facebook changed its News Feed to optimize for what it called meaningful social interactions. The goal was to show you more from friends and family and less from publishers and politicians. The intent was good.

The problem was in the arithmetic. Meaningfulness is not measurable, so they measured engagement instead and called it the same thing. Internal documents later disclosed by whistleblower Frances Haugen, and reported by CNN Business and the Washington Post, show the scale used in the first half of 2018:

A like was worth 1 point.

A reaction emoji or a reshare without text was worth 5.

An RSVP to an event was worth 15.

A comment, message, or reshare judged significant was worth 30.

The total was then multiplied down by how distant the relationship was. Members of the same group, 0.5. Strangers, 0.3.

Read the scale and you can see what it rewards. The cheapest way to score 30 points is to write something someone feels compelled to argue with.

Facebook's own researchers noticed fast. A memo from November 2018, eleven months in, asked directly whether the platform rewarded outrage, and found that posts drawing more negative comments drew more clicks.

By 2019 their data scientists had found that posts collecting angry reactions were disproportionately likely to carry misinformation and low quality news. The angry reaction's weight was cut to zero in late 2020.

None of this was a bug. The system did exactly what the numbers told it to do.

That is the actual lesson, and it is not about Facebook being careless. Any system that asks a single measurement to stand in for two different things will drift toward whichever one is easier to produce. Engagement is easier to produce than meaning. Argument is easier to produce than agreement.

So TownSquare uses three signals and keeps them apart.

Boost means I like this. It is the familiar one. Hold it and you get reactions, and every reaction counts as a boost, because being moved, amused, or angered by something are all ways of saying it was worth seeing.

Coin means this has value. You get 50 a day. A tap gives 5, and you hold for more, up to your whole daily allowance on a single post if you want. They are free, they cannot be bought, and they reset tomorrow. Spending one is a decision with a cost, which is what makes it different in kind from a like.

Downvote means show me less of this.

The rule that holds them apart is short. Coins rank Trending. Boosts and downvotes rank For You. Neither crosses into the other. Popularity does not become value, and value does not become popularity, because they are computed from different numbers in different places.

We got this wrong in our own code and did not notice for eleven weeks. See below.

Downvotes are the mechanic most likely to go wrong. On other platforms they turn into pile-ons: a visible number, a creator watching it climb, and a crowd that has learned it can bury someone by coordinating.

Two things cause that, and neither is the downvote itself. It is the public count, and the creator being able to see it.

So we removed both. A downvote on TownSquare produces no number anywhere. The creator never learns it happened. There is nothing to brigade, because there is nothing to watch move.

What is left is a private instruction to your own feed. You are not judging the post. You are telling us what to bring you less of.

It does still count toward how far a post travels, and we would rather say so than pretend it is purely personal. How much weight it should carry against a boost is not settled. Today the two are equal, one for one, which we think is probably too strong.

Trending answers a different question than For You. Not what are you into, but what did people here think was worth backing.

It runs on coins. Everyone gets 50 a day, free. They cannot be bought, they reset tomorrow, and spending one costs you something, which is what separates it from a like. You can tap to give 5, or hold and give more, up to your whole day's allowance on a single post if you think it deserves it.

The score is one line of arithmetic:

A post's Trending score is its coin count, times two.

That is the whole formula. Boosts do not enter it. Downvotes do not enter it. Watchtime, comments, shares, and follower count do not enter it. There is a second tier called Rising that adds a recency term, so newer posts get a lift that fades to nothing at 48 hours, but the top of Trending is coins alone.

Until recently that was not true. The scoring function also read boosts, at a quarter of a coin's weight, which broke the rule this page just told you about: coins rank Trending, votes rank For You, and they never cross. It was in the code for eleven weeks before we caught it, and it was not there by design. We removed it. Coins alone rank Trending now, exactly as the arithmetic above says. Boosting is a free approval tap that shapes your own For You feed and is walled off from Trending on purpose, so a wave of likes never reaches the community's value leaderboard. And boosting spends no coins: the coins that do rank Trending come from the free fifty everyone gets each day and cannot be bought, so a spot there is never won with a like or a wallet.

Writing the rules down is easy. Knowing what they do at scale is not, and a platform with fifty users cannot find out by waiting.

So we built a simulation: five hundred synthetic posts, growing to seven hundred and twenty over the run, with plausible ages, categories, and coin totals, run through the actual ranking code rather than a description of it, over twelve rounds, all measured before the boost terms came out, which is why the boost finding below reads as a discovery rather than a footnote.

Four things came back.

Featured Today was not about today. It ranks on all-time coin totals with no time component at all, so between 80 and 93 percent of the posts in it were more than 48 hours old, and the median age climbed to around ten days.

It ossified. Zero turnover for ten consecutive rounds. Across twelve rounds, only six distinct posts ever held the five Featured slots.

Rising was redundant. About 87 percent identical to just ranking by coins. The recency term was too weak to reorder anything.

Boosts were decorative. Zeroing every boost in the simulation moved exactly one of thirteen surfaced posts, which is how a rule violation sat in the code for eleven weeks without anyone noticing.

Then we tested alternatives. Ranking Featured on coins received in a trailing window beat the all-time version clearly, and 72 hours was the best of the ones we tried: median age dropped from 181 hours to 80, turnover rose from 10 percent to 18, and the number of distinct posts that ever reached Featured went from six to eight.

A 24-hour window was more volatile. A week was almost no change at all.

One finding did not go our way. Removing the boost terms, which we did because the separation rule required it, made concentration slightly worse: distinct-ever-Featured dropped from six to five and turnover from 10 percent to 8.3. The boost term had been acting as faint noise that occasionally dislodged an incumbent. Removing it was still right, and it makes the case for a real fix stronger rather than weaker.

What we have not done yet. The window is not implemented. What we would rather build is smoother than a hard cutoff: recent coins weighing more than old ones, with old coins losing weight gradually instead of falling off a cliff. Both need the same thing, which is knowing when each coin arrived. That data has been recorded since the beginning and the ranking code simply ignores it, so this is a query change rather than a rebuild.

One more result worth stating, because it is the one we would have preferred to hide. New posts do not break through, and no scoring change we tried fixed that. A post with no coins cannot rank by coins. That is a cold-start problem, and Trending is not where it gets solved.

A page about mechanism should say where the mechanism stops.

How much weight a downvote should carry. Right now it subtracts exactly one, the same as an upvote adds. That is probably too strong. A signal meant to say "show me less" should not cancel someone else's "this is worth seeing" one for one, and we have not chosen what the right ratio is.

How to give smaller accounts a fair shot. We believe discovery should not reward whoever is already large. That is a commitment, not a mechanism, and there is no follower-weighting term in the ranking code today. We have not settled on a formula, and we would rather say so than describe one that does not exist.

What stops someone controlling many accounts. Coins cannot be bought, so buying reach means buying accounts instead. Whether a coin should count for less when the same person spends many of them, and how you would tell, is unsolved. It is the reason identity verification is a live design problem for us rather than a compliance checkbox.

How Trending should handle time. The simulation says a trailing window works and 72 hours was the best we tested. We would rather build something smoother, where recent coins weigh more and older ones fade rather than vanish. Neither is built.

What you see when you turn For You off. The tab is gone, and Following, Trending, and Friends remain. Whether anything should take its place is not written down anywhere.

Live in the beta today. Following, Trending, Friends, and the ranked feed standing in for For You. Coins, upvotes, reactions, private downvotes, and topic categories you can turn up or down. Trending's scoring is the arithmetic above.

Designed, not built. For You itself, the intentional-use controls, creator earnings, and age verification.

Not public yet. The ranking code. Everything on this page is written from it, but you cannot read it directly. We will publish it, and we are not putting a date on it we cannot keep.

The beta is open, no waitlist, and things break. If something here does not match what you find, the page is wrong and we want to know.